Characterization of skin tissue heterogeneity with near-infrared microspectroscopy and its effects on noninvasive measurements of glucose

Transcription

1 University of Iowa Iowa Research Online Theses and Dissertations Fall 2011 Characterization of skin tissue heterogeneity with near-infrared microspectroscopy and its effects on noninvasive measurements of glucose Natalia Victorovna Alexeeva University of Iowa Copyright 2011 Natalia Victorovna Alexeeva This dissertation is available at Iowa Research Online: Recommended Citation Alexeeva, Natalia Victorovna. "Characterization of skin tissue heterogeneity with near-infrared microspectroscopy and its effects on noninvasive measurements of glucose." PhD (Doctor of Philosophy) thesis, University of Iowa, Follow this and additional works at: Part of the Chemistry Commons

2 CHARACTERIZATION OF SKIN TISSUE HETEROGENEITY WITH NEAR-INFRARED MICROSPECTROSCOPY AND ITS EFFECTS ON NONINVASIVE MEASUREMENTS OF GLUCOSE by Natalia Victorovna Alexeeva An Abstract Of a thesis submitted in partial fulfillment of the requirements for the Doctor of Philosophy degree in Chemistry in the Graduate College of The University of Iowa December 2011 Thesis Supervisor: Professor Mark A. Arnold

3 1 ABSTRACT The ability to measure glucose transcutaneously and noninvasively is an exciting prospect. Such a procedure will offer a painless way of glucose self-monitoring improving the lives of people with diabetes by lowering the barriers to optimal glycemic control. Noninvasive measurements involve collecting near-infrared spectra ( cm -1 ; µm) of skin with two optical fibers in a transmission geometry. Previous results indicate that repositioning of the fiber optic interface adversely affects both precision and accuracy of such measurements. Slight movements of the interface increase prediction errors more than 2.5-fold relative to measurements with a stationary rat model. In this dissertation, the chemical heterogeneity of skin tissue is explored as a possible cause for the sensitivity of the measurement to the position of the optical interface. Rat and human skin tissues are mapped by using combination near infrared spectra to provide distributions of the major components of skin: water, collagen type I protein, fat, keratin protein, and two scattering terms of constant and slope. On the basis of the measured heterogeneity, sets of rat and human skin spectra are simulated to investigate the impact of repositioning the fiber-optic interface. Glucose predictions are analyzed for each location of the interface for a series of partial least squares (PLS) calibration vectors established for different locations on the skin. Significant increases in measurement errors are observed for the situation where the PLS calibration models are generated from spectra associated with one location of the interface while subsequent measurements are performed at different locations on the skin matrix. These increases in prediction errors match the 2.5-fold increase observed in vivo. The impact of replacing the spectrum of bovine fat with spectra of native fat for both rat and human skin samples is established. Principal component analysis (PCA) of the spectral residuals reveals that the magnitude of the spectral residuals and the effects of tissue fat content on the quality of the linear regression were decreased.

4 2 The key implication of the research detailed in this dissertation is that chemical heterogeneity of skin tissue must be considered in multivariate models intended for noninvasive glucose measurements. Abstract Approved: Thesis Supervisor Title and Department Date

5 CHARACTERIZATION OF SKIN TISSUE HETEROGENEITY WITH NEAR-INFRARED MICROSPECTROSCOPY AND ITS EFFECTS ON NONINVASIVE MEASUREMENTS OF GLUCOSE by Natalia Victorovna Alexeeva A thesis submitted in partial fulfillment of the requirements for the Doctor of Philosophy degree in Chemistry in the Graduate College of The University of Iowa December 2011 Thesis Supervisor: Professor Mark A. Arnold

6 Graduate College The University of Iowa Iowa City, Iowa CERTIFICATE OF APPROVAL Ph.D. THESIS This is to certify that the Ph.D. thesis of Natalia Victorovna Alexeeva has been approved by the Examining Committee for the thesis requirement for the Doctor of Philosophy degree in Chemistry at the December 2011 graduation. Thesis Committee: Mark A. Arnold, Thesis Supervisor Gary W. Small Lei M. Geng Claudio J. Margulis William I. Sivitz

7 To my family in Russia, United States, and Kazakhstan. ii

8 ACKNOWLEDGEMENTS First and foremost, I would like to express my deep gratitude to my advisor and mentor of many years, Dr. Mark Arnold. He is a man of great vision and integrity, both academical and personal. Without his unwavering support and sound advice this work would have not been possible. Dr. Arnold is the best role model a junior scientist can have and I have been very fortunate to do research under his guidance. Next, I would like to acknowledge my wonderful dissertation committee members, Dr. Gary Small, Dr. Lei Geng, Dr. Claudio Margulis, and Dr. William Sivitz. These professors have shown great interest in my graduate work providing encouragement as well as valuable suggestions. In addition, Dr. Small has been a wonderful teaching mentor sharing teaching tips and strategies and helping me to become a better educator. Dr. Russell Larsen and Dr. Johna Leddy are other two faculty members who have been enormously kind to me helping to improve my teaching skills. This research has largely been explorative and data collection has been made possible by hard work of Dr. Jonathon Olesberg, Dr. Liu, Dr. Chuannan Bai and lovely Terri Graham. They set me up for success with their knowledge of instrumentation and animals. Dr. Chris Coretsopoulos graciously shared his biological microscope for collection of dry skin images. Dr. Bickenbach provided some of the human skin samples. The rest of the samples came from the Deeded Body Program at the University of Iowa. I would like to thank Arnold s (ARG) group members for many fruitful discussions. You, guys, are the best! I wish Joo Young, Jue, Dede, Ryan, Madhuri, Bimali, Yatian, Hankyu, Dr. David Cho, and Sherif good luck in their personal and professional lives. Lastly, these acknowledgements are not complete without mentioning the center and the meaning of my life, my dear family. This dissertation is as much their accomplishment as mine and I love them so much. Thank you all! iii

9 ABSTRACT The ability to measure glucose transcutaneously and noninvasively is an exciting prospect. Such a procedure will offer a painless way of glucose self-monitoring improving the lives of people with diabetes by lowering the barriers to optimal glycemic control. The noninvasive measurements involve collecting near-infrared spectra ( cm-1; μm) of skin with two optical fibers in a transmission geometry. Previous results indicate that repositioning of the fiber optic interface adversely affects both precision and accuracy of such measurements. Slight movements of the interface increase prediction errors more than 2.5-fold when performed with a stationary rat model. In this dissertation, the chemical heterogeneity of skin tissue is explored as a possible cause for the sensitivity of the measurement to the position of the optical interface. Rat and human skin tissues are mapped by using combination near infrared spectra the to provide distributions of the major components of skin: water, collagen type I protein, fat, keratin protein, and two scattering terms of constant and slope. On the basis of the measured heterogeneity, sets of rat and human skin spectra are simulated to investigate the impact of repositioning the fiber-optic interface. Glucose predictions are analyzed for each location of the interface for a series of partial least squares (PLS) calibration vectors established for different locations on the skin. Significant increases in the measurement errors are observed for the situation where the PLS calibration models are generated from spectra associated with one location of the interface while subsequent measurements are performed at slightly locations of the skin matrix. These increases in prediction errors match the 2.5-fold increase observed in vivo. The impact of replacing the spectrum of bovine fat with spectra of native fat for both rat and human skin samples is established. Principal component analysis (PCA) of the spectral residuals reveals that the magnitude of the spectral residuals and the effects of tissue fat content on the quality of the linear regression were decreased. iv

10 The key implication of the research detailed in this dissertation is that chemical heterogeneity of skin tissue must be considered in multivariate models intended for noninvasive glucose measurements. v

11 TABLE OF CONTENTS LIST OF TABLES... vi LIST OF FIGURES... viii CHAPTER I INTRODUCTION...1 Importance of Noninvasive Glucose Monitoring...1 Diabetes...1 Strategies for Self Monitoring of Glucose...2 Methods of Noninvasive Glucose Sensing...4 Noninvasive Glucose Sensing with Near Infrared Spectroscopy...7 Near Infrared Spectroscopy of Bioanalytes...11 Origins of Near Infrared Spectral Bands...12 Near Infrared Vibrational Microspectroscopy...13 Overview of the Thesis...18 CHAPTER II MICROSPECTROSCOPY FOR NEAR-INFRARED ANALYSIS OF BIOLOGICAL SAMPLES...20 Introduction...20 Measurement Modes in Vibrational Microspectroscopy...20 Mapping versus Imaging...22 Resolution in Vibrational Microspectroscopy...22 Chapter Overview...23 Experimental Procedures...24 Instrumental Setup...24 Manual Focusing Procedure...26 Sample Preparation...28 Spectral Data Collection...29 Spectral Noise Calculation...30 Estimation of the Captured Area...31 Standard Components...31 Linear Regression Models...34 Results and Discussion...35 Optimization of Spectral Quality...35 Detector Type Substitution...35 Selection of Wavenumber Range...39 SNR Dependence on the Number of Co-Added Interferograms...41 Diameter of the Captured Area...44 Testing of the Spectral Standards with Biological Samples...46 Rat Nail Spectra...48 Lipid Suspension Spectra...50 Dried Rat Skin...53 Conclusions...57 ii

12 CHAPTER III CHARACTERIZATION OF SPATIAL HETEROGENEITY OF RAT SKIN EX VIVO WITH NEAR-INFRARED MICROSPECTROSCOPY...58 Introduction...58 Rat Tissue Heterogeneity in Relation to Near Infrared Spectroscopy...58 Composition and Structure of Rat Skin Tissue...58 Physiological Basis for Skin Tissue Heterogeneity...60 Chapter Overview...61 Experimental Procedures...63 Rat Skin Samples and Reagents...63 Instrumental Setup...63 Spectral Data Collection...64 Data Processing...68 Absorbance Spectra and Spectral Quality...68 Spectral Fitting...69 Construction of 2D Tissue Maps...69 Correlation between Tissue Components...70 Analysis of Variance...71 Results and Discussion...72 Spectral Quality Assessment...72 Microspectroscopic Data...72 RMS Noise for Air Spectra...73 RMS Noise for Rat Skin Spectra...78 Rat Skin Tissue Absorbance Data...81 Visual Inspection of Rat Skin Absorbance Spectra...81 Six Component Regression Fit...85 Rat Skin Spatial Heterogeneity...88 Spatial Contour Maps...88 Six Component Distributions for Rat Skin...93 Relative Regression Coefficients (%) for Chemical Components Rat Skin ANOVA Data Correlations between Skin Components Conclusions CHAPTER IV CHARACTERIZATION OF SPATIAL HETEROGENEITY OF HUMAN SKIN EX VIVO WITH NEAR-INFRARED MICROSPECTROSCOPY Introduction Optical Heterogeneity of Human Skin Tissue Chapter Overview Experimental Procedures Preparation of Human Skin Samples Instrumental Setup Spectral Data Collection iii

13 Data Processing Absorbance Spectra and Spectral Quality Spectral Fitting and Two-Dimensional Tissue Maps Correlation between Tissue Components Results and Discussion Spectral Quality Assessment Human Skin Tissue Absorbance Data Visual Inspection of Absorbance Spectra Six Component Regression Fit Human Skin Spatial Heterogeneity Spatial Contour Maps Six Component Distributions for Human Whole Skin Relative Regression Coefficients (%) for Chemical Components Correlations between Components of Whole Skin Human Skin Layers Tissue Absorbance Data Visual Inspection of Absorbance Spectra Six Component Regression Fit Human Skin Layers Spatial Heterogeneity Spatial Contour Maps Six Component Distributions for Human Skin Layers Relative Regression Coefficients (%) for Chemical Components Correlations between Components of Skin Layers Conclusions CHAPTER V LOCATION-SPECIFIC PLS CALIBRATION MODELS FOR GLUCOSE IN RAT AND HUMAN SKIN TISSUE Introduction Repositioning Errors in Noninvasive Glucose Measurements in Skin Principal Component Analysis (PCA) of Tissue Spectra Partial Least Squares (PLS) Regression Evaluation of Partial Least Squares (PLS) Regression Chapter Overview Experimental Procedures Simulation of Skin Tissue Spectra Addition of Spectal Variance Addition of Glucose Component Addition of Spectal Noise PLS Models for Glucose Predicton Evaluation of PLS Models for Glucose Predicton Results and Discussion Spectral Simulation Study Optimization of PLS Calibration Model Representative Results of PLS Regression iv

14 Effects of Rat Skin Heterogeneity on Glucose PLS Prediction Effects of Human Skin Heterogeneity on Glucose PLS Prediction Conclusions CHAPTER VI IMPROVEMENTS TO THE SIX COMPONENT SKIN TISSUE MODEL ACHIEVED WITH RAT AND HUMAN FAT SPECTRAL STANDARDS Introduction Lipids in Skin Residual Spectral Features of the Skin Tissue Model Chapter Overview Experimental Procedures Prearation of Fat Tissue Samples Instrumentation and Spectral Data Collection Data Processing PCA on Residual Spectra Results and Discussion Native Fat Standards Six Component Regression Model with Bovine and Native Fat Standards Correlation between Fat Content and SSRES Values PCA of the Spectral Residuals Conclusions CHAPTER VII FUTURE WORK APPENDIX REFERENCES v

15 LIST OF TABLES Table II-1. RMS noise values across the combination region ( cm -1 ) for 1 mm thick water samples collected with a K band filter Table II-2. Table II-3. Table III-1. Linear dependence of SNR on the number of co-added interferograms...4 Regression coefficients summary for dried rat skin sample measured after 5 and 29 days of desiccation Summary of lower and higher microscope signal groups of rat skin maps Table III-2. RMS noise values on 100% lines of air reference spectra Table III-3. RMS noise values on 100% lines of tissue reference spectra for the lower signal group of rat skin data Table III-4. Summary of regression coefficients for rat skin samples Table III-5. Relative regression coefficients (%) summary for rat skin samples Table III-6. Summary of correlation coefficients for male rats Table III-7. Summary of correlation coefficients for female rats Table IV-1. Demographics of human skin samples Table IV-2. RMS noise values on 100% lines of air reference spectra Table IV-3. RMS noise values on 100% lines of skin tissue spectra Table IV-4. Regression coefficients summary for human whole skin samples Table IV-5. Table IV-6. Summary of correlation coefficients for human whole skin samples collected from female Specimens #1 and #3 (± standard error) Summary of correlation coefficients, r, for human whole skin samples collected from male Specimens #2 and #5 (± standard error) Table IV-7. Relative regression coefficients (%) summary for human skin layers Table IV-8. Correlation coefficients, r ± standard error, for human skin layers samples Table V-1. Average regression coefficients for 4 quadrants of map # vi

16 Table V-2. Table V-3. Table V-4. Table V-5. Table V-6. Table V-7. Table VI-1. Table VI-2. Results of PLS regression for 4 quadrants of map #81 with 7 latent variables Summary of standard errors (SEC, SEP) and standard error ratios (SEP/SEC Ratio, %) for male rats List of standard error ratios (SEP/SEC Ratio, %) outliers for rat skin samples, not included in boxplots Summary of standard errors (SEC, SEP) and standard error ratios (SEP/SEC Ratio, %) for female rats Summary of standard errors (SEC, SEP) and standard error ratios (SEP/SEC Ratio, %) for human skin samples List of standard error ratios (SEP/SEC Ratio, %) outliers for human skin samples, not included in boxplots Sum of squares of residuals (SSRES) for rat and human skin samples obtained with the original bovine standard and the new native fat standards Pearson coefficients, R, for correlation between bovine and native fat content and sum of squares of residuals (SSRES) vii

17 LIST OF FIGURES Figure I-1. Figure I-2. Figure I-3. Figure I-4. Glucose (solid line) and other biochemical analytes (dashed lines) absorptivity spectra collected from aqueous solutions at 37 C....6 The instrumental setup for the in vivo glucose measurement on an anesthesized rat. General view is shown in (A); detailed view of the measurement site is added in (B)....8 Effects of fiber repositioning on (A) PLS prediction of glucose concentrations and (B) amounts of water, collagen type I protein, fat, keratin protein, constant, and slope components Schematic representation of a hypercube of microspectroscopic data. 75 Dimensions X and Y define spatial location of each pixel and dimension λ contains absorbance values at the n-wavelengths (λ 1 λ n ) for the spectra collected at each pixel Figure I-5. Schematic diagram of Michelson interferometer...17 Figure II-1. Schematic diagram of the IR Plan Advantage microscope Figure II-2. Figure II-3. Figure II-4. Figure II-5. Manual focusing procedure for the IR Plan Advantage microscope. The first step is shown in (A), when the condenser and the sample stage are moved in sync to reach to focal point of the objective; in (B) the height of the condenser is adjusted to ensure optimal focusing as shown in (C)...27 Estimation of the captured area with selected CCD images of the orange laboratory tape shown at the distance of 0 μm (1), 40 μm (2), and 160 μm (3) Combination near-infrared spectra for water, collagen type I protein, fat, keratin protein, constant, and slope Combination near-infrared single beam spectra for (A) air and (B) water inside the 1-mm thick Infrasil cell. Data for MCT/A detector is plotted in red and for InSb detector in blue...37 Figure II-6. Combination near-infrared absorbance spectra for water inside the 1- mm thick Infrasil cell. Spectra collected with a K band-pass filter is plotted in red for MCT/A detector and in blue for InSb detector. The spectrum plotted in green is for the MCT/A detector without the K band filter...38 viii

18 Figure II-7. Figure II-8. Figure II-9. Signal-to-noise ratio (SNR) dependence on the number of co-added interferograms. The error bars correspond to the standard deviation of the SNR values derived from different pairs of triplicate single beam spectra...42 Determination of the microscope measured area. Absorbance spectrum of the orange laboratory tape is shown in (A); in (B) the normalized absorbance values at 4331 cm -1 are plotted against the distance between the edge of the tape and the focal point. Green dash-dot line defines the radius of the measured area at 160 μm Nail samples modeling with the spectral standards. The schematic diagram of the aluminum plate in (A) shows the dimensions of the sample holder. The opening filled with nail are colored orange, with air white. Combination absorbance spectra in (B) were taken at all of the 9 locations. Distribution map for keratin standard is presented in (C) Each measured location is noted by a green open circle in this map. The lighter shade of color corresponds to the higher keratin content and the darker shade to the lower content Figure II-10. Lipid suspension modeling with the spectral standards. The images in (A) were collected in the visible mode with the CCD camera. The darker regions show the lipid flakes and the lighter regions water. Locations 7 (mostly water) and 12 (lipid particles) are emphasized for comparison. Absorbance spectra in (B) were taken at all of the 15 locations. Distribution maps are presented for (C) water and (D) fat with each measured location denoted. For both components, the lighter shade of color corresponds to the higher content and the darker shade to the lower content Figure II-11. Absorbance spectra for a male rat skin sample dried in the desiccator at 1.3 C for (A) 5 days and (B) 29 days Figure III-1. Figure III-2. Microscopic images of dried rat skin from the area between the shoulders. Images (A) and (B) show the edge of the skin with hair fibers visible; (C) shows the center of the sample with white rectangles marking particularly heterogenous areas. Data collected with a universal photoscope (Carl Zeiss Microimaging, Thornwood, NY) in transmission mode. The size of the white bar is 100 μm Photographs of the microspectroscopic mapping setup showing (A) motorized stage with a sample cell, (B) compression cell with Teflon spacers for air measurements, (C) compression cell with a skin sample inside, and (D) top view of the skin sample inside the cell ix

19 Figure III-3. Figure III-4. Figure III-5. Figure III-6. Figure III-7. Figure III-8. Figure III-9. Schematic of the microspectroscopic mapping setup including (A) the compression cell, (B) the skin sample with the raster pattern for the measured spectral array, and (C) the CCD camera images of skin tissue. Red dot in the center of the spectral array marks the reference location. The length of the white bar on the CCD skin images is 100 μm RMS noise on 100% lines of air reference spectra for the lower signal group of data in the regions between (A) cm -1 and (B) cm -1 ; the higher signal group of data in the regions between (C) cm -1 and (D) cm RMS noise on 100% lines of rat skin tissue spectra for the lower signal group of data between (A) cm -1 and (B) cm -1 ; the higher signal group of data between (C) cm -1 and (d) cm Representative spectral data for rat skin samples: (A) and (B) are the absorbance spectra for point maps of male rat sample #27 and female rat sample #70, respectively; (C) and (D) are the corresponding reference absorbance spectra at the central location of the skin slices; and (E) and (F) are the corresponding absorbance residuals for these maps after the six-component fit...82 Representative regression coefficients for skin samples: (A) and (B) are the results of regression fit for point maps of male rat sample #27 and female rat sample #70 respectively; (C) and (D) are the corresponding regression coefficients at the central location on the skin slices...86 Spatial distribution maps for the male rat sample #27, where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped skin slice. Black contours denote 10% change in regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement...90 Spatial distribution maps for the female rat sample #70, where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped skin slice. Black contour lines denote 10% change in regression coefficient. Two white circles represent the size of the optical fiber interface used in the related in vivo experiment...91 x

20 Figure III-10. Box plots for (A) water, (B) collagen type I protein, and (C) fat regression coefficients of male rats. Black circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent the values between 25% and 75% of the distribution Figure III-11. Box plots for (A) keratin protein, (B) constant, and (C) slope regression coefficients of male rats. Black circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent the values between 25% and 75% of the distribution Figure III-12. Box plots for (A) water, (B) collagen type I protein, and (C) fat regression coefficients of female rats. Black circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent the values between 25% and 75% of the distribution Figure III-13. Box plots for (A) keratin protein, (B) constant, and (C) slope regression coefficients of female rats. Black circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent the values between 25% and 75% of the distribution Figure III-14. Histograms for (A) fat distribution for male rat 1 and B) collagen type I protein distribution for female rat Figure III-15. Box plots for (A) water and (B) collagen type I protein relative regression coefficients (%) of male rats. Black circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent the values between 25% and 75% of the distribution. The data between 25% and 75% from the in vivo experiments on two rats are represented by green boxes Figure III-16. Box plots for (A) fat and (B) keratin protein relative regression coefficients (%) of male rats. Black circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent the values between 25% and 75% of the distribution. The data between 25% and 75% from the in vivo experiments on two rats are represented by green boxes Figure III-17. Box plots for (A) water and (B) collagen type I protein relative regression coefficients (%) of female rats. Black circles represent mean values, red horizontal lines median values, red plus signs xi

21 outliers, and blue box the values between 25% and 75% of the distribution Figure III-18. Box plots for (A) fat and (B) keratin protein relative regression coefficients (%) of female rats. Black circles represent mean values, red horizontal lines median values, red plus signs outliers, and blue box the values between 25% and 75% of the distribution Figure III-19. Examples of correlation plots for (A) strong positive, (B) strong negative, and (C) not significant correlation between rat skin tissue components. Red line on the plot is the linear fit to the data with the correlation coefficient reported in each case Figure IV-1. Microscopic images of dried human skin from the area between on the knuckle of a hand. Images (A) and (B) show the skin with wrinkles visible. Data collected with a biological microscope (Carl Zeiss Microimaging, Thornwood, NY) in transmission mode. The instrument was calibrated with a micrometer (SpectraTech, Inc., Shelton, CT). The size of the red bar is 100 μm Figure IV-2. RMS noise on 100% lines of air reference spectra in the regions between (A) 4,400-4,500 cm -1 and (B) 4,500 4,600 cm -1. The data are arranged in the order the maps were collected. The vertical black lines separate maps collected from specimen 1 (#82-85), specimen 2 (#86-89), specimen 3 (#90-93), specimen 5 (#95-98), epidermis (E1, E2), and dermis (D1, D2) samples Figure IV-3. RMS noise on 100% lines of human skin tissue spectra in the regions between (A) 4,400-4,500 cm -1 and (B) 4,500 4,600 cm -1. The data are arranged in the order the maps were collected. The vertical black lines separate maps collected from specimen 1 (#82-85), specimen 2 (#86-89), specimen 3 (#90-93), specimen 5 (#95-98), epidermis (E1, E2), and dermis (D1, D2) samples. * Higher RMS noise values are observed for map #90 because of its higher collagen type I content and E2 map because it was collected across epidermis folded in half Figure IV-4. Spectral data for whole skin samples from female subjects: (A) and (B) are the absorbance spectra for point maps of sample #82 and sample #90, respectively; (C) and (D) are the corresponding reference absorbance spectra at the central location on the skin slices; (E) and (F) are the corresponding absorbance residuals for these maps after the six-component fit Figure IV-5. Spectral data for whole skin samples from male subjects: (A) and (B) are the absorbance spectra for 100 point maps of sample #88 and sample #95, respectively; (C) and (D) are the corresponding xii

22 reference absorbance spectra at the central location on the skin slices; (E) and (F) are the corresponding absorbance residuals for these maps after the six-component fit Figure IV-6. Representative regression fit data for whole skin samples from female subjects: (A) and (B) are the results of regression fit for point maps of sample #82 and sample #90, respectively; (C) and (D) are the corresponding regression coefficients at the central location on the skin slices Figure IV-7. Representative regression fit data for whole skin samples from male subjects: (A) and (B) are the results of regression fit for point maps of sample #88 and sample #95, respectively; (C) and (D) are the corresponding regression coefficients at the central location on the skin slices Figure IV-8. Spatial distribution maps for the sample #82 from a female subject, where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped skin slice. Black contours denote 10% change in the regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement Figure IV-9. Spatial distribution maps for the sample #95 from a male subject, where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped skin slice. Black contours denote 10% change in the regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement Figure IV-10. Box plots for (A) water, (B) collagen type I protein, and (C) fat regression coefficients of whole skin samples. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by magenta (female specimens) and blue (male specimens) boxes Figure IV-11. Box plots for (A) keratin, (B) constant, and (C) slope regression coefficients of human whole skin samples. Black circles represent mean values, red horizontal lines median values, red plus signs xiii

23 outliers. The values between 25% and 75% of the distribution are represented by magenta (female specimens) and blue (male specimens) boxes Figure IV-12. Box plots for (A) water, and (B) collagen type I protein relative regression coefficients (%) of human whole skin samples. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by magenta (female specimens) and blue (male specimens) boxes Figure IV-13. Box plots for (A) fat, and (B) keratin protein relative regression coefficients (%) of human whole skin samples. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by magenta (female specimens) and blue (male specimens) boxes Figure IV-14. Examples of correlation plots for (A) strong positive, (B) strong negative, and (C) no correlation between human whole skin tissue components. Red line on the plot is the linear fit to the data with the correlation coefficient reported in each case Figure IV-15. Spectral data for epidermis samples: (A) and (B) are the absorbance spectra for 100 point maps of sample #E1 and sample #E2 (folded epidermis), respectively; (C) and (D) are the corresponding reference absorbance spectra at the central location on the sample; (E) and (F) are the corresponding absorbance residuals for these maps after the six-component fit Figure IV-16. Spectral data for dermis samples: (A) and (B) are the absorbance spectra for 100 point maps of sample #D1 and sample #D2, respectively; (C) and (D) are the corresponding reference absorbance spectra at the central location on the sample; (E) and (F) are the corresponding absorbance residuals for these maps after the six-component fit Figure IV-17. Regression fit data for epidermis samples: (A) and (B) are the results of regression fit for point maps of sample #E1 and sample #E2 (folded epidermis), respectively; (C) and (D) are the corresponding regression coefficients at the central location on the sample Figure IV-18. Representative regression fit data for dermis samples: (A) and (B) are the results of regression fit for point maps of sample #D1 and sample #D2 (folded epidermis), respectively; (C) and (D) are xiv

24 the corresponding regression coefficients at the central location on the sample Figure IV-19. Spatial distribution maps for the sample of a single layer of epidermis (map #E1), where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped sample. Black contours denote 10% change in the regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement Figure IV-20. Spatial distribution maps for the sample of a double layer of epidermis (map #E2), where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped sample. Black contours denote 10% change in the regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement Figure IV-21. Spatial distribution maps for the sample of a single layer of dermis (map #D1), where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped sample. Black contours denote 10% change in the regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement Figure IV-22. Box plots for (A), (B) water, (C), (D) collagen type I protein, and (E), (F) fat regression coefficients of epidermis and dermis layers, respectively. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by green (epidermis samples) and cyan (dermis samples) boxes Figure IV-23. Box plots for (A), (B) keratin, (C), (D) constant, and (E), (F) slope regression coefficients of epidermis and dermis layers, respectively. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by green (epidermis samples) and cyan (dermis samples) boxes Figure IV-24. Box plots for (A), (B) water, and (C), (D) collagen type I protein relative regression coefficients (%) of epidermis and dermis layers, respectively. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between xv

25 25% and 75% of the distribution are represented by green (epidermis samples) and cyan (dermis samples) boxes Figure IV-25. Box plots for (A), (B) fat, and (C), (D) keratin protein relative regression coefficients (%) of epidermis and dermis layers, respectively. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by green (epidermis samples) and cyan (dermis samples) boxes Figure V-1. Figure V-2. Figure V-3. Illustration of glucose NAS vector calculation in three-dimensional space Summary of our approach to quantify the influence of skin tissue heterogeneity on noninvasive glucose prediction. 185 White block addresses processing of the in vitro spectral data from Chapters II and III, green block explains simulation of spectra at various locations across the skin, and blue block describes PLS procedures Simulation of optical fiber movement between (A) 4 quadrants and (B) 36 possible locations on the skin slice Figure V-4. Flowcharts illustrating the PLS regression procedures for (A) 4 quadrants and (B) 36 possible locations on the skin slice. Each cycle is reproduced 10 times Figure V-5. Figure V-6. Distribution maps for sample #81 with 4 quadrants delineated for each skin component: (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant offset, and (E) sloping baseline. Location of the four quadrants is denoted in the distribution map for slope coefficients Simulated absorbance spectra for map #81 with the microscopically measured absorbance spectra shown in (A), average spectra for 4 quadrants presented in (B), and the average spectra for 36 random locations on the map featured in (C) Figure V-7. Cross validation standard error of prediction (CV-SEP) values for 4 quadrants of sample # Figure V-8. Figure V-9. B-vectors from 4 quadrants of map #81 obtained with (A) 6 and (B) 7 latent variables PC vectors for simulated skin tissue spectra for 4 quadrants of map#81. Principal components for quadrants 1-4 are shown in plots (A)-(D), respectively xvi

26 Figure V-10. Comparison of NAS and PLS b-vectors (averaged for 4 quadrants of map #81) with an absorptivity spectrum of glucose Figure V-11. Concentration correlation plots for quadrants 1 (A, C, E) and 2 (B, D, F) of map #81. Data is shown for concentrations between 10 and 25 mm to increase the visibility of the offset. The concentrations used for calibration are presented by red circles, the predicted concentrations by blue circles. Red solid line demonstrates 100% correlation and the blue line is the linear fit to the predicted concentration values from 5 to 35 mm of glucose Figure V-12. Concentration correlation plots for quadrants 3 (A, C, E) and 4 (B, D, F) of map #81. Data is shown for concentrations between 10 and 25 mm to increase the visibility of the offset. The concentrations used for calibration are presented by red circles, the predicted concentrations by blue circles. Red solid line demonstrates 100% correlation and the blue line is the linear fit to the predicted concentration values from 5 to 35 mm of glucose Figure V-13. SE ratios for male rats obtained for (A) 4 quandrants and (B) 36 random locations on the tissue. Outlier values for the rats marked with a star (*) that are greater than (A) 2500% and (B) 10000% are not included in these plots for clarity of presentation and can be found in Table V-4. Mean values are shown as green circles. The slashed black line is added for comparison with the in vivo SEP/SEC ratio of 255% Figure V-14. SE ratios for female rats obtained for (A) 4 quandrants and (B) 36 random locations on the tissue. Outlier values for the rats marked with a star (*) that are greater than (A) 2500% and (B) 4000% are not included in these plots for clarity of presentation and can be found in Table V-4. Mean values are shown as green circles. The slashed black line is added for comparison with the in vivo SEP/SEC ratio of 255% Figure V-15. SE ratios for human skin obtained for (A) 4 quandrants and (B) 36 random locations on the tissue. Outlier values for the specimens marked with a star (*) that are greater than (A) 5000% and (B) 50000% are not included in these plots for clarity of presentation and can be found in Table V-7. Mean values are shown as green circles. The slashed black line is added for comparison with the in vivo SEP/SEC ratio of 255% Figure VI-1. Chemical structures of major classes of skin lipids: (A) fatty acid (palmitic acid shown), (B) ceramide, (C) cholesterol, (D) sphingomyelin. 1 Structure of n-hexadecane is shown in (E) xvii

27 Figure VI-2. Near infrared spectrum of n-hexadecane collected with IR microscope in the combination region ( cm -1 ). The black dotted line separates region cm -1 where C-H vibrational peaks are prominent Figure VI-3. Average spectral residuals after the six-component fit for samples of (A) rat skin and (B) human skin. Bovine fat mirror with 1 mm thickness is used as the fat standard Figure VI-4. Six component linear for (A) rat and (B) human skin spectra with high fat, represented by blue solid line, and low fat content, represented by black solid line. The corresponding regressed spectra are shown as the red dashed curve for the low and as the solid green curve for the high fat coefficient Figure VI-5. Comparison of bovine fat standard and the rat fat standard. In (A), pure component spectra are shown; in (B), average residual for rat skin sample #49 is compared with the differential absorbance spectrum between rat and bovine fat standards Figure VI-6. Comparison of bovine fat standard and the human fat standard. In (A), pure component spectra are shown; in (B), average residual for human skin sample #88 is compared with the differential absorbance spectrum between human and bovine fat standards Figure VI-7. Regression coefficients and spectral residuals for rat sample #49 obtained with (A), (B) bovine fat standard and (C), (D) rat fat standard Figure VI-8. Regression coefficients and spectral residuals for human sample #88 obtained with (A), (B) bovine fat standard and (C), (D) human fat standard Figure VI-9. Spatial distribution plots for bovine fat regression coefficients of (A) rat map# 49 and (B) human map #88. The distributions for native fat standards are shown for the same maps in (C) and (D), respectively. Black contours denote 10% change in regression coefficient Figure VI-10. Spatial distribution plots for sum of squares of spectral residuals (SSRES) for rat skin sample #49. SSRES for the whole combination spectral region, cm -1, are plotted in (A) and (B); for the C-H absorption region between 4400 and 4200 cm -1 in (C) and (D); xviii

28 and for the protein related absorption between 4900 and 4400 cm -1 in (E) and (F). Bovine fat standard was used for (A), (C), and (E) plots. Rat fat standard is used for (B), (D), and (F). Black contours denote 10% change in the SSRES Figure VI-11. Spatial distribution plots for sum of squares of spectral residuals (SSRES) for human skin sample #88. SSRES for the whole combination spectral region, cm -1, are plotted in (A) and (B); for the C-H absorption region between 4400 and 4200 cm -1 in (C) and (D); and for the protein related absorption between 4900 and 4400 cm -1 in (E) and (F). Bovine fat standard was used for (A), (C), and (E) plots. Human fat standard is used for (B), (D), and (F). Black contours denote 10% change in the SSRES Figure VI-12. PCA loadings for rat skin sample #49 with 1-6 latent variables obtained with (A), (C), (E) the bovine fat standard and with (B), (D), (F) the rat fat standard Figure VI-13. PCA loadings for human skin sample #88 with 1-6 latent variables obtained with (A), (C), (E) the bovine fat standard and with (B), (D), (F) the human fat standard Figure A-1. Plots of ratios of regression coefficients for map (A) # 27 and (B) # Figure A-2. Gaussian distribution fits for A) water, B) collagen type I protein, C) fat, D) keratin protein, (E) constant, and (F) slope regression coefficients of male rats Figure A-3. Gaussian distribution fits for A) water, B) collagen type I protein, C) fat, D) keratin protein, (E) constant, and (F) slope regression coefficients of female rats Figure A-4. Spatial distribution maps for the sample of a single layer of dermis (map #D2), where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped sample. Black contours denote 10% change in the regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement xix

29 1 CHAPTER I INTRODUCTION Importance of Noninvasive Glucose Monitoring Diabetes Diabetes is a chronic disease that is caused by disruptions in the normal glycolytic pathways. Glucose is a primary source of energy for living cells. 1 In the body, glucose concentrations are regulated by insulin, a pancreatic hormone. Type I diabetes is characterized by an inability of the body to produce insulin. This type of diabetes accounts for 5% of all diagnosed cases. People suffering from Type II diabetes cannot utilize insulin correctly. In other words, they exhibit insulin resistance. Gestational diabetes, a temporary condition in pregnant women, is also observed. Women who had abnormal blood glucose concentrations during pregnancy have a 35 to 60 % chance of developing diabetes in the next 10 to 20 years. The complications of diabetes are numerous and some are particularly serious. They range from mood changes and high blood pressure 2 to various complications of microcirculation, stroke, kidney failure, blindness, and death. In 2011, diabetes is reported as the seventh leading cause of death in the United States. The disease affects more than 25 million people in this country alone, which is more than 8 % of the U. S. population. 3 The cumulative costs related to the treatment of diabetes and its complications, disability, work loss, and increased mortality exceeded 174 billion U. S. dollars in Currently, no cure for diabetes is available. Conventional treatment strategies apart from insulin injections for Type I patients and oral medications for Type II largely rely on lifestyle changes. 4 The most important adjustments include healthy diet and exercise regimen, as well as meticulous self-monitoring of blood glucose (SMBG). 5 Knowing the correct blood glucose concentration is necessary to identify conditions of

30 2 hypo- and hyperglycemia, that is extremely low and high blood sugar. When a person becomes aware of their current glucose levels, the correct actions can be taken to prevent complications and maintain euglycemia. Tight glycemic control invariably maintained throughout life can extend a patient s lifespan by 5 to 8 years. 6 Strategies for Self Monitoring of Glucose Modern conventional methods of SMBG are painful and invasive. A drop of capillary blood collected by piercing the skin with a small lancet is placed onto the test strip of the portable meter. 4 The meter reading is based on the electrochemical reactions involving glucose and the measured current is glucose-selective. Such a device requires periodic recalibration and is not devoid of interferences. For example, electrochemical meters measurements are particularly sensitive to temperature changes. 7 Overall, clinically acceptable standard errors, portability of the meter, and sufficient performance stability have made the finger prick method the route of choice for SMBG. Major disadvantages of the electrochemical meters include pain, risk of infection, and high cumulative cost of materials. A person with Type I diabetes is often recommended to check his/her blood sugar at least 4 times a day or more. Each test strip costs around one dollar depending on the meter. Obviously, these costs accumulate over time. More importantly, only discrete values of glucose concentrations can be obtained and trends in glucose concentration changes are difficult to determine in a reasonable time period. Often, knowing the trend in glucose levels is necessary for proper and timely interventions. Invasive test strip-based meters do not provide the ability of continuous glucose monitoring (CGM). All of these disadvantages jeopardize strict adherence to SMBG routine for some patients. 8 Alternative methodologies of blood glucose measurement have been under development for over 40 years. 9, 10 A large group of alternative sensors achieve continuous glucose monitoring (CGM) by means of implanting a sensing element into the subcutaneous tissue. This modality relies on detection of glucose concentrations in the

31 3 person s interstitial fluid (ISF) that are correlated with blood glucose values CGM sensors are often developed with the ultimate goal of closed-loop glycemic control that would mimic pancreatic function. 5, 15 Specifically, based on the output of the implanted sensor, either insulin or a glucose-increasing medication would be injected into the body to attain euglycemia. The chemistry of the CGM sensing devices involves electrochemical detection of the products of the enzymatic reactions of glucose. 16 Full biocompatibility of these sensors is not yet achieved. 17, 18 Periodic re-calibration or even complete reimplantation of the sensor is needed to ensure acceptable sensitivity. The human subject s immune system reacts to the introduction of the extraneous object by trying to isolate the implant from the tissues. Extensive research is underway to eliminate or at least alleviate the effects of the biological response on the performance of the CGM devices. Although these devices allow continuous monitoring of ISF glucose levels, the implantation procedure is invasive and involves a risk of infection. An ideal device for SMBG would be free of all of the above-mentioned disadvantages. Such a monitor should be able to provide continuous or nearly continuous readings of glucose concentrations in a reliable and noninvasive way. A wealth of research has been done in the area of noninvasive glucose sensing in the last 20 years. 10, 19 Unfortunately, none of the proposed noninvasive sensors have yet been approved by the Food and Drug Administration (FDA). The goal of achieving painless and accurate glucose measurements is very attractive and will greatly benefit the global community by lowering the costs associated with diabetes. However, a number of significant obstacles have retarded commercialization of noninvasive devices. According to Khalil, a reliable noninvasive glucose meter should overcome issues associated with low concentrations of glucose in the body, interpatient variations as well as physical changes at the probe site, effects of probe/body coupling, and repositioning errors. 10 A key question is whether a

32 4 universal multiple-person calibration can be created. Besides, the response to changing glucose levels should be fairly fast and the operating costs comparatively low. Methods of Noninvasive Glucose Sensing Various methodologies have been suggested for noninvasive glucose monitoring. Some of them are not truly noninvasive because they simply involve less painful ways to obtain a sample of ISF but still modify the structure of subject s skin. Examples include skin suction blisters between dermis and epidermis 20, use of micropores in the epidermis, 21 capillaries, 22 and iontophoresis. 23 In the ionophoresis method, a current is applied across the skin to collect glucose molecules transdermally. Skin permeability can also be increased for glucose collection with ultrasound as in the sonophoresis method. 24 The primary challenge for these methods is that the concentrations of glucose in the obtained sample are considerably lower than in blood and sensitivity is limited. Moreover, the risk of infection and inconvenience to the user are still very much present. The calibration procedure is complex with sweat on the skin being a major interference. Non-optical approaches to noninvasive glucose determinations include tissue impedance and temperature measurements. Both of these are indirect approaches because the measured parameters are not intrinsic properties of glucose molecule but are correlated with changes in glucose concentration. 19 The exact relationships between blood glucose concentrations and changes in dielectric properties of tissue and its temperature must be determined in the calibration step for these devices. Physiological conditions of the subject and body motion also introduce measurement errors. Optical noninvasive sensors probe the human body with nonionizing radiation. As photons interact with the tissue, various optical properties related to glucose concentration can be measured. 25 A sample containing glucose does not have to be collected or placed directly on the surface of the sensor. For some methods, optical fiber interfaces can be used. Indirect optical approaches involve measuring changes in tissue scattering, optical coherence tomography (OCT), and fluorescence-based methods

33 5 Selectivity of these devices can be questioned. For instance, changes in temperature, hydration, and protein content can alter the scattering properties of human tissue. Sensor repositioning errors also affect these optical measurements. To realize a fluorescencebased sensing scheme, a sample of body fluid or tear fluid has to come in contact with a substrate containing a fluorophore so this is not a strictly contact-less measurement. An interesting implementation reported in the literature involves a fluorophore-containing contact lens worn by a user. 29 When glucose from the tear fluid binds to the fluorophore, the fluorescence signal decreases as detected by a handheld meter. Obstacles of substrate saturation and reversibility are crucial in fluorescence methods of SMBG. Direct noninvasive optical methods achieve sensitivity by measuring an intrinsic property of the glucose molecule in the tissue matrix. Near infrared (NIR), mid infrared, Raman spectroscopy, photoacoustic spectroscopy, and polarimetry are currently studied. 19, To date, none of these methods is capable of providing a long-term multiuser calibration. The majority of the direct optical methods are centered on vibrational transitions in the glucose molecule. Unfortunately, the glucose vibrational spectrum does not provide distinctive analytical features independent of tissue or ISF matrix interference. The NIR absorptivity spectra of glucose and other bioanalytes are plotted in Figure I-1. Because overlap is present for all peaks of glucose absorption the broad spectral range of the glucose-containing tissue spectrum is statistically analyzed and the concentration information is extracted with the help of multivariate techniques. Collection of a large representative set of independent spectral measurements prior to glucose detection is essential for implementation of a robust multivariate calibration. This may not be practical for SMBG at home. For this reason, ideally the instrumentation has to be pre-calibrated with a universal calibration model that accounts for all or most of the interferences and matrix changes possible in the future. 10, 36 Vibrational spectroscopy has demonstrated sufficiently low standard errors of glucose prediction in simple aqueous solutions but a universal calibration model in human skin tissue still eludes researchers.

34 Absorbance, AU 6 x albumin lactate urea alanine glucose Wavenumber, cm Figure I-1. Glucose (solid line) and other biochemical analytes (dashed lines) absorptivity spectra collected from aqueous solutions at 37 C.

35 7 Noninvasive Glucose Sensing with NIR Spectroscopy Our group s approach to noninvasive glucose sensing involves transmitting a selected band of combination near infrared light ( cm -1 ; µm) through a fold of skin with the use of optical fibers or sapphire rods. The concentration of glucose is extracted from a multivariate analysis of the resulting spectrum. In this configuration, the analytical information is derived primarily from the ISF as the incident light propagates through the skin matrix. In practice, the resulting ISF glucose concentration is related to the corresponding blood glucose concentration through a calibration process that assumes a direct and constant relationship between glucose concentrations in blood and ISF. A careful analysis of these methods verifies that in vivo glucose contained in the interstitial fluid (ISF) of the skin is the origin of the chemical information used in the multivariate calibration models. For the animal model trial of the technology illustrated in Figure I-2, the skin on the upper shoulders of male and female rats was chosen for its similarity to human skin at the back of a hand. 37, 38 Prediction errors below 1 mm of glucose have been achieved for a stationary animal model. 37 However, in these experiments glucose quantification with the same precision is hampered by effects of fiber repositioning. Movement of the interface increases prediction errors more than 2.5-fold as reported for the partial least squares (PLS) model in Figure I-3 (A). In this plot, reference blood glucose concentrations are represented by the solid black line, the blue circles denote predicted glucose concentration in the calibration data set, the pink circles are predicted glucose values in the training data set, and the blue dashed line encloses the period of time when the interface was removed and put back onto tissue after collection of a single spectrum. When a net analyte signal (NAS) multivariate model is used to extract glucose data from the same dataset, similar offsets in predicted concentrations are seen. 39 Obviously, when the optical fiber interface is repositioned, parameters of path length and pressure can change. It was assumed that glucose molecules are distributed in the aqueous medium of

36 8 (A) (B) Figure I-2. The instrumental setup for the in vivo glucose measurement on an anesthesized rat. General view is shown in (A); detailed view of the measurement site is added in (B).

37 Regression Coefficient 9 (A) 2 (B) water collagen I keratin fat constant slope :00 12:00 13:00 14:00 15:00 16:00 17:00 18:00 19:00 Time, hrs Figure I-3. Effects of fiber repositioning on (A) PLS prediction of glucose concentrations and (B) amounts of water, collagen type I protein, fat, keratin protein, constant, and slope components. 38

38 10 the skin only. With this in mind, the effect of path length change should be negated if the concentrations are normalized for the water path length throughout. Such normalization was done but did not help to completely avoid the detrimental effects of fiber shift. Next, physical and biochemical variations in the tissue matrix itself between locations are hypothesized to cause this degradation of the model s performance. Figure I-3 (B) contains the regression coefficients representing thicknesses of major skin components, namely, water, collagen type I protein, fat, keratin, constant, and slope. These numbers were obtained for the duration of the in vivo experiment by regressing the skin absorption spectra with a linear six-component model of the standard pure component spectra. The region of the high scatter in the tissue component thicknesses coincides with the time of probe relocation. Based on the in vivo data, degradation of the calibration model for glucose appears connected to the spatial changes in the biochemical composition of the skin. This came as no surprise because the concentrations of major tissue components, water, collagen type I protein, keratin, and fat, are much greater than glucose concentrations. All of these components absorb in the combination region and their spectra partially overlap with glucose absorption. Then, the crucial question for noninvasive NIR transcutaneous measurements of glucose is if these spatial changes in one or more of these components present in tissue have the potential to cause the increase in standard error for glucose observed in vivo. Before this question can be answered, lateral heterogeneity of whole skin should be characterized in terms of the major sources of light attenuation in the combination NIR region at the spatial resolution superior to the size of the optical fiber (1.8 mm in diameter used in our initial animal experiments). Near infrared vibrational microspectroscopy is a great fit as the tool for characterization of tissue heterogeneity. A brief introduction to the origins of near infrared spectra of biological materials and the implications these have for the NIR vibrational microspectroscopy follows.

39 11 NIR Spectroscopy of Bioanalytes Near infrared or NIR radiation is uniquely suited to probe biochemical components. In fact, water in gelatin and vegetable oil were among the first samples studied by NIR in the first half of the 20th century. 40 As early as the 1960s, when NIR spectroscopy was just an add-on for UV/Vis spectrometers, Karl Norris and colleagues conducted a series of groundbreaking experiments with NIR spectroscopy investigating various topics in food chemistry, involving protein and moisture content in wheat and the composition of eggs. The progress in the field was slow because many researchers were confused by the region s weak and overlapping peaks. Method development in analytical NIR spectroscopy is heavily dependent on statistics and chemometrics and thus was not practical before the mainstream introduction of the microprocessor. However, with the introduction of powerful computers into research lab NIR spectroscopic methods have gained momentum. Currently, near infrared spectroscopy is widely used for measurement of biomolecules, both in vivo and ex vivo. Examples include extensive research for the measurements of glucose and its metabolites, 35, 35 urea, water, , 49, 50 amino acids, protein primary and secondary structures, simpler peptides, 49, 50 and lipids among other components. Tracking the concentrations of biological components can be utilized in general areas of process control, 58 biotechnology and bioengineering, 59 diagnostic medicine, 56 cosmetics, agriculture, pharmaceutical, and food industry. 60, 61 NIR spectroscopy has a crucial advantage as a non-destructive tool for probing of living tissues. When light hits the tissue several processes can happen: some part of the radiation will be reflected off the surface due to the refractive index mismatch between the sample and its environment while the remaining photons will penetrate the tissue and be either absorbed or scattered. However, near-infrared radiation is scattered to a larger degree within living tissue than mid-infrared. Thus NIR photons are able to travel deeper into the tissue providing path lengths on the order of several millimeters depending on the wavelength and sample characteristics. Longer path lengths ensure accumulation of more

40 12 analytical information into the measured spectrum. The lower absorbance of NIR light by biological tissues prevents excessive heating of the measurement site and enables transmission of some portion of this radiation that can be collected for analysis. Moreover, the near-infrared spectrum of water contains three optical windows where water absorptivity is very low. This allows for measurements in living tissue that is largely an aqueous medium. Origins of NIR Spectral Bands At the basis of vibrational spectroscopy is the concept that all intramolecular bonds vibrate with certain frequencies. Electromagnetic radiation corresponding to these vibrational frequencies is absorbed as the molecule is excited to higher energy level and the absorption spectrum can be quantified. Vibrational frequencies of the sample are then determined from the absorption spectrum. Overtones and combinations of these fundamental vibrations originate due to anharmonicity. Being forbidden transitions, the signal magnitude for overtones and combinations is generally from 10 to 1000 times lower than for fundamental vibrations. Overtone bands show up in the spectrum at the frequency twice or three times higher than the fundamental. Also, two or more vibrations can combine in a single combination band. Molar absorptivity at a particular wavelength is proportional to the product of the transition probability and the cross-section area of the molecule. Because absorption maxima in NIR spectroscopy result from forbidden transitions with probabilities lower than 0.01, NIR absorptivity bands, combinations and overtones, are classified as low intensity. 62 The near infrared spectrum spans wavelengths from to 4000 cm -1 (0.78 to 2.5 μm). The major part of the near-infrared absorption in biological matter corresponds to vibrational overtones and combinations of C-H, O-H, and N-H bonds. Specifically, spectral features between 5000 cm -1 and 4000 cm -1 come from combinations of stretching and bending vibrations of OH, CH and NH moieties. These chemical bonds are ubiquitous in biological molecules, which results in a large degree of overlap in the

41 13 combination near-infrared spectra of tissues. The overlap in the combination spectrum of glucose is shown in Figure I-1. Beer s law can be applied to mixtures of absorbers according to Equation I-1: A icil Equation I-1 i where i is the extinction coefficient of mixture component i, c i in the concentration of this component, and l is the sample path length. Beer s equation is used in our regression model for skin tissue with the assumption that the individual skin components do not interact. NIR Vibrational Microspectroscopy Vibrational microspectroscopy combines the advantages of optical spectroscopy with the high spatial resolution and imaging ability of light microscopy Its value for bio-analysis originates in the large amount of quantitative and qualitative information that can be obtained from spectra of individual small areas on the sample Vibrational techniques do not typically require introduction of an extraneous reagent, solvents, dyes or other complex sample preparation procedures. Depending on the region of the electromagnetic spectrum, the measurement can be performed noninvasively and nondestructively causing minimal disturbance to the biological system The result of data collection for vibrational microspectroscopic techniques is a large number of spectra coming from the measured pixels, micro-size locations on the sample, i. e. data hypercube. 72 The concept of a microspectroscopic data hypercube is illustrated in Figure I-4. Here, each pixel on the spectral map contains tissue absorption values at all resolution elements or wavelengths used for measurement. To make this wealth of data comprehensible for non-spectroscopists, spatially resolved colored images can be generated. There are multiple approaches to convert multidimensional spectral information into a colored image. Many researchers exploring the use of vibrational 73, 74 microspectroscopy for optical biopsy adopt a functional group imaging approach.

42 Figure I-4. Schematic representation of a hypercube of microspectroscopic data. 75 Dimensions X and Y define spatial location of each pixel and dimension λ contains absorbance values at the n-wavelengths (λ 1 λ n ) for the spectra collected at each pixel. 14

43 15 Commonly with this method, an area under the characteristic absorption peak or a ratio of peaks is calculated for each location on the sample. The resulting numerical values are then expressed as RGB colors in the image to define the localization of specific compounds. This method is extensively used in pharmaceutical characterizations and is well-suited to the mid-infrared spectral region that contains sharp distinctive peaks for the compounds of interest. However, for complex biological tissue samples the method does not provide sufficient differentiation between sample structures due to extensive spectral overlap. The same is true for near-infrared microspectroscopy where absorption features for most biological components are broad and highly overlapped. The strategy to overcome this complexity is to obtain the colored images of samples with the help of various statistical techniques, such as PCA, 76 multiple least squares (MLS), 77 artificial neural networks 78 and others. These techniques reduce the dimensionality of spectral data and ultimately produce markers that are correlated to various tissue structures within the sample rather than mapping the distribution of specific biomolecules. At the moment, the most common techniques in bio-microspectroscopy employ mid-infrared radiation. 77 The first reported spatially resolved spectra were measured in 1949 with a conventional microscope combined with an IR spectrometer. 79 The reason mid-ir is so attractive for vibrational imaging is the presence of distinctive peaks for most biological compounds in the region between 3 and 50 μm. In the literature, a significant number of studies have been reported that aim at detecting spectral differences between normal and malignant tissue with mid-ir microspectroscopy. 76 IR-maps of tissue are produced that complement histological images (i. e. optical biopsy) based on a unique spectroscopic fingerprint of each tissue type These microscpectroscopic images are oftentimes more informative than their histological counterparts. A histologist only has visible morphological changes available to her as a guide for the conclusions. Whereas with the optical biopsy, chemical information is embedded in the images generated from vibrational spectra. A major disadvantage of using mid-infrared

44 16 spectroscopy is that the radiation in this region is so strongly absorbed by most biological materials that path lengths larger than 15 μm are not practical. In addition, the sensitivity to water may complicate the analysis of most biological samples. Modern Raman microspectroscopic measurements offer complementarity to IR imaging with excellent spatial resolution, as high as 1 μm. 82, Unfortunately, for many biological components (for instance, skin lipids) fluorescence is a major interference in the Raman spectra. The furthermost advantage of NIR microspectroscopy is in its noninvasive and nondestructive nature. Challenges in interpreting broad and overlapping peaks of NIR spectrum can be largely overcome by employing pattern recognition techniques or factor analysis to evaluate spatially resolved NIR data. Among other applications, NIR 68, 69 microspectroscopy has demonstrated its great value in studies of skin hydration, distribution of active ingredients in an end pharmaceutical product, 71 protein kinetics in sol-gel matrices, 85, 86 and chemical composition of artifacts. 70 Because NIR absorptions are much weaker than at mid-infrared frequencies, issues of noise and optical throughput become central in instrument development. Commercially available IR microscopes are coupled with a Fourier transform (FT) spectrometer built around a Michelson interferometer, the schematic of which can be found in Figure I-5. In an FT instrument, spectra are collected in the form of an interferogram that is generated due to constructive and destructive interference of light inside the interferometer. Different wavelengths undergo interference at different rates after light is separated into two beams and travels between moving and stationary mirrors. The three main advantages of this FT design are: throughput (Jaquinot), multiplex (Fellgett), and the frequency precision (Connes). Advantages of FT spectrometry are essential for obtaining a high quality spectrum from a small location on the tissue either in mapping or imaging modality. Further discussion of measurement modalities in microspectroscopy can be found in Chapter II.

45 Figure I-5. Schematic diagram of the Michelson interferometer. 17

46 18 Overview of the Thesis Major advanced in characterization of rat and human skin tissue heterogeneity in relation to noninvasive glucose sensing are presented in this dissertation. First of all, microspectroscopic instrumentation system is optimized for data collection at NIR frequencies and methodology is developed for reliable collection of ex vivo skin tissue spectra in a transmission mode. The spectral data are analyzed with a linear regression model to yield coefficients representing amounts of the major skin components in each measured location. With these values, simulation of numerous skin spectra is achieved and effects of tissue heterogeneity on prediction of glucose concentration with optical fibers in vivo are modeled. Optimization of the microspectroscopic system is described in Chapter II. Advances include substitution of the detector inside the IR microscope, addition of optical filters and a customized sample holder. Spatial localization is determined for lipid, protein, and water molecules with simple samples. Detailed root-mean-square (RMS) noise evaluation is reported for the optimized instrumentation. In Chapter III, the system is employed to quantify heterogeneity of rat whole (i. e. dermis and epidermis) skin. RMS noise for the ex vivo spectra are tabulated and the database of regression coefficients for six components is compiled based on skin samples from 8 male and 4 female rats. A similar database for human skin samples from 2 male and 2 female donors is produced as explained in Chapter IV. In addition, spatial heterogeneity of individual dermis and epidermis as well as a double layer of epidermis tissue is characterized in this fourth chapter. Differences and similarities in NIR combination spectra of rat and human skin, of whole skin and layers are appreciated. Chapter V places the measured spatial heterogeneity in context for noninvasive glucose sensing with an optical fiber interface. Repositioning of the interface across a slice of tissue is modeled with two approaches: for 4 nonoverlapping adjacent quadrant and 36 overlapping random locations on the same map. Both approaches when combined

47 19 with the partial least squares (PLS) regression analysis demonstrated the effects of tissue heterogeneity on glucose prediction errors. The degree of calibration model degradation was assessed to be around the value observed in the in vivo experiment in the majority of cases. Occasionally, fiber movement between two particularly dissimilar locations on the skin was modeled. Then, the relative increase in standard error greatly surpassed the value observed in vivo. Chapter V clearly demonstrates that the effects of skin spatial heterogeneity must be accounted for to achieve universal calibration in transcutaneous NIR optical glucose sensing. In Chapter VI, efforts are presented to create a more accurate linear regression model for combination spectra of skin. Nonrandom features between 4400 and 4200 cm -1 in the spectral residuals of the linear fit are associated with the structure of lipid molecules present in the skin. Native fat standards of rat abdominal fat and human subcutaneous fat are obtained and used in the six component regression model. The substitution decreases the magnitude of the residuals and improves their shapes for two representative samples of rat and human skin. Thus a better fit of skin spectra is achieved for these samples.

48 20 CHAPTER II MICROSPECTROSCOPY FOR NEAR-INFRARED ANALYSIS OF BIOLOGICAL SAMPLES Introduction Measurement Modes in Vibrational Microspectroscopy Microscopes used for vibrational microspectroscopy typically offer the ability to measure in both reflectance and transmission modes. 65 Each modality offers unique advantages for different samples and spectral regions. When a sample on the stage of the microscope is illuminated, several processes can occur. Specular reflection is the process by which the light is reflected off of the topmost surface of the sample at the same angle as the incident angle. 87 Diffuse reflectance is the phenomenon that takes place when light penetrates into the sample and is scattered between internal surfaces, which causes it to exit the sample in a diffused form. Both specular and diffuse reflectance measurements have low sensitivity because the intensity of the radiation collected in the objective is typically from 3 to 10% of the incident value. 67 Oftentimes, a sample is ground and mixed with KBr powder to collect its diffuse reflectance spectrum. Attenuated total reflection (ATR) occurs when the totally internally reflected radiation beam traveling through an optically dense crystal with a high refractive index comes into contact with a sample. An evanescent wave is generated that extends beyond the ATR crystal. The portion of the evanescent wave is attenuated according to the absorption properties of the sample. ATR requires minimal sample preparation and produces vibrational spectra similar to those collected in the transmission mode. ATR is an attractive technique for the collection of microspectra from substances of limited solubility, adhesives, films, threads, beads, and small samples in general. 88 Unfortunately, the distance that light penetrates into the sample in the ATR measurement is very short, on the order of 0.5 to 3.5 μm. 89 Such low penetration depth greatly

49 21 decreases sensitivity of NIR spectroscopy of biological samples performed with ATR attachment. Reflection/absorbance measurements are done with a sample mounted on a reflective surface to allow the light penetrating the sample to reflect off the underlying substrate and travel back to the objective. In this dissertation, reflection/absorbance modality is called reflection. All of the reflection techniques are meaningful only if the key question is answered, that is spatial origin of the spectral information. Depth 90, 91 profiling or penetration depth simulation studies are done to address this question. Monte Carlo simulation can be used to calculate the length over which light attenuation 92, 93 occurs, i. e. path length. The transmission techniques simplify path length determination and offer the means to collect interpretable data free of contaminations from fixatives. Other research groups have successfully used diamond compression cells for collecting transmissionbased micro-spectroscopic FTIR images of complex multi-layered samples. 94 A relevant example is paint fragments, which consist of micron thick layers of organic and inorganic components, including proteinaceous compounds, fats, and resins. When a sample is sandwiched between two sapphire windows, surface defects are eliminated, which helps to decrease artifacts due to specular reflection. Additionally, the sample thickness becomes uniform. This translates into independence of measured chemical composition at a particular location on the sample thickness at that location. Working in transmission mode is complicated by several factors. First, the sample must pressed between two windows to realize a uniform path length and ensure sufficient light throughput. This pressure may displace some of its components as well as broaden details of a soft tissue structure. Pressing layered paint samples between diamond windows has been shown to distort the geometry of layers somewhat. 95 However, this procedure kept the samples intact and free of contaminants. 95 Alternative methods (preparation of KBr pellets, gluing, embedding techniques, etc.) either involve complete destruction of the sample structure or introduction of strong absorbers. 94

50 22 Mapping versus Imaging The term mapping in microspectroscopy is defined as a method of analysis, where the sample is moved stepwise across the area of interest and a single channel detector is used to collect a spectrum after each step. 63, 64 Depending on the area of the sample that is mapped, this procedure may be quite laborious and time consuming. It has the advantage of robustness and reproducibility because a single detector is used. Apertures are used to limit the incident radiation to the area of interest. If these apertures are small the flux of photons to the detector is low and the signal-to-noise ratio (SNR) is poor. To increase the SNR in microspectroscopic applications without compromising resolution, the synchrotron light source is needed. Synchrotron light offers unique brightness that allows smaller areas to be measured with acceptable SNR. 96, 97 The synchrotron source only fills from 10 to 20 μm. 98 For this reason its advantage is not realized in focal plane array (FPA) imaging applications where larger apertures are necessary. Imaging is the alternative method used to rapidly obtain spatially resolved vibrational spectra. An FPA detector is needed to collect spectra at multiple locations within the sample simultaneously. 99 Because each detector element in the array has to be illuminated at the same time the microscope cannot operate in a confocal arrangement and the spatial resolution suffers somewhat. For a static sample without any chemical changes during the course of the measurement, mapping and imaging will provide the same information. 73, 94, 100 Depending on the sample analyzed, mapping may be more convenient and cheaper. In this chapter, mapping is used to collect spectra for a limited number of locations across a sample matrix. Resolution in Vibrational Microspectroscopy Microspectroscopic spatial resolution is the ability to measure the spectrum from an area of the sample limited by apertures without significant interferences from surrounding regions. 65 The least resolvable separation (LRS) of a microscope is obtained as indicated in Equation II-1:

51 LRS N. A. Equation II-1 where λ is the wavelength of light and N.A. is the numerical aperture of the optics. 66 Because the LRS value is directly proportional to the wavelength of light, measurements done at NIR wavelengths have the potential for a better resolution than those collected with mid-ir radiation. The diffraction effect in vibrational microspectroscopy is revealed when the radiation that reaches the detector contains information from a larger physical area than expected. In the infrared microscope used herein, the optical geometry is mostly identical for visible and near-infrared light. The consequence of the diffraction effect is that the area of interest is more clearly seen through the objective and the CCD camera in the visible region of the spectrum than it can be measured over the NIR wavelengths. Control over the region that is illuminated in an infrared microscope is typically done with one or two external apertures. In this study, spectra were collected without the use of apertures to increase optical throughput. Also, relatively large stage steps were used to speed up the data acquisition. These factors limited the resolution of the measurement to the step-size of the motorized stage that is 480 µm and 360 µm in the X and Y dimensions, respectively. As a result, the question of resolution morphed into a question of the diameter of the measured area (i. e., captured area) at a single location on the map. It was also tested if two adjacent locations on a map actually overlap. Chapter Overview The opening goal of this chapter is to demonstrate how the modifications to the infrared microscope have made it well-suited for data collection in the combination nearinfrared ( cm -1 ) region. Detector type, detector oversaturation, noise values, and spectral region are all addressed as part of the optical system s optimization. Next, biologically relevant samples of rat nail, lipid/water suspension, and dried rat skin are microspectroscopically measured. The collected spectra are then fitted with a multi-component linear regression model over the cm -1 near-infrared region.

52 24 From previous research in the group, 19, 38, six standard components responsible for ~99% of skin tissue absorbance have been identified and used in compliance with Beer s law. Fractional regression coefficients for these standard components are computed at each measured location of the sample. The magnitude of these regression coefficients is expressed as the shade of color on the spatially resolved images of a sample. Overall, these preliminary results for rat nail, lipid/water suspension, and dried rat skin reveal sufficient suitability of the multi-component models for near-infrared micromapping of biomaterials. The six-component tissue model is extensively employed further in Chapter III for spatial characterization of multiple ex vivo rat skin samples. Experimental Procedures Instrumental Setup Data collection was performed with an IR Plan Advantage microscope (SpectraTech, Inc., Shelton, CT) coupled with a Nicolet Magna 560 Fourier transform infrared spectrometer (Nicolet Instrument Corp., Madison, WI). The visible light source, the sample stage, and the detector were all a part of the microscope. The spectrometer components that were used for data collection were the Michelson interferometer, the HeNe laser wavelength reference system, the 20 W tungsten halogen lamp used as a white light source, and the calcium fluoride, CaF 2, beam splitter. The microscope featured a binocular viewer (Olympus, Japan) for visible light mode and fixed compensation Cassegrain optics (0.58 NA, 160/0) that provided a 15-fold magnification of the sample. An RGB CCD vision camera module (model XC-711, Sony Electronics Inc., San Diego, CA) was attached on top of the view piece of the microscope to allow collection of images in the view mode. The schematic drawing of the system is presented in Figure II-1. Originally, the IR microscope came equipped with an integral 0.25 mm MCTA detector. However, this detector was found to be not ideally suited for measurements

53 25 image of aperture area CCD camera image upper aperture source sample stage detector interferometer field of view source light source Figure II-1. Schematic diagram of the IR Plan Advantage microscope. 65

54 26 between 4000 and 5000 cm -1 ( µm). Instead, a cryogenically cooled indium antimonide (InSb) detector (Thermo Electron Inc., Waltham, MA) was installed. The active area of the InSb detector was cm 2. InSb detectors are easily saturated. 104 To prevent detector saturation and isolate the spectral region of interest, a multi-layer optical interference K-band filter (Barr and Associates, Westford, MA) was attached to the detector s window. The microscope permitted the selection of external apertures but for the present study these apertures were omitted to maximize the signal. A 3 7 cm 2 motorized XY mapping stage (version 1.0, SpectraTech, Inc., Shelton, CT) was used to hold the sample and move it reproducibly in X and Y directions across the horizontal plane. The stage was controlled via the functions in the Atlμs software and offered ± 0.3 μm repeatability. Mediacapture 1.01 (ATI Technologies, Inc.) software was used to capture images from the camera. Each captured frame with this software measures 480 µm in the X direction and 360 µm in the Y direction. In other words, the dimensions of the image matched the step sizes for tissue maps. The instrument was connected to a Compaq Pentium PC running the Windows 95 operating system and was controlled by OMNIC Atlµs software (Version 1.1, Nicolet Instrument Corp., Madison, WI). Manual Focusing Procedure For all experiments throughout this dissertation, the microscope was manually focused in the visible mode as outlined in Figure II-2. Two 1.2 mm apertures were used for this focusing procedure: one aperture was installed before the objective, and the second one after the condenser. Both the condenser and the objective had the N.A. of First, the condenser and the motorized stage were moved together to align the focal point of the objective with the center of the sample holder filled with air. The position of the condenser was then adjusted to optimize visible light capture and match the visible images of the aperture produced by two apertures. The apertures were removed before the measurement of a sample to maximize the signal-to-noise ratio.

55 Figure II-2. Manual focusing procedure for the IR Plan Advantage microscope. The first step is shown in (A), when the condenser and the sample stage are moved in sync to reach the focal point of the objective; in (B) the height of the condenser is adjusted to ensure optimal focusing as shown in (C). 27

56 28 Sample Preparation In order to test the ability of skin tissue spectroscopic standards to model keratin distribution, four nails from a paw were harvested from a male rat. The nails were subsequently placed in a refrigerated desiccator for a month. A holder for the nail samples was manufactured from a 0.6 mm-thick aluminum plate. Holes, spaced 0.96 mm and 0.72 mm in X and Y directions, were precisely drilled to form a rectangular array. The diameter of each hole was 0.4 mm. These dimensions were used to match the step size of the motorized stage and to provide sufficient light throughput at each location. Water samples were composed of triple-distilled deionized water purified with a Milli-Q reagent water system (Millipore, Bedford, MA). Bovine total lipid liver extract was purchased from Avanti Polar Lipids Inc. and used to prepare an aqueous lipid suspension upon receipt. This suspension was prepared by heating a small amount of lipid to 70 C until melted and then shaking with several drops of water. An emulsion was formed. Upon transfer to the sample holder the emulsion cooled down and formed a suspension with flakes of lipid visible inside the aqueous matrix. A skin tissue sample was harvested from a euthanized retired breeder male Harlan Sprague rat. The details of the skin collection are given in Chapter III. In brief, the animal was sacrificed with the use of ketamine. Once the death was confirmed, the abdominal cavity of the animal was cut open and the rat was laid on its stomach to ensure the blood flowing away from the back to minimize contamination of the skin sample. Skin at the back of the upper back was closely shaved and a 1 1 cm 2 piece was excised. Both epidermis and dermis layers were collected in one piece but the subcutaneous layer and the membrane covering the muscle tissue were both left behind. Immediately after collection, the sample was positioned on top of a sapphire slide and dried in a desiccator maintained at 5 C for 5 days. After the first set of data was collected the sample was returned to the desiccator for the total of 29 days.

57 29 Spectral Data Collection For the measurements of air, the microscope was focused either in the middle of the clear opening of the motorized stage or in the middle of a sample cell, whichever was applicable. The resolution and the number of co-added scans matched those of the samples for which the air spectra were used as the reference. Air spectra were collected with 8 cm -1 resolution and 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, and 1024 interferograms were averaged. Water samples were measured inside the 1-mm thick rectangular cell (Starna Cells, Inc, Atascadero, CA) sealed with a Teflon stopper. Infrasil is a type of fused quartz that offers transmission values of more than 90% in the combination NIR. 105 When the detector response was characterized, 4 cm -1 resolution and 512 scans were used for both air and water spectra. All of the biological samples were positioned between two sapphire windows (Mettler Optics, Inc., Providence, RI). Desiccated rat nails were fitted into the holes at the outer edges of the rectangle on the aluminum sample plate. Then the plate was polished with sand paper to the uniform thickness of 580 μm. Polishing also helped to tightly fill the holes with the nail material. The prepared holder was positioned between two sapphire slides and spectra at the 9 locations on the aluminum plate were collected. Per spectrum, 512 scans were accumulated with a resolution of 4 cm -1. The holes at the outer edges of the aluminum plate yielded spectra of nail. The air spectra were collected in between the nail samples through the empty holes in the plate. A droplet of lipid/water emulsion was transferred with a pipette onto the sapphire slide between two 0.55 mm Teflon spacers. The second sapphire slide was carefully placed on top so that the mixture was in full contact with both slides and the thickness of the sample stayed equal to 0.55 mm. As the emulsion cooled, it formed a suspension with lipid flakes floating in water. The area with one of these flakes was chosen for mapping. For water and lipid samples the spectral resolution was 8 cm -1 and 256 interferograms were averaged per spectrum.

58 30 Dried rat skin slice was measured twice: after 5 and 29 days of desiccation. During both sessions, the sample was positioned between two sapphire windows. Dried skin spectra were collected then from 25 random locations on the sample as 128 co-added double-sided interferograms with a resolution of 8 cm -1. After the measurement of the skin desiccated for 5 days had been completed, the sample was returned to the desiccator in the refrigerator for 24 more days before the second set of data was collected. All of the microspectroscopic data in this thesis were obtained at ambient room conditions. The room temperature was recorded as 23 ± 1 C with a TES 1364 Humidity/Temperature meter around the site of the measurement. The average relative humidity (RH) value by the same instrument was 13 ± 2%. Spectral Noise Calculation Root-mean square (RMS) noise values on 100% lines of water spectra were obtained according to Equation II-2 below: RMS n i 1 ( x xˆ ) i n i 2 Equation II-2 where n is the number of wavelengths in the fitted region, x i is the absorbance value of the 100% line measured at the wavelength i, and xˆ i is the fitted absorbance value at the same wavelength. A 100% line was obtained by calculating the negative logarithm of the ratio between two back-to-back single beam spectra collected from the same sample. SNR values for air were obtained from the RMS noise values in cm -1 and cm -1 spectral regions. Equation II-3 demonstrates the inverse correlation between SNR and RMS noise values. SNR RMS n n i 1 ( x i xˆ ) i 2 Equation II-3 where n is the number of wavelengths in the fitted region, x i is the absorbance value measured at the wavelength i, and xˆ i is the fitted absorbance value. 106

59 31 Estimation of the Captured Area To estimate the measured area at each location within the sample, a piece of orange laboratory tape was attached to the sapphire slide and measured in transmission mode. The position of the edge of the tape was systematically varied. The procedure is illustrated in Figure II-3. In the beginning, the edge of the tape was aligned with the focal spot of the objective and the condenser and the first spectrum was collected. Next, the stage with the sapphire slide was gradually shifted to the left, 10 µm at a time. At each new location an image was captured with the CCD camera and a NIR spectrum was measured, accumulating 21 spectra spanning the distance of 200 µm total. Finally, spectra of a clear sapphire slide were collected in triplicate to be used as a blank. An orange laboratory tape spectrum exhibits several distinct absorption bands over the combination spectral range. To estimate the measured area, the baseline absorbance value at 4504 cm -1 was subtracted from the absorbance at 4331 cm -1 for each location. Once the difference in absorbances became less than the cut off value it was assumed that there was no more contribution to the spectra coming from the absorption of the tape material. The cut off absorbance value was calculated according to Equation I-4. A 3 RMS A cut _ off tape blank Equation II-4 where ΔA cut_off is the absorbance difference value used to define the captured area, RMS tape is the average root mean square noise value on the 100% lines of the tape spectra collected 650 μm inside the tape, and ΔA blank is the absorbance difference for the air blank collected through a clear sapphire slide. The radius of the captured area was determined as the longest distance from the edge of the tape to the focal point of the microscope for when the absorbance difference, ΔA cut_off, was still less than the cut off value. Standard Components Absorbance spectra of the six standard components used for modeling are plotted in Figure II-4. These include water, collagen type I protein, fat, keratin protein, constant

60 Figure II-3. Estimation of the captured area with selected CCD images of the orange laboratory tape shown at the distance of 0 μm (1), 40 μm (2), and 160 μm (3). 32

61 Absorbance, AU collageni * 10 water fat keratin constant slope Figure II-4. Combination near-infrared spectra for water, collagen type I protein, fat, keratin protein, constant, and slope Wavenumber, cm

62 34 offset, and sloping baseline. Data for the first four chemical standards were previously collected in a transmission geometry at room temperature with a MIDAC FT-IR spectrometer. The absorbance spectrum of water (A water ) was obtained from a sealed 1-mm thick transmission cell filled with triple-distilled deionized water purified with a Milli-Q reagent water system (Millipore, Bedford, MA). The collagen spectrum (A collagen ) was collected from a 1-mm thick pressed pellet composed of type-i collagen protein dispersed in a matrix of potassium bromide. The keratin spectrum (A keratin ) was obtained by transmitting the near infrared radiation through the free edge of a human fingernail. Nail is used as the keratin standard because it has been reported to be composed primarily of keratin protein with small amounts of water and lipids The fat spectrum (A fat ) was collected from a 1-mm thick sample of bovine fat as described by this group elsewhere. 101, 111 The constant and sloping baseline terms were added to account for spectral variations associated with scattering and changes in temperature. Linear Regression Models Each absorbance spectrum was fitted with a multi-component linear regression model over the cm -1 spectral range. Least squares regressions for each type of sample incorporated only those components that were actually present in that sample. Equations II-4 through II-6 were used for modeling rat nail samples, lipid/water mixture, and dried rat skin samples respectively: A nail_i ( Akeratin keratin) i ( Aconstant constant) i ( Aslope slope) i i Equation II-4 A A ) ( A ) ( A ) lipid/water_i ( water water i fat fat i constant constant i ( Aslope slope i ) i Equation II-5 A A ) ( A ) ( A ) skin_i ( water water i collagen_i collagen_i i fat fat ) ( Akeratin keratin) i ( Aconstant constant) i ( Aslope slope i i i Equation II-6

63 35 where Alipid/wate r_i, Anail_i, and Askin_iare the measured absorbance spectrum at location i of the lipid/water suspension, rat nail plate, and dried rat skin, respectively; the standard absorbance spectra of four major chemical components as well as the constant offset and slope terms are represented in the equations by A A, A, A, A, water, collagen_i fat keratin constant and A slope. Finally, water, collagen_i, fat, keratin, constant, and slopeare the regression coefficients from the least-squares fit for the corresponding standard absorbance spectra. The spectral residual at location i is expressed as i. The fitting procedure consisted of multiple linear fit iterations performed until the minimal value of the spectral residual i was obtained. The resulting six regression coefficients represent the relative amount of each component (water, collagen, keratin, fat, slope and offset) for each location on the sample. Results and Discussion Optimization of Spectral Quality Reflection measurements with a sample placed on the gold-plated mirror produced reasonable SNR values for water and air. However, attempts to measure rat nail or dried rat skin samples in this arrangement were not successful. The spectra of the skin were especially noisy but, more importantly, were dominated by scattering. Moreover, even with a relatively transparent water droplet, the estimation of the path length was complicated by scattering background. Reflectance setup was abandoned for the rest of this project in favor of transmission measurements. Transmission was consistent with the method used for collection of the modeling standards as well as the optical fiber setup for 37, 38, 112 noninvasive glucose measurement in rats. Detector Type Substitution The detector is one of the key components of the microspectroscopic system responsible for spectral quality in NIR spectroscopy. Random error independent of transmittance value usually arises from Johnson noise in the thermal detector. 113 It becomes especially important when the flux of photons is low because Johnson noise

64 36 occurs in thermal detectors even in the absence of radiation. 114 In Figure II-5 (A) air single-beam spectra are plotted. The spectrum in red was collected with an MCT/A detector and the spectrum in blue with an InSb detector. Both detectors had been cryogenically cooled before the measurement. Single-beam spectra for a 1-mm sample of water are provided in Figure II-5 (B). Consistently with the air spectra, the data for water collected with the MCT/A detector is plotted in red, with InSb detector in blue. Immediately, it is apparent that there is a substantial increase in the signal magnitude for both air and water samples when using the InSb detector. The signal improvement was expected from the technical characteristics of the two detectors. Originally, the IR Plan Advantage microscope came equipped with an MCT/A liquid nitrogen cooled detector optimized for use in the mid-infrared spectral region. This detector type does not offer satisfactory sensitivity over the cm -1 range. 104 Compared to the MCT/A, the InSb detector is a better match for the combination region. In fact, the increase in the signal after the detector substitution was measured to be between 4.0 and 4.5 times based on the peak values for the interferograms. In Figure II-5 the same K band pass filter was attached to both detectors to eliminate contribution from the wavelengths outside the combination region. Limiting the detected radiation to the region of interest prevented detector oversaturation and refined the spectral shape. An example of the detector saturation effect on water absorbance in the absence of the band pass filter is illustrated in Figure II-6. The absorbance spectra for a 1-mm thick aqueous sample are plotted in different colors: blue and red for the InSb detector and MCT/A detector respectively, both collected with the addition of the K-band filter. The absorbance spectrum plotted in green was measured with the MCT/A detector without the interference filter present. For this measurement, the effect of detector saturation clearly manifests itself in the shape of the absorbance spectrum.

65 Signal Intensity, volt Signal Intensity, volt (A) MCT/A InSb Wavenumber, cm (B) MCT/A InSb Wavenumber, cm Figure II-5. Combination near-infrared single beam spectra for (A) air and (B) water inside the 1-mm thick Infrasil cell. Data for MCT/A detector are plotted in red and for InSb detector in blue.

66 Absorbance, AU InSb MCT/A MCT/A no K band Wavenumber, cm Figure II-6. Combination near-infrared absorbance spectra for water inside the 1-mm thick Infrasil cell. Spectra collected with a K band-pass filter is plotted in red for MCT/A detector and in blue for InSb detector. The spectrum plotted in green is for the MCT/A detector without the K band filter.

67 39 First of all, absorbance values are slightly lower across most of the combination region. The effect is especially noticeable at the wavenumbers larger than 4700 cm -1. Saturation impacts the detector response at those larger wavenumbers in a way that makes water absorbance appear to level off instead of sharply rising at that edge of the combination region. Both MCT/A and InSb detectors, when equipped with the interference filter, produced absorption spectra with the correct shape for water as shown in the red and blue spectra presented in Figure II-6. Selection of Wavenumber Range Root mean square (RMS) noise on 100% lines of water spectra is a very important predictor of the overall microscope performance when used to measure biological samples. Most living tissue has water as a primary component, with the possible exception of hair, nails, and stratum corneum. 108 Noise in water spectra was assessed across ten 100 cm -1 intervals between 4000 and 5000 cm -1 for the MCT/A and InSb detectors equipped with the K band filter. The results are summarized in Table II-1. Water absorption is at its minimum between 4600 and 4400 cm -1. In this region, the RMS noise for spectra collected with the InSb detector is 2.2 times lower than the RMS noise collected with the MCT/A detector. The difference between these two detectors is more prominent around 5000 cm -1. At those wavenumbers, the RMS noise obtained with the MCT/A detector is about 10 times higher than the values obtained with the InSb detector. These differences in performance of the two detectors provide an additional basis for the substitution of the original MCT/A detector within the IR Plan Advantage microscope with the InSb detector for measurements over the combination wavelengths. Lowering the degree of random noise in the microspectroscopic data is crucial for successful modeling of biological samples, such as skin tissue. Various data smoothing algorithms, like Savitzky-Golay filtering, can be effective for decreasing spectral noise. 113, 115 However, when the noise level is comparable in magnitude to the spectral

68 40 Table II-1. RMS noise values across the combination region ( cm -1 ) for 1 mm thick water samples collected with a K band filter. Spectral Region, cm -1 RMS Noise, AU InSb detector MCT/A detector ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± 0.3 Note: Values are listed as mean ± 1 standard deviation for 3 observations.

69 41 features of the chemical compounds in the sample, noise-reducing algorithms may lead to errors. RMS noise values larger than 0.01 AU are approaching the absorbance level of the finer spectral feature in the major biochemical component used for tissue modeling in this dissertation. For example, features similar in magnitude to those at 4320 cm -1 for collagen type I protein and 4260 cm -1 for fat in Figure II-4 would be lost or deformed if present in the spectral region with the noise level higher than 0.01 μau. On the other hand, it is impractical to include the regions with the large RMS noise levels into the linear model without spectral smoothing. The modeling algorithm will try to find the best fit for the random values caused by the spectral noise which may lower the quality of the fit across this region of absorption of the chemical constituents in the sample. The adopted solution for this issue is to exclude the regions in the combination spectrum with the RMS noise higher than 0.01 AU. For the MCT/A detector, this means limiting the modeled region to the wavenumbers between 4800 and 4200 cm -1. For the InSb detector, the acceptable region is wider, from 4900 and 4200 cm -1. Notably, these wavelengths were used for noninvasive glucose detection in living rats. 38, 112, 116 In all of the experiments that follow, the data analysis is performed for the cm -1 region. SNR Dependence on the Number of Co-Added Interferograms For a near-ftir spectrometer, the signal-to-noise ratio (SNR) is linearly proportional to the square root of the number of co-added interferogram scans. 113 Essentially, random noise in the spectrum is decreased by averaging multiple interferograms. Such a linear trend is demonstrated in Figure II-7 for the air spectra collected with the microscope. The corresponding numerical values are listed in Table II- 2. In Figure II-7, the SNR data points marked with blue circles were obtained from the spectral region between 4400 and 4500 cm -1. The data for the cm -1 region is plotted in red. In Figure II-6, the error bars correspond to the standard deviation of the SNR values derived from different pairs of triplicate single beam spectra. The trend lines for

70 42 Signal-to-Noise Ratio cm cm -1 Linear Fit SQRT(number of scans) Figure II-7. Signal-to-noise ratio (SNR) dependence on the number of co-added interferograms. The error bars correspond to the standard deviation of the SNR values derived from different pairs of triplicate single beam spectra.

71 43 Table II-2. Linear dependence of SNR on the number of co-added interferograms. Parameter Coefficient Std Error R 2 value cm -1 Slope Intercept 34* cm -1 Slope Intercept 34* 543 Note: * Intercept value is less than its standard error

72 44 the best linear fit are included. For both wavenumber regions, the linear dependence of the SNR on the square root of the number of co-added scans is present. This is indicated by the coefficient of determination R 2 values higher than 0.99 in Table II-2. For a given number of the interferogram scans, the magnitude of SNR values for the microspectroscopic system is about 10 times lower than for the typical near-ftir spectrometer. 117 The experimental significance of this is that a large number of interferogram scans must be accumulated to lower the noise to an acceptable value. Using the existing microspectroscopic setup for analytes with weak absorptivity in the combination region (for instance, glucose) will be impractical, especially if the sample matrix has a high absorbance. On the other hand, such SNR level is sufficient for modeling major biochemical components of tissue (i. e., water, protein, fat). Examples of such modeling are described later in this chapter. Diameter of the Captured Area The use of external apertures was found to be not feasible for biological samples because of the reduced light throughput and decreased signal-to-noise ratio. Hence, it was necessary to estimate the area that the signal is coming from for a given location on a sample. The approximation is that the spectral information collected through the condenser is coming from a cylinder with a diameter defined here as the captured area of the system. The diameter of the captured area was estimated by partially blocking the incident beam with a piece of orange laboratory tape at different positions along the beam cross-section as described in the Experimental section. In Figure II-8 (A) the spectrum of the laboratory tape is provided. In Figure II-8 (B) the absorbance differences are plotted against the distance between the focal point and the edge of the tape. The green dash-dot line in the plot indicates the radius of the measured spot. The distance from the focal point at which this happens was measured as 160 µm. This is the experimental estimate of the radius of the captured area for the measurement. There are limitations to this method of determining the captured area. Even

73 Normalized Absorbance at 4331 cm -1, AU Absorbance, AU (A) cm -1 peak cm -1 baseline Wavenumber, cm (B) cut off point (blank + 3*RMS noise) Distance, m Figure II-8. Determination of the microscope measured area. Absorbance spectrum of the orange laboratory tape is shown in (A); in (B) the normalized absorbance values at 4331 cm -1 are plotted against the distance between the edge of the tape and the focal point. Green dash-dot line defines the radius of the measured area at 160 μm.

74 46 though the laboratory tape was a lot thinner than skin tissue it provided light attenuation of the same order of magnitude as tissue samples. However, it is safe to approximate that at every point of a microspectroscopic map spectral information comes from the cylinder with a diameter of 320 µm. The important finding of this experiment was that the measured area is close to the Y step size of the sample stage, 360 μm. This is important because the step size smaller than the captured area (i. e. redundant aperturing ) is optimal for microspectroscopic imaging as it minimizes the loss of information. 118 In the microspectroscopic setup, the resolution of the spectral map is effectively determined by the step size of the sample stage that is slightly larger than the captured area. Even though redundant aperturing is not realized, this lower resolution of the system is acceptable for the present application. First of all, the data collected from the slices of tissue ex vivo would be used to model the relocation of optical interface in vivo. The diameter of the contact area of the interface is typically at least 4 times larger than the step size of the stage. Tissue spectra across several locations on the skin map will be averaged to simulate a given location of the optical fiber. With this averaging in mind, the lower resolution of several hundred micrometers will be sufficient for successful tissue modeling. Secondly, measuring fewer locations per a given area of the map will decrease data collection time mitigating effects of tissue dehydration and heating. With a lower number of the mapped points it will become possible to improve spectral quality: more interferogram scans will be accumulated in the same measurement time. Testing of the Spectral Standards with Biological Samples Experiments described further in this chapter constitute the preparatory work for rat and human skin tissue microspectroscopic mapping. Two crucial aspects are examined: the suitability of the microscope for collecting the near-infrared spectra of the biological samples (i. e., rat nail, lipid-water suspension, and dried rat skin tissue) and the ability to model these absorbance spectra with a combination of the chemical standards.

75 47 Six standard components were experimentally determined to model 98% to 99% of information in absorbance spectra of rat skin for noninvasive FT-IR study on rats. 38 Four chemical components (water, collagen type I protein, fat, and keratin protein) were selected because of their universally high abundance in whole skin tissue and subcutaneous layer. The non-chemical terms, constant offset and sloping baseline, were necessary to complete the spectral model in the combination region. The constant term accounts for the stray light contribution that uniformly lowers absorbance values. Scattering that happens outside the sample and reflection off the surfaces of mirrors and filters of the microspectroscopic system manifest themselves as stray light. 119 This radiation is detected by the microscope but has not passed through the sample and does not contain analyte information. 120 Because the four standard spectra collected for the chemical components have a minimal amount of scattering in them it is important to include the constant term into the linear regression. The slope term represents both the contribution for the temperature-dependent scattering and absorbance changes. Extensive research has been documented in the literature in regards to the temperature dependence of nearinfrared spectra of aqueous samples , In brief, the vibrational peaks for liquid water at 3800 cm -1 and 5200 cm -1 experience a shift to higher wavelengths (lower wave numbers) with an increase in the temperature of the sample. The magnitude of the absorbance for both of these peaks also changes somewhat, with water absorbance decreasing with an increase in temperature. 45, 124 The negative slope, plotted in orange in Figure II-4, approximates the impact of the changing temperature on the change in spectral shape of water. 124 Because most biological tissues contain a substantial amount of water, integration of the slope term into the linear regression model was imperative. Specific chemical components were chosen based on the content of each individual sample. When a non-existent component was added into the regression, for example, when the term for keratin was added into the model for lipid/water mixture, the

76 48 coefficients produced had negative values. Constant and slope terms were routinely used in modeling of each of the mapped biosamples. Rat Nail Spectra Dried rat nail samples were used to test the modeling performance of the keratin spectroscopic standard. Four nail were placed inside the holes of the aluminum plate and microspectroscopically measured as shown in Figure II-9. The schematic diagram of the aluminum plate is shown in Figure II-9 (A). The dimensions of the sample holder were optimized according to the step size of the motorized stage. The diameter of the openings, 400 μm, was chosen to be larger than the estimated captured area to minimize the loss of analytical information in the nail spectra. Fortunately, rat nails had sufficient diameter to snugly fit into the openings filling most of the empty space. The openings filled with the nail samples are colored orange, with air white. Combination absorbance spectra in Figure II-9 (B) were taken at the nine locations across the aluminum holder. There is a notable difference between the nail and air absorbance spectra. Rat nail spectra reveal broad absorption features around 4865, 4600, 4350, and 4250 cm -1 characteristic for protein samples. 52, 54, 56, 125 Air absorbance profiles demonstrate uniform light attenuation due to scattering properties of the sample holder and sapphire windows, as well as the K band pass filter. The simpler three-component linear regression model was successfully tested with these absorbance spectra. Included in this model were the constant and the slope terms in addition to the only chemical standard of keratin protein as expressed in Equation II-4. Raw absorbance spectra were fitted with the model and the fractional coefficients for keratin protein, constant, and slope were computed. Spatial distribution map for keratin coefficient is presented in Figure II-9 (C). Each measured location is marked by a green open circle in this map. The lighter shade of color corresponds to the higher keratin content and the darker shade to the lower content. The square shape of the shading is an artifact of the plotting function in MATLAB.

77 Absorbance, AU 49 (A) (B) μm 720 μm nail 1 air nail 2 air air air nail 3 air nail μm Wavenumber, cm (C) Figure II-9. Nail samples modeling with the spectral standards. The schematic diagram of the aluminum plate in (A) shows the dimensions of the sample holder. The opening filled with nail are colored orange, with air white. Absorbance spectra in (B) were taken at all of the 9 locations. Distribution map for keratin standard is presented in (C): each measured location is noted by a green empty circle in this map. The lighter shade of color corresponds to the higher keratin content and the darker shade to the lower content.

78 50 As expected, the corner locations filled with the rat nails consistently yield the highest coefficient for keratin with the average value of 0.69 with the standard deviation of This value is larger than the thickness of the sample of 0.58 mm but the chemical standard for keratin was collected from a human nail. The exact thickness of the keratin standard is not known but it is safe to assume that it is less than 1 mm. A literature source lists values from about 300 μm to 600 μm. 126 When these values are multiplied by the average regression coefficient value for keratin, 0.21 mm and 0.41 mm are obtained. Both of these are lower than the thickness of the aluminum sample holder. This fact suggests that the density of the rat nail material in the openings of the holder was different from the density of the human nail standard. When a nail is dried it tends to splinter. 107, 109, 110 This splitting of the nail together with incomplete packing of the holes most probably caused the discrepancy in the keratin values. The openings left empty produce negligible keratin coefficients with the mean value of and the standard deviation of The standard deviation larger than the mean value here indicates the presence of a number of negative coefficient values, which is an indication of the complete absence of protein absorption at those locations. The differences in regression coefficients for constant and slope between locations are very low. The maps for the constant and slope terms are not included in Figure II-9 because they are not illustrative of the sample. Lipid Suspension Spectra Lipid-water suspension was prepared and microspectroscopically measured as described in the Experimental section. Mixtures of lipid and water are used by researchers to model biological tissue in near infrared spectroscopic studies. 101, 127 Tissue phantoms consisting of these two components have been shown to mimic near-infrared spectra of skin reasonably well. 101 Here, bovine lipid extract suspended in water is spatially characterized to examine the suitability of the fat spectral standard for determining areas high in lipid molecules in skin tissue samples.

79 51 The visible images of the suspension, absorbance spectra, and the computed regression coefficient maps for fat and water are shown in Figure II-10. Specifically, in Figure II-10 (A), the visible images of a part of the suspension collected with the CCD camera are assembled to demonstrate the overall mapped area. Slight movements of the lipid flakes in aqueous medium were unavoidable when the sample was repositioned under the objective of the microscope as seen in the slight shifts between fragments in the visible image. Each of these 15 rectangular photos measures 480-by-360 μm. The dimensions of the mapped area are 1920-by-720 μm. The darker regions correspond to the position of the lipid flakes and the lighter areas are mostly occupied by water. In Figure II-10 (B) all of the 15 locations are represented by near-infrared absorbance spectra collected with the focal point in the middle of each rectangle. Among these spectra, the representative spectra for water and lipid were collected at locations 7 and 12, respectively. Absorbance spectrum at location 7 has the shape of water absorbance spectrum with the minimum at 4530 cm -1. At location 12, lipid peaks at 4,330 and 4,255 cm -1 are prominent. After linear regression was applied to the spectra according to the Equation II-5 the spatial distribution maps in Figure II-10 (C) and (D) were created for water and fat coefficients, respectively. Each measured location is denoted by a black circle in these maps. The lighter shade of color corresponds to the higher content and the darker shade to the lower content of the component. The sums of water and fat coefficients at each of the mapped locations are very close to the thickness of the Teflon spacer, 0.55 mm. The mean value of the sum of lipid and water coefficients is with the standard deviation of For both water and fat standards, these fractional coefficients represent the thickness (mm) of lipid and aqueous layers in the beam path. The reason for this is that these particular chemical standards were obtained from 1 mm thick water and bovine fat samples. Small difference between the sum of chemical coefficients and the spacer thickness can be associated with the unmodeled spectral residuals, ε.

80 Absorbance, AU 52 (A) (B) location location Wavenumber, cm (C) (D) Figure II-10. Lipid suspension modeling with the spectral standards. The images in (A) were collected in the visible mode with the CCD camera. The darker regions show the lipid flakes and the lighter regions water. Locations 7 (mostly water) and 12 (lipid particles) are emphasized for comparison. Absorbance spectra in (B) were taken at all of the 15 locations. Distribution maps are presented for (C) water and (D) fat with each measured location denoted. For both components, the lighter shade of color corresponds to the higher content and the darker shade to the lower content.

81 53 Results of spatial distribution of fat and water are very encouraging because they closely match the visible images of the mapped area. The highest amount of fat is found at location 12, corresponding to the position of the lipid flake in the visible image. On the other hand, locations 5, 6, and 15 have negligible fat content and the highest water content. They appear clear in the CCD image because there are no lipid fragments present at those locations. Comparison of visible image to the spatial maps of chemical components obtained with the linear regression fit demonstrates the ability to locate areas high in lipid content in an aqueous matrix with near-ir microspectroscopy. Dried Rat Skin Upon successful testing of the microspectroscopic transmission set up with the simple lipid/water mixture and the rat nail samples, measurements of whole rat skin were attempted. When the whole skin was positioned on the sapphire slide immediately after having been thawed without compression, its absorption and scattering properties dramatically reduced the amount of light that reached the InSb detector. In an IR microscope, it is primarily the specimen itself that limits the optical throughput, not the aperture or the detector. In contrast to an FTIR spectrometer, the throughput (Jacquinot) advantage is limited by the need to analyze a small area of the sample. 65 To circumvent this limitation, measurements of the whole skin were performed after the sample had been dried in the desiccator at 1.3 C for 5 and 29 days. Because prolonged drying and refrigeration possibly compromised sample integrity, tissue mapping was deemed unreliable and was not performed. Instead, too sets of absorbance spectra were obtained from random locations on the sample to track overall dehydration. Water is a major absorber in the near-ir region. Transmission values of the skin were increased after drying which enabled data collection. Absorbance spectra for 25 random locations are shown in Figure II-11 (A) for 5 days and Figure II-10 (B) for 29 days of desiccation. Because the sample was removed from the microscope stage and placed in the desiccator it was impossible to measure exactly the same locations after 5

82 Absorbance, AU Absorbance, AU 54 (A) Wavenumber, cm (B) Wavenumber, cm Figure II-11. Absorbance spectra for a sample of male rat skin dried in a desiccator at 1.3 C for (A) 5 days and (B) 29 days.

83 55 and 29 days of drying. The plots illustrate the overall shape and magnitude of dried tissue absorbance. The spread in these data is due to variations in chemical composition and thickness between locations. Major absorbance features present in all of the spectra are those of proteins, 4600 cm -1, 4400 cm -1, and 4,890 cm -1, and lipid molecules, 4,330 and 4,255 cm -1. Conversely to the spectra in Figure II-10, there is no prominent water background in either set of the dried skin spectra. Up to 50% of light attenuation is due to scattering in these samples. This is revealed by the prominent linear background in the spectra and the large regression coefficients for the constant term in Table II-5. This table lists results of six component linear regression as expressed in Equation II-6 for the initial drying of the rat skin sample (5 days) and the subsequent drying (29 days). For each measurement session, coefficients obtained from 25 locations across the sample were averaged and the standard deviations tabulated. Changes after additional drying in the average content of collagen type I protein, fat, keratin protein, and the magnitude of the constant term are all within the standard deviation values for these components. Conversely, changes in coefficients for water and slope fall outside of their standard deviation values. These values are highlighted in bold font in Table II-5. Change in water content is in agreement with the additional water loss between 5 and 29 days of drying. Because the slope term largely reflects the shifts in water absorption due to temperature changes, the magnitude of the slope coefficient is also expected to change with the decrease in water content. No characterization of spatial heterogeneity was attempted for the skin sample at this point. The tissue appeared very dense and the native internal structure had probably been largely lost because of prolonged drying. In Table II-5, regression coefficients for the 25 locations on the skin are treated as representing a single sample. The testing of the six component model with a sample of dried rat skin demonstrated sensitivity of the model to the changes in water content in the presence of other chemical constituents. The ability to quantify water effectively in the skin samples

84 56 Table II-3. Regression coefficients summary for dried rat skin sample measured after 5 and 29 days of desiccation. Regression coefficients, mean ± standard deviation* Water a Collagen Fat Keratin Constant Slope a Type I days ± ± 0.04 ± ± 0.02 ± days ± ± 0.04 ± ± 0.02 ± 0.07 Note: * Mean and standard deviation values calculated for 25 separate locations on the sample; a the change in the coefficients of these components falls outside of their standard deviation values. ± ± 0.004

85 57 is crucial for accurate simulation of tissue spectra. The assumption used for the multivariate calibration methods is that glucose is dissolved in the aqueous medium in the skin and is not present in the space occupied by lipid and protein compounds. What this implies is that to build a robust calibration model one has to estimate water path length for each given location of the measurement. Fitting skin absorbance spectra with the six component model will accomplish this goal. Conclusions This chapter summarizes preparatory work that enables spatially resolved microspectroscopic transmission measurements of rat whole skin. To begin with, optical components of the system were modified to ensure optimal RMS noise values and to avoid the effects of detector saturation. Linearity of SNR dependence of the square root of the number of co-added interferograms has been demonstrated for the IR microscope. Next, a linear regression strategy was tested with biosamples of dried rat nail, lipid and water mixture, and dried rat whole skin sample. The additive nature of Beer s law provided regression coefficients for each standard constituent as it was assumed that the components of the sample do not interact with each other. 128 Overall, the results demonstrate the ability to localize major biological components of water, proteins, and fat in simple matrices. This ability will be employed to spatially characterize skin tissue in subsequent chapters.

86 58 CHAPTER III CHARACTERIZATION OF SPATIAL HETEROGENEITY OF RAT SKIN EX VIVO WITH NEAR-INFRARED MICROSPECTROSCOPY Introduction Rat Skin Tissue Heterogeneity in Relation to Near Infrared Spectroscopy Complexity of the skin matrix imposes constraints on non-invasive near-infrared measurements of glucose. As explained in Chapter I, accurate predictions of glucose concentration in vivo are possible when the fiber-optic interface remains stationary during the collection of rat skin spectra. 37, 38 When the interface is removed and repositioned before each spectrum, the predicted glucose concentrations are highly scattered. 38 A major source of spectral variance is the tissue sample itself. The composition of the tissue matrix and the physical distribution of the principal chemical constituents have a major impact on the spectra collected during a noninvasive experiment. 129 Many biomolecules abundant in the skin have higher absorptivities and/or higher concentrations than the glucose molecule. The relative amount of each chemical constituent within the path of propagating photons and the heterogeneous distribution of these chemicals heavily impact the absorption and scattering of photons that compose each noninvasive spectrum. Furthermore, the spectroscopic and scattering properties of living tissue are strongly affected by temperature and pressure, 130, 131 thereby producing additional sources of tissue-related spectral variance. The crucial issue addressed in this chapter is the extent of spatial heterogeneity of the chemical composition of rat skin and how this heterogeneity impacts the spectral variance of noninvasive near-infrared transmission spectra. Composition and Structure of Rat Skin Tissue Skin is the largest organ of the body that serves as a protective and sensing barrier between the organism and the environment. The heterogeneous stratified structure of skin can be roughly divided into three major layers: the outer surface layer of epidermis, the

87 59 underlying dermis, and the subcutaneous tissue. 132 These layers can be further separated into a number of finer sub-layers. Two examples of sub-layers are the stratum corneum (the cornified outer-most layer of the epidermis) and the papillary dermis (the dermalepidermal junction). Notably, each of these layers has different optical properties including absorption, scattering, and refractive indices. 91 For the purposes of nearinfrared spectroscopy, the complexity of the skin matrix can be characterized reasonably well with six major components: water, collagen type I protein, fat, keratin protein, constant offset and sloping baseline. These six components are responsible for the bulk of the skin absorption spectrum over the combination region of the near infrared spectrum. Epidermis is a tightly bound outer layer of the skin which is rich in keratin protein. Its thickness in rats ranges from roughly 14 to 80 μm depending on the location on the body. 133, 134 This layer is especially thick over the regions of the body subjected to high friction forces, such as palms. Composition of epidermis is highly dependent on diet and environmental factors. 133, 133, , 137 This dependence is not surprising as it takes just under a month for epidermis to completely renew itself. 132 The predominant cells in the epidermis are keratinocytes embedded in a lipid and protein rich extracellular matrix. As keratinocytes mature and move upward to join the stratum corneum (SC), keratin filaments within the cells become denser. 138 In case of injury, keratinocytes are also capable of horizontal locomotion with the help of actin protein microfilaments. 132 Another important cell in the epidermis is the melanocyte. Its primary function is to produce melanins, pigments that protect skin from ionizing effects of radiation and determine skin color. 139, 140 Even though melanins are generally considered to be important skin chromophores in skin UV-Vis and near infrared spectroscopic studies, 69, 141 variations in melanins content were previously shown to have little if any effect on glucose prediction in the combination region. 37 Moreover, these compounds have limited 139, 142 solubility that complicates the direct collection of high quality spectra. Consequently, melanins are omitted from the skin tissue model presented in this

88 60 dissertation. Keratin protein and fat components are used to account for the absorption properties of the epidermis layer of skin. The dermis layer provides protection from physical injury as well as the ability to sense the environment. This layer is generally thicker than the epidermis and can be several millimeters thick in rats. Dermis tissue is dense with fibrous proteins, nerves, and blood vessels. 143 The dermis also contains a number of important appendages such as hair, sebaceous and sweat glands. These offer additional regulatory functions, protection to the organism, and help to contain and dissipate heat. Despite its large volume, the dermal structure is mostly made up of connective tissue 144 and, to a lesser extent, of cells. The major cell type in the dermis is fibroblast which produces the connective substances. Collagens I and III are the predominant protein types in the dermis, collagen I being more abundant in the lower-lying reticular dermis and representing as much as 90% of the mass of the skin tissue. 145 In addition to collagens, elastin protein is found in the dermal layer. Although composition of the dermis is complex, for the purpose of skin spectra, the dermis is represented in this dissertation by water and collagen type I protein standards. Elastin protein spectra were collected in the combination region but did not improve the quality of the fit and are not included in the model. Finally, the subcutaneous layer is rich in fat and serves to insulate the body and buffer trauma. The 1 to 1.5 mm thick fold of rat skin used for the in vivo experiments reported in this thesis did not have a substantial amount of subcutaneous tissue. 38 Care was taken to remove the subcutis from the skin samples used for the microspectroscopic studies described herein. For this reason, rat skin samples contain only minor amount of fat that is contained within the epidermis layer. Physiological Basis for Skin Tissue Heterogeneity Spatial skin heterogeneity can be characterized at various levels: from intracellular structures to millimeter-size structures , 148 Considering the diameter of

89 61 the fiber optic setup and the step size of the motorized microspectroscopic stage, skin structures with sizes on the order of several hundred microns are of relevance in the present investigation. Such structures are plentiful in the rat skin and include wrinkles, 149 sebaceous glands, 148 and hair follicles. 150 Images of desiccator dry rat skin samples were recorded with a universal photoscope (Carl Zeiss Microimaging, Thornwood, NY) in transmission mode. The images in Figure III-1 show some hair fibers visible around the edge of the skin samples. Even though the optical density of dry tissue was high, spatial heterogeneity on the order of several hundred microns can be inferred, as is particularly noticeable in the image presented in Figure III-1(C). Near-infrared spectroscopy is sensitive to the chemical composition of appendages which is dissimilar to the surrounding matrix. For instance, hair follicles may be the source of elevated keratin and lipid signals, 132, 151 whereas wrinkles have a different protein-to-water ratio. Appendages are not the only source of skin lateral inhomogeneity. The dermalepidermal junction has been shown to resemble a wavy surface with the dermal papillae invading the epidermis. 90 These features can contribute to heterogeneity in the spectroscopic tissue maps. Stratum corneum in itself is an inhomogeneous structure with differences in the lateral distribution of lipids and proteins. Keratinocytes tend to arrange in stacks or columns up to a hundred microns in diameter. 146 Chapter Overview This chapter builds up on the work detailed in Chapter II and demonstrates the ability to map the distribution of the major chemical components (water, collagen type I, fat, keratin, constant, and slope) within the skin matrix semi-quantitatively. Tissue maps are generated using microspectroscopy for near-infrared spectral mapping of mm sections of excised rat skin samples. An analysis of the resulting chemical 2 distribution maps reveals information pertaining to the spectral variance within selected measurement sites and between animals. Sex related differences are also noted. This experiment provides the database for subsequent rat skin tissue simulation studies.

90 62 (A) (B) hair hair (C) Figure III-1. Microscopic images of dried rat skin from the area between the shoulders. Images (A) and (B) show the edge of the skin with hair fibers visible; (C) shows the center of the sample with white rectangles marking particularly heterogenous areas. Data collected with a universal photoscope (Carl Zeiss Microimaging, Thornwood, NY) in transmission mode. The size of the white bar is 100 μm.

91 63 Experimental Procedures Rat Skin Samples and Reagents Skin tissue samples were harvested from euthanized retired breeder Harlan Sprague rats weighing between 400 and 500 grams, 8 male and 4 female. The sacrificed rats were part of a 7-hour in vivo glucose experiment performed on the same day. The sampled skin tissue was excised from the site where the noninvasive glucose spectra were collected as outlined in Figure I-1(A). The animals were anesthetized for surgery with ketamine and the 10-fold increase in this anesthetic agent was used to sacrifice the rat. The effects of the drug had no significant impact on the measured skin composition as its been shown to be very similar for anesthesized and non-ansthesized animals. 37 All procedures were performed according to a protocol approved by the University of Iowa Animal Care and Use Committee. All the personnel had completed appropriate training. Immediately following sacrifice, the abdominal cavity was cut open and the animal was left lying on its belly until blood moved down from the neck area. This was done as a precaution against contamination of our samples with excessive blood. Next, the skin between the shoulders was shaved with an electric razor and approximately 6 6-mm 2 piece of whole skin tissue was excised in one piece. This single sheet of skin was then cut into 4-5 separate pieces that where immediately snap frozen in liquid nitrogen. The frozen samples were submerged in chilled phosphate buffer saline (Fischer) and stored at -18ºC. Instrumental Setup Mapping was performed with the setup described in detail in Chapter II. Briefly, spectra were collected by using a Nicolet Magna 560 FTIR (Thermo-Nicolet Corp., Madison, WI) coupled with an IR-Plan Advantage microscope (SpectraTech, Inc., Shelton, CT) in the combination near-infrared region ( cm -1 or µm). All spectra were collected as 8k, double-sided interferograms treated with a triangular apodization and standard Mertz phase correction. To increase the signal-to-noise ratio,

92 scans were accumulated and averaged per spectrum. Incident radiant powers were maximized by removing all of the external apertures within the microscope. A custom-made compression cell was used to hold the tissue samples on the motorized XY stage of the microscope. This cell was based on a 1-inch optical mirror holder with threads cut into the inner surface. Two round sapphire windows (0.020 inch thick, 1 inch in diameter, Mettler Optics, Inc., Providence, RI), were placed inside the holder with a skin sample or a Teflon spacer sandwiched between. Two holes, each about 3 mm in diameter, were drilled into the opposite sides of the holder to relieve air pressure when a sample was squeezed. The widows and the sample were held in place with a metal ring-screw. The thickness of the sample was varied by screwing the ring-screw tighter and compressing the tissue between the sapphire windows. Lastly, the compression cell was tightly wrapped in a ribbon of Parafilm to cover the holes and slow the rate of tissue dehydration. Photographs of the instrumental setup are presented in Figure III-2. The motorized stage with a sample cell inserted is shown in Figure III-2 (A). Figures III-2 (B) through (D) show the compression cell with Teflon spacers assembled for air measurements, the cell with a skin sample inside, and the top view of the skin sample inside the cell, respectively. Knowing that the inner diameter of the compression cell was 1 inch, the dimensions of the skin tissue slice can be inferred from these images. Spectral Data Collection Before data collection, a skin sample was thawed, brought to room temperature, removed from the buffer solution and patted dry with Kimwipes. Meanwhile, background air spectra were collected in triplicate at room temperature. The background sample was the empty compression cell with a 0.92 mm thick Teflon spacer inserted to control the distance between the two windows. In addition, a neutral density filter A, supplied with the spectrometer, covered the inner aperture to decrease the light intensity at the detector for the air background measurements. For the microscope setup this screen was

93 65 (A) (B) objective sample cell (C) (D) Figure III-2. Photographs of the microspectroscopic mapping setup showing (A) motorized stage with a sample cell, (B) compression cell with Teflon spacers for air measurements, (C) compression cell with a skin sample inside, and (D) top view of the skin sample inside the cell.

94 66 experimentally determined to transmit 21% of the incident radiation. Following background collection, the A filter was detached, the sample cell was disassembled, the Teflon spacer removed, and the skin sample was placed between the two sapphire windows. The pressure applied to the sample was adjusted manually until sufficient level of near-infrared signal was detected. The adjustment of sample thickness was done with care so that the skin components were minimally displaced. Typically, sample thickness after the adjustment ranged from 0.4 to 1.0 mm. The compression cell filled with the sample was fitted into the opening of the motorized XY stage and left at room temperature for at least 15 minutes to stabilize the temperature of the tissue. During the course of the data collection, the largest degree of ambient temperature change was between 21 to 24 C. More exact control of skin sample temperature was not attempted because heating of the skin due to irradiation was deemed to be small. A schematic diagram of the cell is presented in Figure III-3(A). Each spectral map was collected in a 10-by-10 raster pattern illustrated in Figure III-3(B). The step size of the stage was 480 μm in the X direction and 360 μm in the Y direction. The stage had to be controlled via mapping software throughout the experiment because every time the next skin sample was positioned in such a way that the starting coordinate for the map was always different. The array corresponded to 100 evenly spaced locations across an area of mm 2. It took about two minutes to collect a single tissue spectrum and move the stage to the next location. Every 30 minutes the stage was returned to the central location and reference spectra were collected in triplicate. Data collection for each map required approximately three hours, including 30 minutes for the tissue to thaw to room temperature and 2.5 hours to collect spectra. Visible images of tissue are added in Figure III-3 (C). These rectangles are snapshots of the upper surface of skin samples taken during data collection. Each rectangle is μm 2 in size as indicated by the scale bar. Some structures (mostly, hair and trapped bubbles) provide a scale for the

95 67 (A) (C) (B) Figure III-3. Schematic of the microspectroscopic mapping setup including (A) the compression cell, (B) the skin sample with the raster pattern for the measured spectral array, and (C) the CCD camera images of skin tissue. Red dot in the center of the spectral array marks the reference location. The length of the white bar on the CCD skin images is 100 μm.

96 68 mapped area. It should be noted that these images are of the top surface only and do not reflect the heterogeneity obtain with microscopic data collected through all layers of these skin samples. Tissue mapping was performed with OMNIC Atlus software (Thermo-Nicolet Corp.) and MATLAB 7.0 (The Mathworks, Inc., Natick, MA) was used for all statistical analyses. Spectral data were transferred to a Pentium V PC running the Windows Vista operating system. The interferograms were Fourier transformed to yield singlebeam spectra covering the cm -1 ( μm) range with a nominal point spacing of 3.8 cm -1 and a resolution of 7.6 cm -1. All further analysis was done in MATLAB (Version 7.0, The MathWorks, Inc., Natick, MA). Data Processing Absorbance Spectra and Spectral Quality Air reference spectra used to compute tissue absorbance spectra were collected in triplicate by using a blank compression cell equipped with a 0.92 mm Teflon spacer with air between the two sapphire windows. Absorbance spectra of skin tissue were calculated as the negative logarithm of the ratio of the tissue single beam spectrum relative to an average air reference single beam spectrum. Spectral quality was assessed as the root-mean-square (RMS) noise on 100% lines in two adjacent regions: cm -1 and cm -1. A 100% line is obtained by calculating the negative logarithm of the ratio between two back-to-back single beam spectra collected from the central location. The resulting 100% line was then fitted to a second order polynomial, and the root-mean-square (RMS) error of this fit was obtained by using Equation II-1. The RMS noise values were determined for the single set of triplicate air single beam spectra and the multiple sets of triplicate single beam spectra collected at the reference point of the tissue slice. For each case, the average value and the standard deviation are calculated for the three possible combinations of these single beam spectra (e.g., ratios of first-to-second, second-to-third, and third-to-first).

97 69 Spectral Fitting Quantitative information can be extracted from each skin spectrum by fitting it to a set of standard spectra that represent the main chemical components of the skin matrix. Similarly to the lipid/water and nail samples in Chapter II of this work, each skin spectrum was fitted to a set of standard spectra representing water, collagen protein, keratin protein, and fat. Additional terms were included in the regression analysis to account for a constant offset and a sloping baseline. Each spectrum was fitted by least squares regression over the cm -1 spectral range following Equation II-6. The resulting six regression coefficients represent the relative amount of each component (water, collagen, keratin, fat, slope and offset) for each location of the skin slice. For comparisons between animals and sexes, weighted percent coefficients (i. e., relative regression coefficients) were calculated for each of the four chemical components according to the following equation: comp (%) i ( water collagen_i comp_i fat keratin 100% ) i Equation III-1 where (%) comp i is the relative amount of a given chemical component at location i on the skin map, comp_i is the measured regression coefficient for this component at location i, and water, collagen_i, fat and keratin represent the measured regression coefficients for all four of the chemical components in the tissue model (water, collagen I protein, fat, and keratin protein, respectively). Construction of 2D Tissue Maps Chemical distribution maps for all six tissue components were created in MATLAB software by analyzing 100 near-infrared spectra collected over an area of mm 2, as schematically shown in Figure III-3(B). Two-tone false-color distribution maps were used to represent three-dimensional data in two dimensions. This type of plot uses a density of color (the lighter the shade the higher the value at the particular

98 70 location) at each rectangular μm 2 pixel to report the average value of the four surrounding tissue locations. This averaging step can blur the image. The contour plot was added to reveal domains where the difference across the line either increases or decreases by 10% of the overall change across the mapped skin slice. 152 Both of these plots were used to construct spatial distribution maps for the six major tissue components. Correlation between Tissue Components In this work, coefficients of correlation (r), or Pearson correlation coefficients, were calculated for 15 possible combinations of the six fitted components for each map on the basis of the regression coefficients calculated from all 100 spectra collected across each tissue slice. Correlation coefficient is a measure of the linear relationship between pairs of components, represented here as x and y, of the six-component model for a given skin map. The value for each pair is reported only once, that is r for water versus fat is the same coefficient as the r for fat versus water. Equation II-2 used to calculate the correlation coefficients follows. r n i 1 n i 1 ( X X ) i ( X X )( Y i 2 i n Y ) i 1 ( Y Y ) i 2 Equation III-2 where n is the number of data pairs (here n is 100 for each skin map), i is the number of a data point on the map, X i and Y i are the values for components x and y, X and Y are the mean values for component x and y, respectively. 153 Values for component x are used as a reference plotted on the horizontal axis. If the r value approaches unity, the components x and y are perfectly correlated and their correlation plot is a line. On the other hand, if r is equal to 0 no correlation between two components is present. Thus it is said that r is the measure of the strength of correlation. The slope of the fitted line determines if the correlation is positive or negative. This property would be lost if the squared coefficients of determination r 2 were used.

99 71 To assess the significance of the computed correlation coefficients, standard errors were calculated according to the Equation III-4 below. SE r 2 1 r n 2 Equation III-3 where r is the coefficient of correlation between two skin components, and n represents the number of data points (for a skin map n = 100). 154 Whenever the correlation coefficient r value is lower than its standard error it is assumed there is no correlation between the components. Analysis of Variance A balanced one-way ANOVA was performed to statistically compare changes in the magnitude of the regression coefficients between locations in one sample and between skin samples obtained from different animals. Maps from each of the four samples from each animal constituted a group that consisted of four sets of 100 spectra per tissue slice. A linear model was used that assumes the means of the different groups are statistically the same. The ANOVA model is described by the following equation: yij = αj + εij Equation III-4 where yij is a matrix of coefficients with each column representing data for a different group, αj is a matrix whose columns are the group means (mean values for a given animal), and εij is the matrix of unmodeled residuals. 155 MATLAB software (Version 7.0, The MathWorks, Inc.) was used to calculate probability (p values) for the null hypothesis that the regression coefficients for a particular chemical component obtained from the four animals belong to the same population. In other words, this p value corresponds to the probability of the null hypothesis being true or that the mean values are not statistically distinguishable between these groups. If the p value is equal to one, this indicates that sample means of all groups come from the same population distribution and that the variance between groups (in this case, between animals) is statistically larger

100 72 than the variance within groups (between map locations for a single tissue slice). A p value near zero indicates that at least one group is significantly different than the others. Results and Discussion Spectral Quality Assessment Microspectroscopic Data First method of data collection for rat skin tissue attempted was by simply positioning the slice on top of a single sapphire window without any compression. These attempts failed because without pressure applied to the skin sample path lengths were too great to achieve meaningful signal-to-noise ratios. Moreover, uncovered and unfixed tissue dried out too quickly when exposed to air. This arrangement did not match the noninvasive optical setup either where a fold of skin is tightly squeezed between two sapphire rods. By compressing the skin, the spacing between the ECM components decreases and the average light scattering is reduced. 156 In order to increase the light throughput and even out the path length, a compression cell shown in Figure III-3 was built. In a sense, the procedure was similar to noninvasive in vivo experiment where the distance between two rods was also controlled to optimize the signal from the fold of rat skin. The compression cell setup provided means to reduce the rate of dehydration of the sample during the lengthy mapping procedure. In all, 5,808 single beam spectra were collected for this study corresponding to both air and skin tissue spectra taken from 12 separate animals with 4 samples per animal. For each skin slice, spectra were collected at 100 locations over a mm 2 area. Reference spectra were measured repeatedly from the center position on the map to determine the integrity of the tissue matrix. Even though the sample cell was tightly wrapped in Parafilm, some water was able to escape during mapping procedure. The amount of water loss was determined by periodically weighing the tissue slice together with the compression cell. The relative weight loss was measured to be 4% for

101 73 the first two hours of the mapping experiment. However, the water coefficient data do not reflect a decrease in water content. The reason is that the map was collected at the central part of the skin slice whereas the dehydration occurred primarily at the outer edges of the sample. RMS Noise for Air Spectra The quality of the spectral data was judged by the intensity of collected radiation for the air spectrum. As this intensity increases, the RMS noise decreases, indicating improved signal-to-noise ratio. During the course of these experiments, it was discovered that the instrumental alignment was subject to change and required frequent adjustment to maintain an optimal signal and high SNR. Table III-1 reveals how the data can be split according to quality of the spectra as judged by the need to include a neutral density filter (A screen) to avoid detector saturation while collecting air spectra. Early data were collected with poorer alignment, resulting in lower radiant powers at the detector and no need for the neutral density filter. During data collection for rat # 6, it was discovered that misalignments of the InSb detector associated with the drastic temperature changes when the detector was cryogenically cooled. Adjustment of the filter position within the microscope housing improved performance, thereby increasing the magnitude of the detected signal and requiring the neutral density filter for air spectra. With time it became evident that routine re-alignment was needed to be performed to get the optimal signal. That is why the samples collected later on generally have a higher signal level. Spectral quality is of paramount importance for successful noninvasive glucose monitoring. 19, 125 A key indicator of spectral quality and basic spectrometer performance is the RMS noise on 100% lines. For the air spectra, RMS noise levels ranged from 20 to 115 μau. The wavelengths between cm -1 and cm -1 were chosen for noise assessment because they correspond to the maximum signal (i. e. minimum absorbance value) of rat and human skin. 37 For the lower signal group, no neutral density

102 74 Table III-1. Summary of lower and higher microscope signal groups of rat skin maps. Maps with lower signal Maps with higher signal (no neutral density filter) (with neutral density filter) Male Rats Rat 1 14, 15, 16, 17 Rat 2 18, 20, 22, 23 Rat 3 24, 25, 26, 27 Rat 4 32, 34, 35, 36 Rat 5 37, 39, 40, 41 Rat , 48, 49 Rat 7 50, 51, 53, 54 Rat , 59, 60 Female Rats Rat 1 62, 63, 64, 65 Rat 2 66, 67, 68, 70 Rat 3 71, 72, Rat , 80, 81

103 75 filter was necessary for quality air spectrum collection. On the other hand, air spectra for the higher signal group were obtained with an A neutral density filter to prevent detector saturation. As a result, the noise values for air from the higher signal group were slightly greater than for the first three rats from the lower signal group as detailed in Table III-2. The microspectrometer performance was similar for samples associated with the first three male animals and then performance degraded slightly for the fourth and fifth male rats. As indicated in Table III-2, the mean values were 28.8, 32.7, 42.1, 65.3, and 76.3 micro-absorbance units (µau) respectively. Male rats # 4, 5, and female rat # 1 correspond to the move of the equipment following the flood of the Iowa Advanced Technology Laboratories. Hence RMS noise values for these air measurements are uncharacteristically large. Noise values for air are plotted in Figure III-4. The decrease in signal quality for maps is illustrated in Figure III-4 (A) and (B). On all of the RMS noise level plots in this chapter, the noise values on 100% lines obtained from the first and the second air spectra are denoted with a red cross, the second and the third air spectra are presented as a blue circle, and the first and third spectra are shown as a green plus. No trend was observed in terms of the relative noise magnitude for different pairs of spectra, which demonstrates that there was no drift in spectrometer alignment during collection of these air spectra. Even though the RMS noise on 100% lines of air spectra for the group with the higher microscopic signal was larger than the values for the first three male rats before the move of the equipment, this elevation is misleading because an A filter was used to attenuate 79% of the incident radiation. The instrumental performance in this higher signal group was in fact improved for skin samples collected without an A filter and more stable than for the lower signal group as demonstrated in Figure III-4 (C) and (D). In general, the levels of the RMS noise for air obtained in the microscopic system are at least 10 to 40 times higher than either those available with the use of Nicolet

104 76 Table III-2. RMS noise values on 100% lines of air reference spectra. Rat # 1 to 2 2 to 3 1 to 3 Mean ± Std Dev a, μau ± Std Dev a, μau ± Std Dev a, μau ± Std Dev, μau Lower signal group* cm -1 1 (male) 27.2 ± ± ± ± 3.6 b 2 (male) 32.2 ± ± ± ± 4.9 b 3 (male) 43.2 ± ± ± ± 9.6 b 4 (male) 67.1 ± ± ± ± 8.1 b 5 (male) 90.0 ± ± ± ± 11.6 b 6 (male) ± (male) ± (female) ± cm -1 1 (male) 30.1 ± ± ± ± 5.4 b 2 (male) 40.3 ± ± ± ± 10.5 b 3 (male) 46.3 ± ± ± ± 7.9 b 4 (male) 66.9 ± ± ± ± 12.7 b 5 (male) 73.6 ± ± ± ± 14.9 b 6 (male) ± (male) ± (female) ± 5.5 Higher signal group* cm -1 6 (male) 47.7 ± ± ± ± 9.5 b 7 (male) 55.9 ± ± ± ± 19.2 b 8 (male) 48.7 ± ± ± ± 7.9 b 1 (female) 55.7 ± ± ± ± 8.4 b 2 (female) 61.4 ± ± ± ± 9.9 b 3 (female) 59.6 ± ± ± ± 15.3 b 4 (female) 55.7 ± ± ± ± 9.5 b cm -1 6 (male) 48.2 ± ± ± ± 7.6 b 7 (male) 73.8 ± ± ± ± 18.6 b 8 (male) 61.7 ± ± ± ± 11.2 b 1 (female) 59.0 ± ± ± ± 13.8 b 2 (female) 50.0 ± ± ± ± 10.4 b 3 (female) 66.1 ± ± ± ± 17.0 b 4 (female) 64.0 ± ± ± ± 9.0 b Note: a The number of degrees of freedom is 3, except for male rats 6 and 8, and female rat 4 (single map) * Defined in Table III-1. b Pooled standard deviation for the mean value.

105 RMS Noise Level, microau RMS Noise Level, microau RMS Noise Level, microau RMS Noise Level, microau to 2 2 to 3 1 to 3 (A) to 2 2 to 3 1 to 3 (B) Map Number Map Number to 2 2 to 3 1 to 3 (C) to 2 2 to 3 1 to 3 (D) Map Number Map Number Figure III-4. RMS noise on 100% lines of air reference spectra for the lower signal group of data in the regions between (A) cm -1 and (B) cm -1 ; the higher signal group of data in the regions between (C) cm -1 and (D) cm -1.

106 78 Magna 560 spectrometer for bulk sample measurements (~ 3 µau for the same spectral regions) or in the noninvasive near-infrared measurement setup used for the in vivo experiments on rats and human subjects. Multiple optical interfaces and inaccuracies in alignment, as well as the use of restrictive apertures all cause the loss of light intensity in the microscope. As a result, the microspectroscopic system was impractical to use for measurements of thicker skin samples. RMS Noise for Rat Skin Spectra Table III-3 presents a summary of the RMS noise levels for rat skin tissue spectra. Microscope performance was not the main factor determining the RMS noise level for these skin samples. Absorption and scattering properties of the tissue as well as the thickness of the sample chiefly determined the spectral quality. Mean RMS noise values ranged from to μau for male rats and from to μau for female rats. The spectra for female rats tended to have lower noise levels primarily because these samples were thinner than male skin samples and they were collected later in the experiment. Figure III-5 presents the plots for the RMS noise values obtained for rat skin. The increase in RMS noise for air seen for rat maps # 32 56, as shown in Figures III-4 (A) and (B), does not appear for rat skin samples in Figure III-5 (A) and (B). Unrelated to the quality of optical alignment, properties of the tissue primarily determined the noise level. It is important to realize that spectra collected through the microscope are not of sufficiently high quality to permit glucose measurements. The microspectroscopic setup has not been intended for glucose detection. The lowest measured level of RMS noise for rat skin is 90 μau, which is approximately a factor of ten too high for successful measurement of glucose in these slices of tissue. Specifically, in a 1-mm thick cell filled with aqueous solution, RMS noise level should be less than 20 μau to make glucose detection feasible. 123 Such low noise level has been obtained by using a more powerful light source and optical fiber interface in the in vivo experiment, 38 whereas the

107 79 Table III-3. RMS noise values on 100% lines of tissue reference spectra for the lower signal group of rat skin data. Rat # 1 to 2 ± Std Dev a, μau 2 to 3 ± Std Dev a, μau 1 to 3 ± Std Dev a, μau Mean ± Std Dev, μau Lower signal group* cm -1 1 (male) ± ± ± ± 76.8 b 2 (male) ± ,003.3 ± ,010.2 ± ± b 3 (male) ± ± ± ± b 4 (male) ± ± ± ± b 5 (male) ± ± ± ± b 6 (male) ± (male) ± (female) ± cm -1 1 (male) ± ± ± ± b 2 (male) ± ± ± ± b 3 (male) ± ± ± ± b 4 (male) ± ± ± ± b 5 (male) ± ± ± ± b 6 (male) ± (male) ± (female) ± 26.2 Higher signal group* cm -1 6 (male) ± ± ± ± 15.1 b 7 (male) ± ± ± ± b 8 (male) ± ± ± ± b 1 (female) ± ± ± ± 66.0 b 2 (female) 91.1 ± ± ± ± 22.6 b 3 (female) ± ± ± ± 25.4 b 4 (female) ± ± ± ± 61.9 b cm -1 6 (male) ± ± ± ± 12.9 b 7 (male) ± ± ± ± b 8 (male) ± ± ± ±192.1 b 1 (female) ± ± ± ± 52.4 b 2 (female) 90.0 ± ± ± ± 21.7 b 3 (female) ± ± ± ± 28.1 b 4 (female) ± ± ± ± 59.6 b a The number of degrees of freedom is 3, except for male rats 6 and 8, and female rat 4 (single map).*defined in Table III-1. b Pooled standard deviation for the mean noise value.

108 RMS Noise Level, microau RMS Noise Level, microau RMS Noise Level, microau RMS Noise Level, microau to 2 2 to 3 1 to 3 (A) to 2 2 to 3 1 to 3 (B) Map Number Map Number to 2 2 to 3 1 to 3 (C) to 2 2 to 3 1 to 3 (D) Map Number Map Number Figure III-5. RMS noise on 100% lines of rat skin tissue spectra for the lower signal group of data between (A) cm -1 and (B) cm -1 ; the higher signal group of data between (C) cm -1 and (d) cm -1.

109 81 microscope has not been optimized for such performance. Nevertheless, the RMS noise levels in Table III-3 are sufficiently low to enable estimation of the major biochemical components shown in Figure II-2 within the skin matrix. These components have higher absorptivity values (for instance, fat) in the combination region and/or higher concentrations than glucose (water, collagen type I protein, keratin protein), which provide a means to map their spatial distribution with the microscope in its current configuration. Rat Skin Tissue Absorbance Data Visual Inspection of Rat Skin Absorbance Spectra To illustrate the appearance and quality of the microspectroscopic mapping data, results are presented in Figure III-6 for two maps: sample # 27 from male rat 3 and sample # 70 from female rat 2. The absorbance spectra for these tissue slices can be considered representative for their respective sexes. Figure III-6 (A) and (B) are the absorbance spectra for point maps of male rat sample #27 and female rat sample #70 respectively; (C) and (D) are the corresponding reference absorbance spectra at the central location on the skin slices; and (E) and (F) are the corresponding absorbance residuals for these maps after the six-component fit. The dominant absorption features of skin tissue for both samples correspond to water absorption. The water spectrum over this region corresponds to a minimum between the overtone OH vibration mode centered at 5130 cm -1 and the fundamental OH stretching mode centered at 3850 cm These features can be observed for both male and female rats in Figure III-6 (A) through (D) as a U-shaped spectrum. The absorptivity value 43 for water between 4,400 cm -1 and 4,600 cm -1 is at its minimum of about mm -1 mm -1 but as the amount of water in biological tissue is so high (about 70%) that it determines the overall shape of the skin spectrum. Major classes of lipid present in rat epidermis are ceramides (45-50%), cholesterol (25%), and free fatty acids (10-15%). 158 The sharper features at 4330 cm -1 and

110 Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU 82 (A) (B) Wavenumber, cm Wavenumber, cm (C) (D) Wavenumber, cm Wavenumber, cm (E) (F) Wavenumber, cm Wavenumber, cm Figure III-6. Representative spectral data for rat skin samples: (A) and (B) are the absorbance spectra for point maps of male rat sample #27 and female rat sample #70, respectively; (C) and (D) are the corresponding reference absorbance spectra at the central location of the skin slices; and (E) and (F) are the corresponding absorbance residuals for these maps after the six-component fit.

111 cm -1 are characteristic of lipid molecules. 56 Absorption at these wavelengths originates from the combination of CH stretching and bending modes associated with the long aliphatic chains. 101 It should be noted that there is lack of substantial subcutaneous fat on the animal s upper shoulders as was verified by postmortem dissections. However, absorptivity values at 4330 cm -1 and 4255 cm -1 are sufficiently large to result in features in these rat skin spectra. Rat skin is known to have a higher fat content in epidermis among several species. 158 Higher lipid content of rat skin was also be seen in the form of two near-ir peaks around 4200 and 4300 cm -1 in Figure I-2. In that figure the aforementioned lipid absorption features were more prominent for rat than human tissue. Also, female rats tend to have a higher lipid percentage in their epidermal layer, 137 which is reflected in the higher magnitude of lipid peaks for sample 70 compared to sample 27 in Figure III-6 (C) relative to overall skin absorbance level. Protein and peptide IR spectra have been studied in the literature somewhat broadly but not systematically enough. 47, 49, 50, 52, 54, For this reason, band assignment for proteins throughout the whole region of the near-infrared still remains an active area of research. These appear in the whole skin spectra as a broad feature around 4600 cm -1, the shoulder at 4400 cm -1, and as a contribution from the peptide bond at 4890 cm -1. No major differences are noted between protein absorptions for male and female samples. Similarly to the models for biological samples in Chapter II, incorporating terms for both constant and slope into the regression model of rat skin is an important step for several reasons. First of all, both of these parameters model the scattering background in the multi-layer whole skin spectra. The level of scattering depends on skin composition and structure at the specific locations on the map. This information about spatial distribution of scattering is very important for future simulation studies. That is why no mathematical correction has been attempted in pre-processing the tissue spectra but rather

112 84 the scattering element was modeled together with the biochemical components. Instrumental deviations due to changes in alignment, stray radiation, and reflection losses at the sapphire window surfaces add to the spectral background, but these components are constant across all locations on the tissue map. The near-infrared spectrum of water is known to be affected by temperature with 46, 123 absorbance bands shifting to higher wavenumbers with an increase in temperature. Because no temperature control was used during the mapping procedure, the temperature effect must be estimated by a change in the slope of the baseline in the skin model. In skin tissue, scattering also changes with temperature. 131 This slope of the change in scattering is opposite in sign to the water absorption changes. Overall, the cumulative slope changes are relatively small and not apparent in Figure III-6. For successful modeling, the spectral standards used to fit the tissue slice spectra must accurately match the spectral features associated with the main chemical components within the skin tissue matrix. Structure of the residual spectra presented in Figures III-6 (E) and (F) indicates that the tissue spectra are largely modeled by the six component spectra used in this study. Based on the magnitude of the residuals, the undermodeled part of the spectral information is generally as little as 0.5-2%. In other words, 99.5% to 98% of the skin absorption is accounted for by the six-component model. The exact value depends on the wavelength and the sample composition. Obviously, the most intense nonrandom features in residuals are associated with the lipid-related bands at 4330 and 4255 cm -1. Previous research in Arnold s group has addressed variations in combination lipid spectra. 111 The differences noted were small and bovine fat was deemed an adequate standard for modeling lipid features in skin spectra. However, fat originating from rat skin samples was not measured at that time. The effect of the RMS noise on the shape of absorbance residuals for sample maps of male (# 27) and female (# 70) rats can be visually assessed. In Figure III-6 (E) and (F) noise level shows up in residual absorbance spectra after subtraction of the six

113 85 noiseless standard spectra for major skin components. Evidently, in the case of map # 27 with a higher noise level of about 1200 μau random fluctuations of the spectral noise largely contribute to the shapes of residuals. Map # 70 has a lower noise level of 100 μau. Consequently, the shape of its residuals is dominated by nonrandom features still present after the six component linear fit. The nonrandom differences between the fitted spectra and the skin spectra can correspond to either a mismatch between the standard spectra and the spectra of the tissue components in situ or an insufficient number of standard spectra. In addition to the uniqueness of lipid absorbance, it is known that secondary structure of the protein and its physical state have an effect on its near-infrared spectrum. 51, 53 Thus collagen type I protein fiber bundles in their hydrated state in tissue may have modified absorption properties compared to those in the standard dry collagen KBr pellet used to prepare the standard spectrum. The residual features due to protein molecules are 2 to 4 times lower than the contribution from the fat. For this reason, the effect of species-specific lipid was the first parameter studied to improve the regression model fitting as described in Chapter VI. Six Component Regression Fit After the linear fit to the absorbance data at each location on the microspectroscopic map, regression coefficients were generated for water, collagen type I protein, fat, keratin protein, constant, and slope according to Equation II-2. Coefficient profiles for samples 27 and 70 are summarized in Figure III-7. In the figure, (A) and (B) are the results of regression fit plotted for the 100 points of male rat map 27 and female rat map 70, respectively; and (C) and (D) are the corresponding regression coefficients for spectra collected periodically at the central location on these skin slices for reference purposes. Although such plots provide no direct spatial information, they display the degree of variation of each regression coefficient and, hence, the contribution of each skin component. Clearly, the coefficient for the constant offset term has the highest

114 Regression Coefficient Regression Coefficient Regression Coefficient Regression Coefficient (A) water collageni fat keratin constant slope Mapping Location Number (B) water collageni fat keratin constant slope Mapping Location Number (C) water collageni fat nail constant slope (D) water collageni fat nail constant slope Reference Spectrum Number Reference Spectrum Number Figure III-7. Representative regression coefficients for skin samples: (A) and (B) are the results of regression fit for point maps of male rat sample #27 and female rat sample #70 respectively; (C) and (D) are the corresponding regression coefficients at the central location on the skin slices.

115 87 magnitude relative to all other constituents. The nature of this large scattering background in the microscopic measurements has been already discussed. The degree of variation in the constant term is found to be proportional to its magnitude. In other words, its relative variation is similar to that of other components as can be seen in the plots of the change in coefficients for the standard components across the 100-point map relative to the coefficient measured at the starting location on the map (see Appendix Figure A- 1). In fact, the largest relative variation was found for the fat component even though the magnitude of the fat regression coefficient is generally very low. Indeed, small changes in fat will appear large relative to the negligible overall fat content. Apart from the constant, water and collagen type I protein are the major contributors to the skin spectra. The values for their regression coefficients are similar and are approximately 0.5 for male rats and 0.3 for female animals. Once again, these numbers do not imply that the concentrations of collagen and water in the beam path are the same. Regression coefficients should be thought of as thicknesses of the standard components layers. Because the standard for water was a 1 mm thick sample and the standard for collagen type I protein was its KBr pellet, these two components will have very different concentrations from their regression coefficients. Regression coefficients are used throughout this dissertation for characterization of spatial heterogeneity of tissue samples with near infrared microspectroscopy and subsequent simulation of the skin spectra based on the six standard spectra as shown in Figure II-1. The plots in Figure III-7 give some general ideas of the degree of spatial heterogeneity in male and female rats. Evidently, all the coefficients are lower for the female rat because the female skin is thinner and more easily squeezed. However, for both male and female skin the variations in all six components are about an order of magnitude larger for the map on the whole than for the central reference location. The principal difference between Figures III-7 (A), (B) and (C), (D) is that the former contain spatially resolved data, whereas the latter demonstrate the degree of temporal changes in

116 88 the skin sample and the reproducibility in the spectroscopic measurements. Parafilm wrapped around the compression cell helped to minimize dehydration for the time period of data collection. Rearrangement of interstitial fluid inside the sample after the compression in the sample cell has taken place before mapping while the sample was left to equilibrate to the room temperature. Consequently, the changes in skin components featured in Figures III-7 (A) and (B) are treated as predominantly coming from the inherent spatial heterogeneity of the skin. Small temporal changes still evident in Figures III-7 (C) and (D) (for example, in the collagen type I protein profiles) can be attributed to the slight misalignment of the motorized stage every time it returned to the central reference location. An exception is the gradually decreasing scattering background for the female sample. This decrease shows up both in the mapping and the reference data as the sloping profile for the constant term. Analogously, a gradual decrease in components has been seen in the in vivo experiment with the living rat. 37 Structural rearrangement within the skin after squeezing is offered as the explanation for this observation. Despite the slight variations in the six components at the reference site, the overall findings indicate that the bulk of variations observed in Figure III-7 correspond to spatial differences in the skin composition and not variations caused by chemical changes of the tissue matrix or variations in performance of the instrumentation. Rat Skin Spatial Heterogeneity Spatial Contour Maps Fourier transform near infrared microspectroscopic technique used in this study provides micrometer-scale resolution with chemically sensitive detection. Measurement in the near-infrared range requires no special tissue treatment or staining. Thus, the sample is studied in a condition that is close to the actual in vivo environment. Two main modes are used to obtain spatial information about tissue: mapping and imaging. These have been already described in Chapter II. The key difference between

117 89 the two approaches is the time required to collect the data from a sample of tissue. Although the imaging approach reduces the collection time dramatically, mapping and imaging will supply essentially the same chemical information for a stationary sample in the absence of rapid chemical changes. In this chapter, large areas of skin are characterized by limiting each map to 100 locations and spacing them sufficiently apart to provide distribution maps in reasonable time periods. Aside from comparing the magnitude of the different regression coefficients at different positions, the degree of tissue heterogeneity can be assessed by evaluating spatial distribution maps for the six components. Distribution maps are presented in Figure III-8 and Figure III-9 for the representative tissue slices: male sample 27 and female sample 70, correspondingly. In both figures, separate maps illustrate the distribution of water (A), collagen type I protein (B), fat (C), keratin protein (D), constant term (E), and sloping baseline (F). These twotone distribution maps plot a density of color at each pixel to represent the average of the four surrounding tissue locations. 152 Contour lines delineate regions where the difference across the line either increases or decreases by one-tenth, or 10%, of the change in that coefficient over the whole 100-point map. Finally, a pair of white circles is included in the distribution maps for water to indicate the size of the optical fibers used to collect the noninvasive spectra reported from living rats. 112 In this case, the fiber diameter is 1.8 mm. In Figures III-8 and III-9 larger amounts of each component are indicated by a lighter shade of orange color. A higher density of contour lines in these plots indicates greater differences in a certain component between locations on the tissue slice. Based on visual inspection, all of the components demonstrate comparable degree of variation with keratin and fat consistently producing somewhat more heterogeneous maps. This fact is notable considering that keratin and lipid are two primary components of the epidermis layer, which is known to be a highly heterogeneous structure in rats. 132

118 90 (A) (B) (C) (D) (E) (F) Figure III-8. Spatial distribution maps for the male rat sample #27, where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped skin slice. Black contours denote 10% change in regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement. 112

119 91 (A) (B) (C) (D) (E) (F) Figure III-9. Spatial distribution maps for the female rat sample #70, where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped skin slice. Black contour lines denote 10% change in regression coefficient. Two white circles represent the size of the optical fiber interface used in the related in vivo experiment. 112

120 92 For each of the measured components, the degree of heterogeneity is relevant in relation to the diameter of the fiber-optic probe, or the dimension of the incident nearinfrared radiation. These maps illustrate that moving the fiber-optic probe will result in the probing of skin tissue with different chemical compositions. Such variations in the chemical composition of the sample will certainly alter the nature of the underlying nearinfrared absorption spectra collected noninvasively. With a diameter of 1.8 mm, light exiting the optical fibers will encounter different distributions of water, protein, and fat, depending on the specific location of the fiber-optic probe. The degree of chemical heterogeneity is on the same length scale as other common noninvasive glucose measurements, including Raman spectroscopy and methods based on changes in the scattering properties of tissue. Indeed, as optical throughout is critical for all spectroscopic approaches, the distributions presented in Figures III-8 and III-9 clearly illustrate that many optical fiber-based methods may be sensitive to the exact location of the measurement interface on the skin. The degree to which this magnitude of chemical heterogeneity impacts noninvasive glucose measurements will be determined in a quantitative manner in Chapter V. Squeezing the tissue between sapphire windows does not necessarily destroy the skin structure but may expand all features. This compression effect explains why the more easily squeezed skin of female rats exhibits larger domains for all of the tissue components. Additionally, the near infrared spectra in this study are collected through all layers of the whole skin causing broadening of heterogeneous features because of photon scattering. 162 Multiple skin layers and protein fiber bundles present interfaces for light scattering events and decrease sensitivity of the measurement. One alternative to squeezing would be using microtome or separation procedures to divide skin into layers. If these layers are measured individually and then summed up, one cannot be sure that the maps would come from exactly the same volume of the whole skin. Even more significant is the fact that scattering properties of the whole skin sample

121 93 would be altered. Light propagation modeling in human skin has shown that a large portion of scattering takes place at the interfaces between layers due to difference in refractive indices of parts of dermis and epidermis. 91 This phenomenon is further supported by our findings described in Chapter IV of this thesis. Elimination of this original scattering component from the spectra will not allow to realistically simulate skin tissue absorption. Six Component Distributions for Rat Skin Heterogeneity and homogeneity have multiple definitions in the literature. Homogeneity is usually defined as the randomness of the distribution of a particular property. For the purpose of tissue characterization in this chapter, heterogeneity of skin tissue is based on two components: the property of interest of the skin and its variability. 163 The properties of interest are the regression coefficients for the six major absorbers over the combination near infrared spectrum. Variability refers to quantitative descriptors of the six component distributions within different animals, such as mean, median, and standard deviation. More in depth analysis of heterogeneity is certainly possible but has not been attempted to date. In this chapter, structural heterogeneity is addressed in the sense of the variability of a system property measured without reference to any functional effects. 163 In Chapter V functional heterogeneity is explored as the variability of a system property that can affect glucose concentration predictions. The box plot representation of spatial distributions of six major skin components was chosen as the preferred means to report these data because box plots enable side-byside comparisons of the separate data groups (i. e. animals). Also, box plots provide a number of valuable statistical characteristics of the distributions. 164 Box plots for water, collagen type I protein, and fat are presented in Figure III-10 for male rats and in Figure III-12 for female rats. Keratin protein, constant and slope term distributions follow in Figure III-11 for male rats and in Figure III-13 for female rats. Data for different rats are positioned along the horizontal axis of the box plots. Coefficients are ordered from the

122 Regression Coefficient for Fat Regression Coefficient for Collagen Regression Coefficient for Water 94 (A) Rat Number (B) Rat Number (C) Rat Number Figure III-10. Box plots for (A) water, (B) collagen type I protein, and (C) fat regression coefficients of male rats. Black circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent the values between 25% and 75% of the distribution.

123 Regression Coefficient for Slope Regression Coefficient for Constant Regression Coefficient for Keratin 95 (A) Rat Number (B) Rat Number (C) Rat Number Figure III-11. Box plots for (A) keratin protein, (B) constant, and (C) slope regression coefficients of male rats. Black circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent the values between 25% and 75% of the distribution.

124 Regression Coefficient for Fat Regression Coefficient for Collagen Regression Coefficient for Water 96 (A) Rat Number (B) Rat Number (C) Rat Number Figure III-12. Box plots for (A) water, (B) collagen type I protein, and (C) fat regression coefficients of female rats. Black circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent the values between 25% and 75% of the distribution.

125 Regression Coefficient for Slope Regression Coefficient for Constant Regression Coefficient for Keratin 97 (A) Rat Number (B) Rat Number (C) Rat Number Figure III-13. Box plots for (A) keratin protein, (B) constant, and (C) slope regression coefficients of female rats. Black circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent the values between 25% and 75% of the distribution.

126 98 smallest to the largest value along the vertical axis. Each rat group includes 400 data values from 4 skin maps collected from adjacent locations. In each box plot, red horizontal lines represent median values and distribution means are indicated by black circles. Other components of box plots are solid blue traces that include the values between the first (25%) and third (75%) quartiles of the sorted data. In this regard, median value is the second (50%) quartile. Whiskers extend from the boxes up to cover all the data within the 1.5 interquartile range from the lower and the larger quartiles. All the values higher or smaller than 1.5 interquartile range are considered outliers and noted with red plus signs. For the majority of cases, different median and mean values are observed for the same animal. This difference reveals a skew in the distributions from the perfect Gaussian shape or even the presence of two distributions within one. Because the data box for each animal is compiled from four separate samples collected at one time but measured on different days, changes in instrument performance, different pressures on the tissue, and different tissue thickness may contribute to a bimodal distribution or extreme values. Examples of this are fat distribution for male rat #1 in Figure III-10(C) and collagen type I distribution for female rat #4 in Figure III-12 (B), respectively. To support this observation, the fat and collagen data for these animals is plotted as histograms in Figure III-14 (A) and (B), respectively. Bimodal distributions are evident in both these plots. The groups of regression coefficients with higher values mostly originate from particular mapped samples out of the four skin samples characterized. The mean and standard deviation values for regression coefficients are summarized in Table III-4. Samples from the female rats have lower coefficients for all four chemical components: water, fat, and the two proteins. The reason for this is that skin from female animals is generally thinner and/or less rigid than from male rats. Thus, female rat skin is easier to compress and the resultant sample thickness is thinner.

127 Density Density (A) Regression Coeffficient for Fat 140 (B) Regression Coefficient for Collagen Figure III-14. Histograms for (A) fat distribution for male rat 1 and B) collagen type I protein distribution for female rat 4.

128 100 Table III-4. Summary of regression coefficients for rat skin samples. Regression coefficients, mean ± standard deviation Collagen Water Fat Keratin Constant Slope type I Rat 1 Rat 2 Rat 3 Rat 4 Rat 5 Rat 6 Rat 7 Rat 8 Average* ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± Male Rats Female Rats ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± Rat ± ± ± ± ± ± Rat ± ± ± ± ± ± Rat ± ± ± ± ± ± Rat ± ± ± ± ± ± Average* ± ± ± ± ± ± Note: * Average regression coefficients for species are reported with their pooled standard deviations.

129 101 Remarkably, the regression coefficients for fat are very low and in some cases the standard deviation for fat in a particular map is larger than its average value. This can be explained by the presence of negative values for the fat coefficient. Coefficients below zero are an artifact of the spectral fit. When no fat is present at a particular location, fitting the noise may result in a very small negative coefficient for fat at that location. If a sample is overall mostly depleted of fat, the number of negative values will be high and will make the standard deviation larger than the mean and/or median coefficient value for fat. For example, negative values are evident in samples from male rats 1, 3, 7, and 8 in Figure III-10 (C). The number of mapped locations with a negative coefficient for fat was very low in female rats. Relative Regression Coefficients (%) for Chemical Components In general, female skin samples were thinner and the coefficients for all of their chemical components were found smaller than for male rats. Additionally, in the course of the microspectroscopic measurements, the thickness of each sample was adjusted depending on its optical density. 129 These two factors explain why comparison between rats in terms of the absolute values of the regression coefficients may not be conclusive. Calculating relative coefficients for four chemical components according to Equation III-1 eliminated the influence of sample thickness or path length. These data are summarized in Table III-5. Small differences between male and female rats were found. Namely, female rats exhibited an overall mean of 44.5% for water and 39.9% for collagen protein. The male specimen had a lower relative water content of 40.1% and a higher relative collagen type I percentage of 45.5%. The elevated collagen content in male skin samples may explain why this skin is more rigid and harder to squeeze between the sapphire windows. The small difference in protein content between species is also consistent with literature reports (12). Keratin protein for both sexes was also slightly higher for male rats.

130 Table III-5. Relative regression coefficients (%) summary for rat skin samples. Relative regression coefficients (%), mean ± standard deviation Water Collagen type I Fat Keratin Male Rats Rat ± ± ± ± 1.3 Rat ± ± ± ± 1.5 Rat ± ± ± ± 1.3 Rat ± ± ± ± 1.6 Rat ± ± ± ± 2.3 Rat ± ± ± ± 1.5 Rat ± ± ± ± 1.8 Rat ± ± ± ± 1.7 Average* 40.0 ± ± ± ± 1.6 Female Rats Rat ± ± ± ± 1.2 Rat ± ± ± ± 1.1 Rat ± ± ± ± 2.0 Rat ± ± ± ± 1.9 Average* 44.5 ± ± ± ± 1.6 Note: * Average relative regression coefficients for species are reported with their pooled standard deviations. 102

131 103 From the start, we expected to see a higher percentage of fat in female skin. 135 However, this is not evident from Table III-5. For male rats #2 and #4, fat content is on the same level with the female animals. An important difference between sexes was that for female rats, fat percentage never went below 3%, while male samples could have either high of low fat percentage depending on the animal and/or location. It is unclear how much of this phenomenon comes from the incomplete removal of subcutaneous fat from the rigid male tissue. One of the central issues in noninvasive glucose measurements is the ability to create universal calibration models that will consistently provide acceptable predictions across animals and across days (2). Figures III-14 and III-15 give a clue as to why such a universal calibration model may be complicated by tissue heterogeneity. The first two groups outlined in green in these figures represent the relative chemical regression coefficients obtained by fitting noninvasive skin spectra collected in vivo over the course 5 days for two different male rats. The rest of the blue boxes correspond to the microspectroscopic data. In case of the in vivo experiment, the researcher took care to position the optical interface at the same contact location on the back of the rat s neck. Of course minimal realignments and variations in pressure were possible. Because the values in these box plots represent relative coefficients, the pressure related path length changes between measurements are eliminated from the data and the variations left are those due to physiological changes in skin matrix. Most notable changes are related to fat in the diet, instrumental variations, and relocation of the probe probe. Surprisingly, the spread and magnitude of the in vivo regression coefficients are similar to those measured from tissue heterogeneity for water, collagen type I protein, and fat. The average values for keratin protein in Figure III-15 (B) are smaller in the case of the in vivo measurement. The keratin component in the skin model represents epidermis. Consequently, the lower value for keratin protein may indicate a higher contribution to the spectra from the dermal layer of the skin. The relative contribution of

132 Relative Regression Coefficient, % Relative Regression Coefficient, % 104 (A) [1] [2] Rat Number (B) [1] [2] Rat Number Figure III-15. Box plots for (A) water and (B) collagen type I protein relative regression coefficients (%) of male rats. Black circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent the values between 25% and 75% of the distribution. The data between 25% and 75% from the in vivo experiments on two rats are represented by green boxes.

133 Relative Regression Coefficient, % Relative Regression Coefficient, % 105 (A) [1] [2] Rat Number (B) [1] [2] Rat Number Figure III-16. Box plots for (A) fat and (B) keratin protein relative regression coefficients (%) of male rats. Black circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent the values between 25% and 75% of the distribution. The data between 25% and 75% from the in vivo experiments on two rats are represented by green boxes.

134 106 epidermis and dermis may be different for the living rat and ex vivo tissue because the data from the former were collected through a fold of skin. Epidermis that is composed primarily from keratinous structures and lipids may be less affected by the strength of tissue squeezing. As the single skin slice inside the cell for the microscopic measurements may experience stronger compression, the dermis layer thins out and spreads to a larger degree and at the end contributes less to the total absorbance value of the skin, so that the relative contribution of the epidermis is increased. The shift of the dermis layer is visible in Figure III-2 (D) as the white tissue around the edges of the darker colored epidermis. Collection of the data across several days for a living female rat has not yet been attempted. Therefore, only ex vivo results are presented in Figures III-16 and III-17 for female animals. Essentially, the box plots provide graphic illustration to the findings discussed earlier and summarized in Table III-5. For example, results for female rat #4 demonstrate the advantages of reporting the relative regression coefficients. The absolute values plotted in Figures III-12 and III-13 indicated a skew in the distribution for water, proteins, and fat with a number of outliers on top of the upper quartile limit. Apparently, normalizing the data by the sum of the four chemical components reduced the effect of thickness variation for this particular animal. Even though the relative regression coefficients may be a convenient format to compare skin tissue heterogeneity, absolute values in Table III-4 will be used to mathematically synthesize multiple rat skin tissue spectra in Chapter V. Rat Skin ANOVA Data The coefficients for water, collagen type I and keratin proteins, and fat collected for each tissue slice were pooled to examine the tissue heterogeneity between animals. ANOVA evaluation was performed for the following two groupings: (1) variability between animals and (2) variability between locations for an individual animal.

135 Relative Regression Coefficient, % Relative Regression Coefficient, % 107 (A) Rat Number (B) Rat Number Figure III-17. Box plots for (A) water and (B) collagen type I protein relative regression coefficients (%) of female rats. Black circles represent mean values, red horizontal lines median values, red plus signs outliers, and blue box the values between 25% and 75% of the distribution.

136 Relative Regression Coefficient, % Relative Regression Coefficient, % 108 (A) Rat Number (B) Rat Number Figure III-18. Box plots for (A) fat and (B) keratin protein relative regression coefficients (%) of female rats. Black circles represent mean values, red horizontal lines median values, red plus signs outliers, and blue box the values between 25% and 75% of the distribution.

137 109 This analysis indicates that animal-to-animal variation for all four chemical components is significantly higher than variations between samples for a given animal or for variations between locations within a single tissue slice, with more than 99% probability (p values < 0.01). On the basis of these results, skin tissue samples from different animals must be treated as different distributions with significantly different values for both mean and standard deviation. ANOVA Gaussian distribution curves are summarized in Appendix Figure A-2 and A-3. Correlations between Skin Components Correlations between the different components can be investigated through an analysis of the collected data. Inspection of the distribution maps presented in Figures III- 8 and III-9 (A) and (D) reveals an inverse correlation between the distributions of water and fat. The darker shaded areas in the water distribution map (lower water content) correspond to the lighter shaded area in the fat map (higher fat content). A more complicated picture is revealed from a correlation analysis between all measured regression coefficients. In some cases, a strong correlation is obtained between two components, while the same components have limited or no correlation in other skin samples. Tables III-6 and III-7 provide an indication of the magnitude of each correlation for male and female rats, respectively. Both tables list the maximum, minimum, and median correlation coefficients as well as the standard errors in the maximum and minimum r values for each component correlation over all 48 tissue samples. In this dissertation, a strong correlation refers to the correlation coefficient absolute value larger than 0.5. Examples of strong positive (A), strong negative (B), and no correlation (C) are provided in Figure III-19. Values listed in Tables III-6 and III-7 indicate that the strongest correlations observed are negative and, in fact, most of the median correlation coefficients are negative. In almost every case, maximum or minimum r values for at least one slice of tissue can be found that suggests a strong correlation. For this reason, the median value gives a better

138 110 Table III-6. Summary of correlation coefficients for male rats. Components Maximum r Median r Minimum r Water vs. Collagen Type I 0.63 ± ± 0.02 Water vs. Fat ± ± 0.02 Water vs. Keratin 0.83 ± ± 0.07 Water vs. Constant 0.19 ± ± 0.04 Water vs. Slope 0.63 ± ± 0.08 Collagen Type I vs. Fat 0.85 ± ± 0.08 Collagen Type I vs. Keratin 0.65 ± ± 0.06 Collagen Type I vs. Constant 0.32 ± ± 0.07 Collagen Type I vs. Slope 0.38 ± ± 0.05 Fat vs. Keratin 0.09 ± ± 0.02 Fat vs. Constant 0.66 ± ± 0.1 Fat vs. Slope ± ± 0.04 Keratin vs. Constant 0.33 ± ± 0.06 Keratin vs. Slope 0.64 ± ± 0.10 Constant vs. Slope 0.12 ± ± 0.03

139 111 Table III-7. Summary of correlation coefficients for female rats. Components Maximum r Median r Minimum r Water vs. Collagen Type I ± ± 0.04 Water vs. Fat ± ± 0.02 Water vs. Keratin 0.61 ± ± 0.06 Water vs. Constant 0.55 ± ± 0.04 Water vs. Slope 0.52 ± ± 0.06 Collagen Type I vs. Fat 0.70 ± ± 0.1 Collagen Type I vs. Keratin 0.41 ± ± 0.08 Collagen Type I vs. Constant 0.53 ± ± 0.07 Collagen Type I vs. Slope 0.49 ± ± 0.05 Fat vs. Keratin ± ± 0.03 Fat vs. Constant 0.76 ± ± 0.1 Fat vs. Slope ± ± 0.05 Keratin vs. Constant 0.32 ± ± 0.09 Keratin vs. Slope 0.79 ± ± 0.1 Constant vs. Slope 0.72 ± ± 0.05

140 Regression Coefficient for Constant Regression Coefficient for Fat Regression Coefficient for Constant 112 (A) 0.8 Map 77 (Female Rat #4) r = / Regression Coefficient for Fat (B) Map 35 (Male Rat #4) r = / Regression Coefficient for Water (C) r = / Map 65 (Female Rat #1) Regression Coefficient for Collagen Type I Figure III-19. Examples of correlation plots for (A) strong positive, (B) strong negative, and (C) not significant correlation between rat skin tissue components. Red line on the plot is the linear fit to the data with the correlation coefficient reported in each case.

141 113 estimate of each correlation s consistency among different skin samples. In the future more rigorous analysis with a larger sampled population will be needed to identify outliers. Pronounced correlations (median magnitude > 0.5) that were obtained for both sexes are between water and fat, fat and keratin, as well as fat and the sloping baseline. The first two reflect the expected hydrophilic/hydrophobic interactions and corneocytes/lipids distributions within the tissue matrix. Correlation of slope with fat content may be more complicated as the sloping baseline term describes temperature dependence of water absorption as well as scattering properties of the tissue. Fat cells are indeed responsible for a large fraction of scattering in biological tissue. 101 For the female rats only water and collagen type I protein have a median correlation of This result may be due to the female skin being thinner and containing less protein than male skin. Conclusions Microspectroscopic methods are established to map the distribution of the major chemical components along two dimensions of 48 samples of excised male and female rat skin tissue. The resulting distribution maps clearly reveal chemical heterogeneity on a length scale that is similar to the dimensions used to collect skin spectra for noninvasive optical glucose measurements. Results show chemical heterogeneity with individual domains on the order of μm for water, collagen protein, keratin protein, and fat for both male and female rats. A similar structure is observed for the terms that characterize changes in the scattering properties of the tissue, as well as changes in the temperature of the sample matrix. An analysis of variance reveals that the animal-to-animal variation is significant and that the chemical variance cannot be treated as coming from a single distribution. Consistent correlations were noted between several of the chemical components. The data obtained serve as the basis for a mathematical simulation of realistic absorption properties of rat skin, for analysis of spectral features at various locations.

142 114 Future noninvasive glucose techniques must recognize the possible existence of this level of tissue heterogeneity. The implications of such heterogeneity are significant and cannot be ignored. Appropriate corrections are needed to diminish the impact of tissue heterogeneity on the accuracy of future noninvasive glucose measurements.

143 115 CHAPTER IV CHARACTERIZATION OF SPATIAL HETEROGENEITY OF HUMAN SKIN EX VIVO WITH NEAR-INFRARED MICROSPECTROSCOPY Introduction Optical Heterogeneity of Human Skin Tissue Human skin tissue has long been recognized by researchers as a complex heterogeneous multi-layered medium. 132, 144 Investigations of the absorption and scattering properties of human skin have been of great interest for the cosmetic and pharmaceutical industries, as well as for medical diagnostics. Among others, a number of microspectroscopic studies have been reported that aimed to characterize the distribution of water, 68 urea, 41, 165 hemoglobin, 69, 166, lipids 158, 167, 168, and glucose 13, 169 in human skin. It is universally accepted that light attenuation properties vary for the different layers of skin tissue because of differences in biochemical composition, particle types, shapes, and sizes, and host media. 108, 147, 166, 170 Scattering properties of human skin have received particular attention because of the importance of scattering and anisotropy coefficients in light propagation modeling. 131, 171 Light propagation models are typically applied to calculate path length and penetration depth of photons in tissue that enable quantification 90, 92 of analytes with noninvasive diffuse reflectance-based spectroscopic methods. The majority of the modeling studies of the optical properties of human skin assume that the skin layers are laterally homogeneous in their scattering and absorption properties and focus on the vertical heterogeneity of the tissue. For example, scattering properties of human skin were modeled between individual skin layers 91 and within each layer 170 but in either of these models horizontal spatial heterogeneity was not taken into account. Besides the assumption of lateral homogeneity, skin models have been published for a limited range of wavelengths. Usually attention is focused on the visible and overtone near infrared regions because these are important for tissue oxygenation,

144 116 sun-related damage, and cancer studies. 141 In this chapter, human skin tissue heterogeneity is characterized in terms of major absorbers in the combination near infrared ( cm -1 ) spectral region as it pertains to the noninvasive 10, 19 measurements of glucose. Morphologically human skin contains sub-millimeter size heterogeneous structures, such as hair follicles, hair shafts, wrinkles, and sebaceous glands, similar to rat skin. These structures may contribute to the heterogeneity picked up at the scale of several hundreds of micrometers. Images of dried human skin tissue collected with a biological microscope from the knuckle area from the back of a hand are featured in Figure IV-1 (A) and (B). Wrinkles are clearly seen transversing the image in Figure IV-1 (A) and captured in the center of Figure IV-1 (B). Surface wrinkles are associated with stacking of keratin-rich corneocytes in the lipid matrix of the stratum corneum. 172 Larger domains are formed from 10 of more columns of keratinocytes that manifest themselves as the wrinkles. 146 Such heterogeneities are typically more pronounced in rat skin. 132 Chapter Overview This chapter summarizes efforts to characterize lateral spatial heterogeneity for human whole skin and for separated epidermis and dermis human skin layers. Intraindividual differences in skin composition and structure comprise those between different locations on the human body, 92 age related changes, 149, 161, 173 and change brought about by disease, for example, diabetes. 174, 175 These, as well as the interindividual differences including gender-specific skin characteristics, are beyond the scope of the present investigation. The data set measured in this chapter demonstrates heterogeneity within the skin matrix over the combination region of the NIR spectrum. The six components linear regression model described and tested in Chapter II is used to measure the spatially distribution of water, collagen type I protein, fat, keratin, constant, and slope components within samples of human skin samples. In addition, correlation coefficients are calculated between the components of the whole skin and the skin layers.

145 117 (A) (B) Figure IV-1. Microscopic images of dried human skin from the knuckle of a hand. Images (A) and (B) show the skin with wrinkles visible. Data collected with a biological microscope (Carl Zeiss Microimaging, Thornwood, NY) in transmission mode. The instrument was calibrated with a micrometer (SpectraTech, Inc., Shelton, CT). The size of the red bar is 100 μm.

146 118 Experimental Procedures Preparation of Human Skin Samples Samples of human skin from four subjects were generously donated by Dr. Bickenbach at the Department of Biochemistry at the University of Iowa. These skin samples included Specimen 1 (88 y. o. female, skin from the thigh) and Specimen 3 (82 y. o. female, abdominal skin), Specimen 4 (57 y. o. male, location unknown), and Specimen 5 (19 y. o. male, skin from the leg). The samples were frozen in the culture medium immediately after the surgery and were stored frozen at -18 C. When a sample was received, it was thawed and thoroughly washed with phosphate buffer saline to remove traces of culture medium. The subcutaneous fat layer was removed with a surgical blade as completely as possible. After this, the sample was divided into smaller fragments of tissue, each measuring about 6 6 mm 2 and one of the smaller samples was measured with the microscope on the same day. The other prepared skin samples were frozen in phosphate buffer saline until needed. Before data collection, a sample was thawed if frozen and patted dry with a piece of Kimwipe to remove excess saline. The sample was then positioned between two sapphire slides of the compression cell described in detail in Chapter III. The metal screw was tightened until sufficient near infrared transmission signal was obtained. Four microspectroscopic maps for four adjacent skin fragments were accumulated per subject. Specimen 2 (96 y. o. male, skin from the knuckle area of both hands) was obtained from the Deeded Body Program at the Carver College of Medicine, University of Iowa. The sample was removed from the body of a male subject that had been kept frozen since death in Because of the nature and length of the preservation procedure, this sample was largely dehydrated. Skin samples from Specimen 2 had to be rehydrated for 1 hour before drying with Kimwipes and loading inside the compression cell as described above. Three maps were collected for Specimen 2. The demographics of the skin samples are summarized in Table IV-1.

147 119 Table IV-1. Demographics of human skin samples. Sample Age, y. o. Location Maps # Female Subjects Specimen 1 88 leg 82, 83, 84, 85 Specimen 3 82 abdomen 90, 91, 92, 93 Male Subjects Specimen 2 a 96 back of a hand 86, 88, 89 Specimen 4 b 57 unknown D1, D2, E1, E2* Specimen 5 19 leg 97, 98, 95, 96 Note: * Epidermis sample E2 was folded in half and the mapping was performed across the folded sample. a The specimen has been kept frozen since 2007; thawed and rehydrated before measurements. b The specimen was too thick to collect data with reasonable SNR and was separated into layers to yield samples of human dermis (D1, D2) and epidermis (E1, E2).

148 120 Specimen 4 was too thick to produce a detectable signal under the IR microscope. It was separated instead into dermis and epidermis by briefly placing it in a solution of phosphate buffer saline heated to 60 C. 176 Upon warming up, the whole skin sample split into the dermis and epidermis layers that were removed and blotted dry with a Kimwipe. The epidermis sample was cut into a smaller 6 6 mm 2 fragment (used for the collection of E1 map) and a larger 6 12 mm 2 fragment that was folded in half and measured inside the compression cell as E2 map (double epidermis). The dermis was divided into two equal 6 9 mm 2 rectangles that were measured as D1 and D2 maps. Instrumental Setup Human skin tissue samples were microspectroscopically measured with the same instrumentation as had been used for the rat skin samples. The details for the procedure are supplied in Chapters II and III. In short, a Nicolet Magna 560 FTIR spectrometer (Thermo-Nicolet Corp., Madison, WI) was used to collect maps in conjunction with an IR-Plan Advantage microscope (SpectraTech, Inc., Shelton, CT). Individual spectra were collected over the 4000 and 5000 cm -1 spectral range as 8k, double-sided interferograms. These interferograms were subsequently treated with a triangular apodization and standard Mertz phase correction. Similarly to the rat skin data, 128 scans were averaged per location on the sample. No external apertures were inserted in the beam path of the microscope. The use of the compression cell with two sapphire windows was needed to ensure a uniform sample thickness as described in Chapter III. The same cell served as a sample holder for the human skin. Extra care was taken not to displace the components of the folded epidermis sample while tightening the cell s ring-screw. The data for the map #E2 were only collected from the area that had two layers of the epidermis overlapping each other. Dehydration of the samples was slowed to 3 w/w%, as measured by weight loss after the mapping, by wrapping the circumference of the cell with a piece of Parafilm.

149 121 Spectral Data Collection After a tissue sample had been prepared and loaded into the compression cell, the sample was exposed to incident radiation from the microscope for 15 minutes. Exposure to the incident light prior to mapping was done to minimize temperature fluctuations in the sample during the mapping procedure. Reference air spectra were collected in triplicate with appropriate resolution prior to the skin measurements. An empty compression cell with a 0.92 mm thick Teflon spacer was used collecting these background air spectra. All of the measurements in this chapter, microspectra for air, whole skin, and skin layers, were performed at the room temperature that varied in the range between 21 and 24 C. Because the human skin maps were collected after the rat skin maps, the signal had been optimized by aligning the detector inside the microscope multiple times. A neutral density filter A was needed to prevent detector saturation when air spectra were collected. The filter was not used for collection of the tissue spectra. After the background spectra were collected and the temperature of the tissue sample had been equilibrated, the mapping of the tissue was performed as presented in Figure III-3. Identically to the mapping procedure for the rat skin, each spectral map was collected in a raster pattern with each step size of the stage being 480 μm in the X direction and 360 μm in the Y direction. The stage was controlled by mapping software OMNIC Atlμs (Thermo-Nicolet Corp.). The array corresponded to 100 evenly spaced locations across an area of mm 2. Every 30 minutes, three reference skin tissue spectra were collected at the central location of the measured skin sample. Microspectroscopic data were transferred to a Pentium V PC running the Windows Vista operating system. Air and tissue single-beam spectra covered the cm -1 ( μm) range with a nominal point spacing of 3.8 cm -1 and a resolution of 7.6 cm -1. MATLAB 7.0 (The Mathworks, Inc., Natick, MA) was used for all spectral and statistical analyses.

150 122 Data Processing Absorbance Spectra and Spectral Quality The root-mean-square (RMS) noise on 100% lines was calculated for air and tissue spectra in two adjacent regions: cm -1 and cm -1. A 100% line is obtained by taking the negative logarithm of the ratio between two back-to-back single beam spectra collected from the same location. The 100% lines for background air spectra and the tissue spectra collected from the central location for each map were fitted to a second order polynomial according to Equation II-1. For both air and the tissue samples, the average RMS noise level and the standard deviation were calculated for the three possible combinations of these single beam spectra (e.g., ratios of first-to-second, second-to-third, and third-to-first). Absorbance spectra for human skin tissue in each map were obtained by taking the negative logarithm of the ratio of the tissue single beam spectrum relative to the air background single beam spectrum calculated by averaging the three air spectra collected prior to the collection of the tissue map. Spectral Fitting and Two-Dimensional Tissue Maps Similar to the samples described in Chapters II and III, quantitative information about human skin tissue samples heterogeneity was obtained by fitting each skin absorbance spectrum to the set of six standard spectra of water, collagen protein, keratin protein, fat, constant offset, and slope, over the cm -1 spectral region as expressed in Equation II-6. For comparison between samples, weighted percent coefficients (i. e., relative regression coefficients) were calculated for each of the four chemical components, water, collagen type I protein, keratin protein, and fat, using Equation III-1. Absorbance spectra for the whole skin samples as well as dermis and epidermis skin layers were fitted with the same set of six major components. If a chemical component was not present at a particular location on a map, a coefficient value less than or equal to zero was expected to be produced by the regression.

151 123 Lateral heterogeneity of human skin tissue was demonstrated with the spatial distribution maps for the six tissue components in the regression fit. These were created with MATLAB software for an area of mm 2 as explained in Chapter III. Briefly, the density of yellow color (the lighter the shade the higher the value at a particular location) at each rectangular μm 2 pixel on the map corresponds to the average value of the four surrounding measured locations. In this manner, all four corners of a 2D map represent four corner locations of the area on the sample. Edges of the pixel rectangle appear smudged as the result of the MATLAB function employed. The contour plot helps to reveal the difference across the tissue location where the value for a given component either increases or decreases by 10% of the overall value range for that component across the map. 112, 152 The resultant spatial distribution maps for human skin and the skin layers contain both of these plots. Correlation between Tissue Components Consistent with the work on the rat tissue, Pearson correlation coefficients (r), were computed for all possible combinations of the six constituents for each spectral map. The regression coefficients calculated from the 100 spectra collected across each tissue slice were fitted to a line and the r coefficient was used to assess the strength of the correlation as given in Equation III-2. The number of possible pairs of the six components is 15. For each of the 15 pairs, if the r value approached unity, the respective components are perfectly correlated and their correlation plot is a line. If r is equal to 0, there is no correlation between the components and they change across the tissue sample independent of each other. The slope of the linear fit from the Equation III-2 determines the sign of the correlation. If the correlation is negative, the value for one component increases when the other decreases. As previously mentioned, this indication of how the components change relative to each other would be lost if the squared coefficients of determination r 2 were used. Standard errors for the r values were computed according to Equation III-3. These standard errors were used to identify cases when there was no

152 124 correlation between the components in the tissue map (i. e., the value for the r coefficient was smaller than its standard error). Results and Discussion Spectral Quality Assessment The experiments described in this chapter serve as the follow-up to the microspectroscopic characterization of rat skin tissue. The level of the spatial heterogeneity observed for rat skin, as summarized in Chapter III, will have important implications for noninvasive glucose sensing with the optical fiber interface in an animal model. For successful human applications of the noninvasive technology it is imperative to explore the lateral heterogeneity of human skin. For the initial trial, several samples of human whole skin were obtained from various locations as reported in Table IV-1. These were generally found thicker than rat skin, which prompt the use of the compression cell setup. The thickness of the skin samples after squeezing inside the compression cell were between 1.0 and 1.8 mm. Instrumental performance of the microspectroscopic system had been previously optimized for the measurements of the biological samples in Chapter II and the rat skin samples in Chapter III. An identical setup was used for the measurements of the human skin samples. The RMS noise levels for the air background spectra are listed in Table IV- 2 and the plots for the air RMS noise values follow in Figure IV-2. The values in Table IV-2 reflect the detector limited performance of the microscopic system and demonstrate that the performance was optimized for both cm -1 and cm -1 regions. In Figure IV-2 (A) and (B), the data are arranged in the order the maps were collected. The vertical black lines separate maps collected from specimens 1 (#82-85), 2 (#86-89), 3 (#90-93), 5 (#95-98), epidermis (E1, E2), and dermis (D1, D2) samples. Generally for the plots of RMS noise in this dissertation, red x-signs are used for the values calculated between the first and the second spectra in a triplicate, blue circles

153 125 Table IV-2. RMS noise values on 100% lines of air reference spectra. Specimen # 1 to 2 ± Std Dev a, μau 2 to 3 ± Std Dev a, μau 1 to 3 ± Std Dev a, μau Mean ± Std Dev*, μau cm -1 1 (female) 50.6 ± ± ± ± (female) 54.4 ± ± ± ± (male) 38.3 ± ± ± ± (male) 49.4 ± ± ± ± 9.5 epidermis a 55.1 ± ± ± ± 0.7 dermis a 58.6 ± ± ± ± cm -1 1 (female) 63.9 ± ± ± ± (female) 63.4 ± ± ± ± (male) 54.8 ± ± ± ± (male) 55.9 ± ± ± ± 9.3 epidermis a 55.5 ± ± ± ± 3.4 dermis a 66.1 ± ± ± ± 2.5 Note: * Pooled standard deviation for the mean value. a Single degree of freedom; 3 degrees of freedom for other specimens.

154 RMS Noise Level, microau RMS Noise Level, microau to 2 2 to 3 1 to 3 (A) E1 E2 D1 D2 Map Number (B) to 2 2 to 3 1 to E1 E2 D1 D2 Map Number Figure IV-2. RMS noise on 100% lines of air reference spectra in the regions between (A) 4,400-4,500 cm -1 and (B) 4,500 4,600 cm -1. The data are arranged in the order the maps were collected. The vertical black lines separate maps collected from specimen 1 (#82-85), specimen 2 (#86-89), specimen 3 (#90-93), specimen 5 (#95-98), epidermis (E1, E2), and dermis (D1, D2) samples.

155 127 denote values between the second and the third spectra, and the green plus-signs are for the RMS noise on the 100% lines obtained from the first and the third spectra. No sample-related or time-related trends were observed for the noise values. Slightly higher noise values were recorded for map #90 (female Specimen 3) in cm -1 region. The reasons for this were not clear and the elevation was attributed to the fluctuations in the alignment of the system. The detector was manually re-aligned prior to the collection of the map #91. Overall, the noise levels for air spectra were on the order of those for the higher signal group of rat skin maps (see Chapter III for more details). RMS noise values calculated from the reference spectra collected at the center of each tissue map reveal the limitations of the microspectroscopic system when applied to the measurements of biochemical components in the near infrared region. Table IV-3 and Figure IV-3 below summarize these values for all skin tissue samples. As a rule, the RMS noise for both male and female samples was between and μau, indicating that the RMS noise levels increase 10-fold compared to the corresponding values for air. This increase is, of course, not directly proportional to the change in optical throughput because an A neutral density filter (20%T) was removed prior to the collection of skin data whereas it was present for all of the air background spectra. As the human skin samples were typically thicker than the rat skin samples, the RMS noise values were on the order of those for the lower spectroscopic signal group for rats (Tables III-3 and 4) even though the system s performance had been optimized. Notable exceptions are samples for male Specimen 2, epidermis (E1, E2), and dermis (D1, D2) layer samples. First of all, the skin for the Specimen 2 maps was harvested from the body of a donor that had been frozen for more than two years and highly dehydrated. The age of the donor (96 y. o.) and the location of the skin (the knuckle area on the back of a hand) were other factors that contributed to the lower thickness of skin samples #86, 88, and 89. For these maps, the RMS values in both

156 128 Table IV-3. RMS noise values on 100% lines of human skin tissue spectra. Specimen # 1 to 2 ± Std Dev a, μau 2 to 3 ± Std Dev a, μau cm -1 1 to 3 ± Std Dev a, μau Mean ± Std Dev, μau 1 (female) ± ± ± ± (female) ± ± ± ± (male) a 66.3 ± ± ± ± (male) ± ± ± ± epidermis ± 8.0 double epidermis ± 5.3 dermis b ± ± ± ± cm -1 1 (female) ± ± ± ± (female) ± ± ± ± (male) a 68.4 ± ± ± ± (male) ± ± ± ± epidermis ± 4.7 double epidermis ± 7.1 dermis b ± ± ± ± 9.6 Note: * Pooled standard deviation for the mean value. a 2 degrees of freedom; 3 degrees of freedom for other whole skin specimens.

157 RMS Noise Level, microau RMS Noise Level, microau to 2 2 to 3 1 to 3 (A) * E1 E2*D1 D2 Map Number to 2 2 to 3 1 to 3 (B) * E1 E2*D1 D2 Map Number Figure IV-3. RMS noise on 100% lines of human skin tissue spectra in the regions between (A) 4,400-4,500 cm -1 and (B) 4,500 4,600 cm -1. The data are arranged in the order the maps were collected. The vertical black lines separate maps collected from specimen 1 (#82-85), specimen 2 (#86-89), specimen 3 (#90-93), specimen 5 (#95-98), epidermis (E1, E2), and dermis (D1, D2) samples. * Higher RMS noise values are observed for map #90 because of its higher collagen type I content and E2 map because it was collected across epidermis folded in half.

158 130 spectral regions are very low, around 65 μau. This value is very close to the RMS noise for single epidermis sample #E1 (average RMS level of 75.4 AU). It has been shown in the literature that RMS noise on 100% lines of aqueous samples is proportional to the thickness of the sample. 177 Higher RMS noise values for the spectra for a fold of epidermis #E2 (average of μau) are consistent with that trend. Both of the epidermis samples were found to transmit near infrared radiation better than the whole skin, primarily because they were thinner but also because they contained less water and collagen type I protein. When visually inspected, dermis samples (#D1 and #D2) appeared thicker and more rigid than the epidermis layer samples. The RMS noise values for these samples (average of μau) approached the values found for map #83 for the female Specimen 1. Notably, for the epidermis layer samples the RMS noise is consistently higher in the range between 4500 and 4600 cm -1 whereas for the dermis samples, larger values for noise are found in cm -1 region. This difference may be explained by the composition of the two layers. Epidermis is the upper skin layer, where keratins and lipids are abundant and water and collagen protein content is low. 158, 162, 167, 173, 178 On the contrary, water and collagen protein are the two major constituents of the dermis layer. 131, 141, 143, 149, 178 The standard spectrum of the collagen type I protein in Figure II-3 features higher absorbance values between 4460 and 4400 cm -1 than between 4500 and 4600 cm -1. Keratin protein absorbance is very similar for both regions. That is why the samples of dermis rich in collagen type I transmit less radiation in the region with lower wavenumbers. In fact, this observation is true for some of the whole skin samples as well. For example, Specimen #3 in Figure IV-3 (A) and (B) was found to produce larger regression coefficients for collagen protein than the other female Specimen #1. The RMS noises for Specimen #3 are the highest of those reported in this chapter but most importantly, the noise is higher for the region of cm -1 as seen in Table IV-3 even though the noise for the corresponding air spectra is not. For the maps of Specimen #1 the trend is

159 131 opposite with cm -1 being the region with more noise. Male Specimen #5 was found to be similar to dermis and Specimen #3 in this respect. Higher relative collagen type I protein content in the tissues of Specimen #5 may be related to the younger age (19 y. o.) of the human donor. 149, 179. On the whole, the RMS noise levels for human whole skin tissue samples were found similar to the lower signal group values for rats in Chapter III. As previously demonstrated, these levels of noise, though not low enough for successful measurement of glucose in skin, are sufficient for quantification of the bulk tissue components with the use of the six component linear regression model. The lower noise values for epidermis and dermis samples are consistent with higher transmission values for these tissues. Human Skin Tissue Absorbance Data Visual Inspection of Absorbance Spectra For the purposes of testing the robustness of multivariate calibration models for noninvasive measurements of glucose, samples from the back of a hand of a human subject, the actual location of the optical interface, would have been ideal. Only one such sample was obtained in its frozen state. Other samples came from different locations on the body. That is why in this initial set of five microspectroscopic experiments with human skin, the results from different donors were not pooled into the same distribution but instead were treated separately. The differences in the origin of the tissue may be partly responsible for the differences in the shape of the absorbance spectra illustrated in Figures IV-4 and IV-5 for male and female whole skin respectively. In Figure IV-4 (A) and (B) 100 absorbance spectra in the combination near infrared region of cm -1 are presented for the representative maps for female specimens: map #82 for Specimen 1 and map #90 for Specimen 3; (C) and (D) are the corresponding reference absorbance spectra at the central location on the skin slices; and (E) and (F) are the absorbance residuals for these maps after the six-component fit.

160 Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU 132 (A) (B) Wavenumber, cm Wavenumber, cm (C) (D) Wavenumber, cm Wavenumber, cm (E) 0.1 (F) Wavenumber, cm Wavenumber, cm Figure IV-4. Spectral data for whole skin samples from female subjects: (A) and (B) are the absorbance spectra for point maps from samples #82 and #90, respectively; (C) and (D) are the corresponding reference spectra at the central location on these skin slices; and (E) and (F) are the corresponding absorbance residuals for these maps after the six-component fit.

161 Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU 133 (A) (B) Wavenumber, cm Wavenumber, cm (C) (D) Wavenumber, cm Wavenumber, cm (E) (F) Wavenumber, cm Wavenumber, cm Figure IV-5. Spectral data for whole skin samples from male subjects: (A) and (B) are the absorbance spectra for 100 point maps of samples #88 and #95, respectively; (C) and (D) are the corresponding reference spectra at the central location on the skin slices; and (E) and (F) are the corresponding absorbance residuals for these maps after the six-component fit.

162 134 The general shape of the absorbance spectra for Specimen 1 and 3 samples in Figure IV-4 (A) and (B) is similar to that of rat s skin (see Figure III-4). The well-shaped overall profile of the combination spectra suggests the presence of water in the tissue. Absorbance peaks for protein are also present at 4650, 4400, and 4890 cm -1. For some spectra, lipid peaks can be seen at 4330 and 4255 cm -1. These prominent lipid signals most probably originate from the incomplete removal of subcutaneous fat and are not present for all spectra of the maps. For instance, lipid peaks are not present in the reference absorbance spectra collected from the central location in Figure IV-4 (C) and (D). The major difference between samples #82 and #90 is in the magnitude of absorbance; it is noticeably larger for sample #90. Also, the slope of the spectral shoulder between 4460 and 4400 cm -1 is greater for map #90 consistent with a larger amount of collagen in that sample. After the six component linear fit of the map data, unmodeled spectral features are concentrated in the spectral residuals shown in Figure IV-4 (E) and (F). Similarly to the rat skin absorbance data, the linear fit for the human skin accounts for 95 to 99%, depending on the wavenumber, of the absorbance signal and the other 5 to 1% of spectral information are revealed in the residuals. There is a marked difference for the two maps in the relative size of the residual peaks related to proteins (4450 and 4650 cm -1 ) and lipids (4330 and 4255 cm -1 ): for map#82 the lipid bands are sharper and more intense whereas for map #90 protein and lipid peaks are close in intensity. This can be explained by lower fat content in map #90 provided by a more thorough removal of the subcutaneous tissue from that sample. Spectral residuals also contain all of the random noise from the tissue spectra. The RMS noise values were larger for map #90 as seen in Figure IV-3. Consequently, higher noise is evident in the spectral residuals presented in Figure IV-4 (F). Spectral data in the combination NIR spectra for whole skin samples collected from male human donors are represented in Figure IV-5. The same lay out is used to

163 135 present the data as in Figure IV-4 for female donors. The representative maps of the male specimens are map #88 for Specimen 2 and map# 95 for Specimen 5. These two specimens are very different in nature: firstly, they were harvested from different locations on the body (back of a hand for Specimen 2 and a leg for Specimen 5), secondly, the age of the human donors varied (96 y. o. for Specimen 2 versus 19 y. o. for Specimen 5), thirdly, the samples for Specimen 2 have be frozen for a long time and are largely dehydrated. These differences help to explain the variations in skin absorbance spectra seen in Figure IV-5 (A) and (B). First of all, the absorbance is higher for the skin sample #95 collected from the leg of a younger male donor. Also, the usual well-shape of the skin spectra is very shallow in map #88 because of the low water content. Most of the water measured in that sample resulted from re-hydration of the skin tissue in the saline solution before mapping. On the contrary, the male sample #95, freshly preserved after surgery, has absorbance spectra representative of a strong water content. Apart from water absorbance, both of the samples demonstrate broad bands for protein consistently present for every location on these maps. Fat-related peaks at 4330 and 4255 cm -1 are only visible for some of the locations and cannot be seen in Figure IV- 5 (C) and (D) for the central reference location on the tissue slices. In agreement with the data for the female human skin, the variance in the spectra across all wavenumbers is about 10 times larger for the 100 location maps than for the reference skin spectra recorded at the single central location. This spatial heterogeneity of the human skin components will is discussed below. The unmodeled absorbance residuals for male human skin samples are plotted in Figure IV-5 (E) and (F). Larger random noise is present in spectra for map #95 in keeping with the larger thickness and higher collagen type I and water content of that sample. Additionally, the relative height of the residual features in the combination region for C-H vibrations characteristic of lipid molecules is higher for map #88. Once again, that sample was received frozen and because of its fragility was not completely

164 136 stripped of the subcutaneous fat, which resulted in a larger number of locations with a measurable amount of fat still present. The samples for Specimen 5 were cleaned underneath the dermis layer to ensure removal of the fatty tissues, which explains the lower relative impact of the fat on the shape of the residuals in Figure IV-5 (F). Six Component Regression Fit The absorbance spectra for the maps described in the previous section were linearly fitted at each location to the six major components of water, collagen type I protein, fat, keratin protein, constant, and slope, following Equation II-2. Coefficient profiles obtained are plotted in Figure IV-6 for the female and in Figure IV-7 for the male representative maps. In these figures, (A) and (B) are the regression coefficients for the 100 points of the human whole skin maps, and (C) and (D) are the regression coefficients for spectra collected every 30 minutes at the central location on these skin slices. These plots express the heterogeneity of the skin samples reflected in the absorbance data in terms of numerical values for the six regression coefficients. The relative differences in composition between different skin samples can be clearly seen because the coefficients are plotted on the same scale. The two major absorbers of near infrared radiation for all of the whole skin samples are water and collagen type I protein. As mentioned earlier, for maps #90 and 95 collagen regression coefficients are larger than those for water at all points on the map, including the central reference location. For maps #82 and 88, the fractions of water and collagen are very similar. The next largest regression coefficient is for the constant term. This observation is different from the data obtained for rat skin, where the constant had the largest relative values among the six components. The constant term is used to model the wavenumberindependent scattering component in the tissue. Apparently, a part of its value is dependent on the properties of the sample setup and microscope optics, as a fraction of light is lost when light travels across air/sapphire and sapphire/tissue interfaces. Another

165 Regression Coefficient Regression Coefficient Regression Coefficient Regression Coefficient 137 (A) (B) 2 water collageni fat keratin constant slope 2 water collageni fat keratin constant slope Mapping Location Number Mapping Location Number (C) (D) 2 water collageni fat keratin constant slope 2 water collageni fat keratin constant slope Reference Spectrum Number Reference Spectrum Number Figure IV-6. Representative regression fit coefficients for whole skin samples from female subjects: (A) and (B) are the results of regression fit for the point maps of samples #82 and #90, respectively; and (C) and (D) are the corresponding regression coefficients at the central location on the skin slices.

166 Regression Coefficient Regression Coefficient Regression Coefficient Regression Coefficient 138 (A) (B) 2 water collageni fat keratin constant slope 2 water collageni fat keratin constant slope Mapping Location Number Mapping Location Number (C) (D) 2 water collageni fat keratin constant slope 2 water collageni fat keratin constant slope Reference Spectrum Number Reference Spectrum Number Figure IV-7. Representative regression fit coefficients for whole skin samples from male subjects: (A) and (B) are the results of regression fit for the point maps of samples #88 and #95, respectively; and (C) and (D) are the corresponding regression coefficients at the central location on the skin slices.

167 139 consideration is the inter-species variations in skin composition (apart from the subcutaneous layer) between rats and humans: human skin has lower fat content than rat. 132 Positive correlation was found for male and female rat samples (see Figure III-18) between the fat and constant terms. This may contribute to the differences in the relative magnitude of the constant between rat and human skin samples. The third consideration would be the pressure applied on the samples inside the compression cell. Since the rat tissue was thinner than human, less pressure was needed to achieve a detectable signal. The extra pressure applied to the human skin samples might have changed the internal skin structure, possibly affecting tissue scattering properties. 156 Smaller coefficients were obtained for keratin and fat. Keratin protein content is demonstrated to be much lower than collagen protein for all maps, except map #88 (96 y. o. donor), where the values for keratin are about 30% of the collagen type I values. Coefficients for keratin were found positive for all locations on all of the whole skin maps. Fat coefficients, however, were often negative, which indicated the absence of fat at those mapped locations. A number of occasionally positive values for fat can be seen in all the representative maps in Figures IV-6 and IV-7 (A) and (B). These are probably due to leftover spots of fatty subcutaneous tissue not removed prior to the measurement with a surgical blade. Figures IV-6 and IV-7 illustrate the degree of spatial heterogeneity of the human skin tissue. For all of the components, variations due to the change of the measurement location were about 10 times larger than the temporal variation at the central location on the tissue demonstrated in (C) and (D) plots in Figures IV-6 and IV-7. Human Skin Spatial Heterogeneity Spatial Contour Maps Identically to the rat skin presented in Chapter III, mm 2 areas of human whole skin were characterized by collecting spectra in a raster pattern from 100 locations on the tissue slice. Distribution maps are presented in Figures IV-8 and IV-9 for the

168 140 (A) (B) (C) (D) (E) (F) Figure IV-8. Spatial distribution maps for sample #82 from a female subject, where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped skin slice. Black contours denote 10% change in the regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement. 38

169 141 (A) (B) (C) (D) (E) (F) Figure IV-9. Spatial distribution maps for sample #95 from a male subject, where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped skin slice. Black contours denote 10% change in the regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement. 38

170 142 representative tissue slices #82 for female Specimen 1 and #95 for male Specimen 5. The degree of tissue spatial heterogeneity can be assessed by evaluating spatial distribution maps for the six components. As previously described, these distribution maps plot a saturation of orange color at each pixel to represent the value average of the four surrounding locations on the map. Black contour lines define domains where the difference across the border either increases or decreases by one-tenth, or 10%, of the total change in that coefficient for the map. In some distribution maps, a pair of white circles is included that corresponds to the size of the optical fibers used in the noninvasive spectra collection. 38 The fiber diameter used for the computations in this dissertation is 1.8 mm. The most pronounced difference between the distribution maps for rat and human skin is in the size of tissue domains of the six components. These are generally around 500 and 800 μm for rat samples and increase to around a 1000 μm for human skin. An example of such a domain of high keratin protein content can be observed as a light yellow shape about 1 mm in length in Figure IV-8 (D). The size of the tissue domains is a reflection of the relative tissue heterogeneity because the black contour lines divide the areas with 10% change in a given coefficient. The increase in the domain size can be related to the fact that human skin samples were thicker than the rat skin. As the light was passing through the layers of human skin, photons had experienced more scattering events and the resulting structures could be blurred to a greater degree than similar structures in maps from the thinner rat skin. In regards to the fat distribution, plots (C) in Figures IV-8 and IV-9 clearly display the locations of the residual fat from the incompletely removed subcutaneous tissue as oval domains of lighter color. The majority of the locations on maps #82 and #95 were devoid of this extra subcutaneous fat and hence produce very small or even negative fat coefficients. This is in accordance with the concentration of lipid-like 167, 172, molecules in the epidermis and the upper-most layer of the skin, stratum corneum,

171 , 181 that are only about 7% of the human whole skin (dermis and epidermis) in thickness. 147 The coefficient for the fat contained in these top skin layers would be very low indeed and these are shown at the locations other than the residual fat areas. Lipid content of human skin is on the lower side among several species and is considerably lower, in relative terms, than the lipid content of rat skin. 158, 167 Larger relative coefficients for fat were found for rats in Chapter III than for these human samples characterized in this chapter. Even with the larger size of the tissue domains for the six components, the degree of heterogeneity found in these samples of human skin can potentially affect the transcutaneous measurements of glucose. For instance, let us consider the white circles representing two locations of an optical fiber in Figure IV-9 (A). These demonstrate the fiber shifting about 1 mm to the side. Apparently, the composition of the water changes with this shift from 0.77 to 0.87 (13% increase). In the multivariate model for glucose, the sugar is assumed to be present in the aqueous medium only. Such a change in the aqueous path length in the skin with a slight movement of the interface could have a large effect on the glucose prediction. The influence of the rat and human tissue heterogeneity on the PLS models for glucose is investigated in Chapter V. Six Component Distributions for Human Whole Skin In all, 12 spectral maps were collected from 12 skin slices associated with Specimens 1, 3, and 5. This number corresponds to 4 maps, or 400 individual locations on each skin sample, per specimen. Specimen 2, harvested from the donor s body that had been frozen from 2007, only produced three skin maps with acceptable quality of the spectroscopic signal, so the number of spectra fitted with the linear regression for Specimen 3 is only 300. The results of the six component fit are plotted in Figure IV-10 (A), (B), and (C) for water, collagen type I protein, and fat, respectively. Figure IV-11 follows with the boxplots for (A) keratin protein, (B) constant offset, and (C) slope coefficients. In both of the figures, magenta boxes define the range between 25% and

172 Regression Coefficient for Fat Regression Coefficient for Collagen Regression Coefficient for Water 144 (A) Specimen Number (B) Specimen Number (C) Specimen Number Figure IV-10. Box plots for (A) water, (B) collagen type I protein, and (C) fat regression coefficients of whole skin samples. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by magenta (female specimens) and blue (male specimens) boxes.

173 Regression Coefficient for Slope Regression Coefficient for Constant Regression Coefficient for Keratin (A) Specimen Number (B) Specimen Number (C) Specimen Number Figure IV-11. Box plots for (A) keratin, (B) constant, and (C) slope regression coefficients of human whole skin samples. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by magenta (female specimens) and blue (male specimens) boxes.

174 146 75% of the values for the female specimens and blue boxes do the same for the male specimens. Black circles represent mean values, red horizontal lines median values, and red plus signs outliers. The largest coefficients are calculated for water and collagen type I protein for all samples. For Specimens 3 and 5, the collagen coefficients are unvaryingly larger than water coefficients. Specimen 5 was collected from the leg of a , 182 y. o. male donor. As the collagen content of skin is known to decrease with age, higher values of collagen are expected for younger human subjects. However, Specimen 3 that also exhibits elevated collagen levels was obtained from an older female donor. Unfortunately, the size of the sample population is too small to make definitive conclusions about relationship between collagen content in skin and the parameters of age and gender. Generally, the values for the coefficients are larger than those found for rat skin as can be explained by the greater thickness of the human samples. The notable exception to this is featured in plot (C) in Figure IV-10, which contains the data for fat. The majority of the rat data for fat are positive values of low magnitude while the data for human samples are predominantly at or below zero with a large number of positive values rendered as outliers in the boxplots. This result is in line with previously explained origins of the fat signal in the human maps from the presence of residual subcutaneous tissue. Remarkably, values for the constant offset for human skin samples are very close to those for male and female rats reported in Figure III-13. Considering that the constant accounts for changes in scattering properties of the tissue, the similarity of the term across the species suggests that the origins of the scattering are similar for rat and human samples. The assumption here is that the scattering between the particles within the layers of the skin, that would depend on the thickness of these layers, has less effect on the magnitude of the scattering coefficient. In other words, the interlayer scattering events that occur when radiation travels through the interfaces of media with different refractive

175 147 coefficients (i. e. air/sapphire, sapphire/epidermis, epidermis/dermis, and dermis/sapphire) constitute the main part of the overall sample scattering and are independent of either the species or the tissue thickness. Relative Regression Coefficients (%) for Chemical Components The absolute values for regression coefficients for water, collagen type I and keratin proteins, and slope terms appear to follow a pattern for the specimens in Figures IV-10 and IV-11: Specimens 1 and 2 producing lower values for all of the four components and Specimens 3 and 5 higher values. To test if this trend is simply due to variations in thickness, relative (%) regression coefficients for the chemical components were computed for water, collagen type I, keratin, and fat as described in the Experimental section. The results are plotted in Figures IV-12 and IV-13. Both absolute and relative regression coefficients are reviewed in Table IV-4 where the mean values for each coefficient are reported when applicable, together with the standard deviation values that provide an idea of the variability in skin composition within a specimen. Surprisingly, the pattern between the specimens holds up in just one case, for collagen protein. It is safe to conclude that the relative amount of collagen type I protein is indeed larger for Specimens 3 and 5 (57.2% and 58.6%) irrespective of the sample thicknesses. The trend in the relative water content is reversed from the absolute values, with Specimens 1 and 2 containing a higher relative percentage of water (47.6% and 45.6%). The reversal indicates that water contribution in the spectra varies proportionally to the thickness of the sample. The samples of Specimen 3 and 5 simply have longer aqueous path lengths because they are thicker even though their relative water content is lower (35.8% and 33.6%). A plot of the relative coefficients for fat in Figure IV-13 (A) supports the contamination of the skin samples with the subcutaneous fat tissues. Typical fat content in combined human dermis and epidermis is about 2 to 4%. 165, 171 The outliers, shown in red, have values from 5 to 30% of fat that are too high and cannot be explained by the fat present in human whole skin tissue.

176 Relative Regression Coefficient, % Relative Regression Coefficient, % 148 (A) Specimen Number (B) Specimen Number Figure IV-12. Box plots for (A) water, and (B) collagen type I protein relative regression coefficients (%) of human whole skin samples. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by magenta (female specimens) and blue (male specimens) boxes.

177 Relative Regression Coefficient, % Relative Regression Coefficient, % 149 (A) Specimen Number (B) Specimen Number Figure IV-13. Box plots for (A) fat, and (B) keratin protein relative regression coefficients (%) of human whole skin samples. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by magenta (female specimens) and blue (male specimens) boxes.

178 150 Table IV-4. Regression coefficients summary for human whole skin samples. Collagen Water Fat Keratin Constant Slope type I Specimen 1 Specimen 3 Specimen 2 Specimen ± ± ± ± Regression coefficients, mean ± standard deviation ± ± ± ± 0.12 Female Specimens ± ± ± ± Male Specimens ± ± ± ± ± ± ± ± ± ± ± ± Relative regression coefficients (%), mean ± standard deviation Female Specimens Specimen 1 Specimen ± ± ± ± ± ± ± ± Male Specimens Specimen 2 Specimen ± ± ± ± ± ± ± ±

179 151 Figure IV-13 (B) demonstrates the uniqueness of samples collected from Specimen 2. These skin slices produced the highest percent coefficients for keratin (12.0%) among all specimens. Moreover, the ratio of keratin to collagen type I protein is 0.31 for Specimen 2 and less than 0.22 for all other specimens. Depletion of collagen pool with age and UV-exposure is a phenomenon well-documented in the literature for human skin. 144, 149, 161 The greater relative content of keratin in the skin obtained from the back of a hand of the oldest donor would be expected based on skin aging research. Correlations between Components of Whole Skin Two-dimensional plots in Figures IV-8 and IV-9 demonstrate how not all of the skin tissue components change independently. Take, for instance, spatial distribution for water and keratin in Figure IV-8 (A) and (D). Clearly, the domains of high water content shown in light yellow color approximately correspond to the domains of low keratin content shown in dark orange shade. Even though the spatial domains for human skin are usually larger than for rat skin, correlations between various components exist. In this work, strong correlation between two components is assumed if the Pearson correlation coefficient, r, is larger than 0.5. No correlation is present when the value of r coefficient is lower than its standard error. Examples of a strong positive (A), a strong negative (B), and no correlation (C) between human skin components follow in Figure IV-14. Median, maximum, and minimum values for r are provided in Table IV-5 for female specimens and Table IV-6 for male specimens of human skin. Consistently strong positive correlations were obtained from most of the female skin samples for water/keratin and collagen/keratin, strong negative correlations for fat/keratin and fat/slope pairs. Strong correlations have not been seen in rats for protein terms with other components of tissue. Another difference from rat samples is the presence of strong positive correlations. The same pairs of components are also strongly correlated for Specimen 5. In addition, a strong negative correlation was found for fat and water as expected. Samples for Specimen 2 generally exhibit weaker correlations between components. The only two

180 Regression Coefficient for Constant Regression Coefficient for Fat Regression Coefficient for Slope Map 98 (Specimen #5) (A) r = / Regression Coefficient for Water (B) Map 92 (Specimen #3) r = / Regression Coefficient for Water (C) 0.7 Map 97 (Specimen #5) r = / Regression Coefficient for Water Figure IV-14. Examples of correlation plots for (A) strong positive, (B) strong negative, and (C) no correlation between human whole skin tissue components. Red line is the linear fit to the data with the correlation coefficient reported in each case.

181 153 Table IV-5. Summary of correlation coefficients for human whole skin samples collected from female Specimens #1 and #3 (± standard error). Components Specimen 1 Specimen 3 r (max) r (median) r (min) r (max) r (median) r (min) Water vs. Collagen Type I ± ± ± ± 0.08 Water vs. Fat ± ± ± ± 0.02 Water vs. Keratin 0.64 ± ± ± ± 0.09 Water vs. Constant 0.63 ± ± ± ± 0.03 Water vs. Slope 0.48 ± ± ± ± 0.02 Collagen Type I vs. Fat 0.36 ± ± ± ± 0.06 Collagen Type I vs. Keratin 0.86 ± ± ± ± 0.10 Collagen Type I vs. Constant ± ± ± ± 0.01 Collagen Type I vs. Slope 0.60 ± ± ± ± 0.08 Fat vs. Keratin ± ± ± ± 0.01 Fat vs. Constant 0.26 ± ± ± ± 0.10 Fat vs. Slope ± ± ± ± 0.05 Keratin vs. Constant ± ± ± ± Keratin vs. Slope 0.85 ± ± ± ± 0.05 Constant vs. Slope 0.43 ± ± ± ± 0.06

182 154 Table IV-6. Summary of correlation coefficients, r, for human whole skin samples collected from male Specimens #2 and #5 (± standard error). Components Water vs. Collagen Type I Water vs. Fat Water vs. Keratin Water vs. Constant Water vs. Slope Collagen Type I vs. Fat Collagen Type I vs. Keratin Collagen Type I vs. Constant Collagen Type I vs. Slope Fat vs. Keratin Fat vs. Constant Fat vs. Slope Keratin vs. Constant Keratin vs. Slope Constant vs. Slope Specimen 2 Specimen 5 r (max) r (median) r (min) r (max) r (median) r (min) 0.29 ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± 0.04

183 155 strong negative correlations were found for collagen/constant and fat/slope. Water content is poorly correlated with any of the other components, which could be a result of uniform re-hydration of the skin samples from Specimen 2. In case of both collagen and keratin proteins, larger correlation coefficients were typically calculated for the samples that contained more of these proteins, such as Specimens 3 and 5. However, the physiological reason for this observation is not clear. It should be reinforced that the number of human skin samples analyzed in this chapter is much smaller than the sample size for the rat skin in Chapter II. With this in mind, one would expect that some of the mentioned strong correlations may not hold once more human skin samples are investigated. Human Skin Layers Tissue Absorbance Data Visual Inspection of Absorbance Spectra Apart from the collection of data from the human whole skin samples, microscopic measurements of a single epidermis layer, folded epidermis, and dermis from male Specimen 4 were performed. The epidermis and dermis layers in the skin have been shown to possess distinctive absorption and scattering properties in the visible and near infrared regions of the spectrum. 170 These differences in light extinction originate from unique structure and composition of both the epidermis and dermis layers. First of all, the water content is very low in the epidermis, especially in the top-most layer of the epidermis, stratum corneum, whose principal function is to serve as a selective water 146, 168, 172, 183 barrier and whose major components are lipids and keratinized structures. Lower epidermis layers also contain a large percentage of keratin protein in addition to melanins. 147 The size of the particles (keratin-producing cells, melanocytes) increases for the lower epidermis compared with the stratum corneum 147 causing the change in its refractive index and scattering properties. 170 When light travels from the epidermis to the dermis layer, a large change in the refractive coefficient is observed. Also, the average size of the scattering particles increases as does the content of water and collagen type I

184 156 protein. 170 These two components happen to be the major absorbers of the near infrared radiation in the dermis layer. To our knowledge, no studies have been published yet that characterize the lateral spatial distribution of the major absorbers of the near infrared light in the combination region ( cm -1 ) for the individual layers of human epidermis and dermis. The absorbance microspectra of 100-point maps are plotted in Figures IV-10 for a single epidermis layer and double layer of epidermis, and in Figure IV-11 for two samples of human dermis layer. The layers were obtained by gently placing the slice of whole skin for about 2 minutes in water heated to 60 C. This method of separation has been reported in literature to cause minimal disturbance to the composition and structure of both dermis and epidermis, apart from dissolution of some water-soluble components. 176 In Figure IV-10 spectral data for the epidermis samples E1 and E2 reveal light attenuation properties sixteen times weaker than those measured for whole skin. The characteristic shape of the absorption spectrum must originate from the water the epidermis was immersed in for separation. This background of residual water is approaching a flat line for a single layer of dermis. The intensity of the broad vibrational bands for proteins in the cm -1 region are weak and comparable to the peaks at 4330 and 4255 cm -1 for lipids. Absorbance signal for the epidermis folded in half is about twice greater than for the signal layer and the degree of variation in the spectral offset also increases. When an epidermis sample is folded, an extra interface for scattering between layers of tissue is introduced hence the coefficients for the scatteringrelated constant term are larger. The magnitude of the spectral residuals obtained after fitting increases for the folded epidermis but the shape of the residual spectra remains the same. The shape of the nonrandom residuals for the epidermis and dermis layers of the human skin maps is dissimilar to that for whole skin. Protein peaks represent the predominant bands in

185 Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU 157 (A) (B) Wavenumber, cm Wavenumber, cm (C) (D) Wavenumber, cm Wavenumber, cm (E) (F) 8 x x Wavenumber, cm Wavenumber, cm Figure IV-15. Spectral data for epidermis samples: (A) and (B) are the absorbance spectra for 100 point maps of sample #E1 and sample #E2 (folded epidermis), respectively; (C) and (D) are the corresponding reference absorbance spectra at the central location on the sample; (E) and (F) are the corresponding absorbance residuals for these maps after the sixcomponent fit.

186 Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU Absorbance, AU 158 (A) (B) Wavenumber, cm Wavenumber, cm (C) (D) Wavenumber, cm Wavenumber, cm (E) (F) Wavenumber, cm Wavenumber, cm Figure IV-16. Spectral data for dermis samples: (A) and (B) are the absorbance spectra for 100 point maps of sample #D1 and sample #D2, respectively; (C) and (D) are the corresponding reference absorbance spectra at the central location on the sample; (E) and (F) are the corresponding absorbance residuals for these maps after the six-component fit.

187 159 the human epidermis residual spectra. For example, the strong unmodeled residual peak is present at 4650 cm -1 in Figure IV-15 (E) and (F) for both E1 and E2 samples. This 49, 52 band is located in the region of amide, N-H, combination vibrational modes. Dermis spectral data presented in Figure IV-16 for samples D1 and D2 are distinctly different from the epidermis absorbance spectra. A strong water background is the most notable feature of these spectra. The magnitude of light attenuation is twelve times larger than for the epidermis layer. The second most obvious difference is the sharper peaks for collagen type I around 4400 cm -1 and 4650 cm -1. Sharper spectral bands associated with lipids are mostly absent from these spectra as can be clearly seen in Figure IV-16 (C) and (D) for the central locations on the tissue slices. The unmodeled residuals for the dermis layer samples in plots (E) and (F) have a steeply rising peak at 4890 cm -1 that is associated with the hydrogen bond vibrations in protein molecules 52 and is less noticeable in the epidermis and whole skin residuals. Most probably, the appearance of nonrandom spectral features in the residuals for the epidermis and dermis maps is due to discrepancy between the standard protein spectra of keratin and collagen type I and the absorption of these molecules in situ. Interactions with water and lipids affect the secondary structure of these proteins and may cause deviations from the Beer s 47, 51 law or shifts in the location or shape of these protein absorption spectra. Six Component Regression Fit Figures IV-17 and IV-18 demonstrate the traces of the regression coefficients of the six components plotted versus their respective locations on the map for the epidermis and dermis samples, respectively. These plots corroborate the conclusions made based on the absorbance spectra. Very low values are obtained for all six components for the single epidermis layer, E1, whereas coefficients for water and keratin protein are doubled in magnitude when the folded epidermis sample, E2, was measured. The constant offset term, mostly below zero, is dramatically increased by folding the tissue. The degree of spatial changes in coefficients is lower for the epidermis compared to the whole skin,

188 Regression Coefficient Regression Coefficient Regression Coefficient Regression Coefficient 160 (A) water collageni fat keratin constant slope (B) water collageni fat keratin constant slope Mapping Location Number Mapping Location Number (C) water collageni fat keratin constant slope Reference Spectrum Number (D) water collageni fat keratin constant slope Reference Spectrum Number Figure IV-17. Regression coefficients for epidermis samples: (A) and (B) are the results of regression fit for point maps of sample #E1 and sample #E2 (folded epidermis), respectively; (C) and (D) are the corresponding regression coefficients at the central location on the sample.

189 Regression Coefficient Regression Coefficient Regression Coefficient Regression Coefficient 161 (A) water collageni fat keratin constant slope (B) water collageni fat keratin constant slope Mapping Location Number Mapping Location Number (C) (D) water collageni fat keratin constant slope water collageni fat keratin constant slope Reference Spectrum Number Reference Spectrum Number Figure IV-18. Regression coefficients for dermis samples: (A) and (B) are the results of regression fit for point maps of sample #D1 and sample #D2, respectively; (C) and (D) are the corresponding regression coefficients at the central location on the sample.

190 162 which makes it safe to assume that the bulk of the spatial heterogeneity of human skin will be governed by the composition of the dermis layer. Indeed, regression coefficients for the dermis layers are very close to the values found for the whole skin samples. The largest coefficient is invariably for collagen protein and the second largest is for water. Some keratin protein is also quantified with the measured coefficients ten times larger than for the single epidermis sample, E1. The degree of heterogeneity for most components of the dermis samples, D1 and D2, is about ten times greater compared to the central reference location in Figure IV-18 (C) and (D). Other two important points about the dermis samples are that their fat content and the constant coefficients appear negligible. Fat coefficients are below zero at all times, which indicates the absence of the fat signal detectable with the linear regression model in the spectra for human dermis. Morphologically, fat is in fact concentrated in the subcutaneous layer of human skin 132 The coefficients for the constant offset term are also very close to zero. This is numerical evidence of how a single skin layer would produce lower scattering than a multilayered sample, such as a fold of epidermis or a slice of whole skin. Moreover, when the epidermis and the dermis of human skin tissue are measured separately and their constant terms summed up, the cumulative offset term is still lower than the constant coefficients for the whole tissue. Human Skin Layers Spatial Heterogeneity Spatial Contour Maps To demonstrate lateral heterogeneity across the layers of human skin, distribution maps were constructed using the absolute values for the six components in the regression model. These maps are shown in Figures IV-19 through IV-20 for the single epidermis, E1, double epidermis, E2, and single dermis, D1, layers, respectively. Spatial distributions found for the second dermis sample, D2, were very similar to sample D1 and are presented in Figure A-4 of this thesis. The single epidermis sample, E1, is characterized by a higher level of

191 163 (A) (B) (C) (D) (E) (F) Figure IV-19. Spatial distribution maps for the sample of a single layer of epidermis (map #E1), where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped sample. Black contours denote 10% change in the regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement. 38

192 164 (A) (B) (C) (D) (E) (F) Figure IV-20. Spatial distribution maps for the sample of a double layer of epidermis (map #E2), where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped sample. Black contours denote 10% change in the regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement. 38

193 165 heterogeneity in fat, keratin, and slope but these coefficient values are so small that any 10% change in a substance will be overshadowed by changes coming from the dermis layer at the same location. For instance, the total change across the map of the single epidermis is equal to 0.01 for keratin; for the dermis layer, D1, the value is three times larger, So even though the relative keratin content may be higher in the epidermis, the larger absolute values for keratin spatial changes in dermis will determine the distribution of this component in the whole skin. Likewise, the values for collagen protein are low, and often negative, in the epidermis maps. Certainly, the collagen distribution of a whole skin sample will reflect the distributions of the protein in the dermis layer. The distribution of fat will be strongly affected by any subcutaneous tissue present in the beam path after squeezing the tissue in the sample cell or positioning a fold of skin between the optical fiber of a noninvasive interface. 38 The sizes of the domains for all six components measured for the dermis layers are very close to those for whole skin. The white circles in Figure IV-21 (A) represent the contact areas of optical fibers on the tissue and were added for comparison. The structure of the dermis tissue may explain why the size of the heterogeneous domains is larger than in the epidermis: dermis layer is known to contain fewer cells with the bulk of the layer occupied by amorphous extracellular matrix (ECM). 184 Both the single and double epidermis maps in Figures IV-19 and IV-20 (A) reveal uniformly distributed coefficients for water. However, the dermis water distribution map in Figure IV-21 (A) contains oval domains similar in shape and size to those found for the whole skin samples. The water seen in the epidermis is for the most part the left over liquid from the separation procedure. Such residual moisture would be uniformly distributed across all locations on the map. The gradual decrease in water content from the bottom right hand corner to the top left hand corner has to do with the dehydration of the sample. The mapping was performed in that direction in a raster pattern. By the time the top left hand corner locations were probed, some water escaped from the epidermis.

194 166 (A) (B) (C) (D) (E) (F) Figure IV-21. Spatial distribution maps for the sample of a single layer of dermis (map #D1), where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped sample. Black contours denote 10% change in the regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement. 38

195 167 The effects of the dehydration process are so noticeable in the epidermis maps because the layer is very thin and has very little water to begin with. Spatial differences in water content dominate over dehydration in the thicker samples of human dermis and whole skin. Six Component Distributions for Human Skin Layers A side-by-side comparison of the epidermis and dermis layers can be found in Figure IV-22 for (A), (B) water, (C), (D) collagen type I protein, and (E), (F) fat regression coefficient absolute values. The other three components are compared in Figure IV-23 where (A), (B) are the boxplots of keratin, (C), (D) constant, and (E), (F) slope term coefficients. In both of the figures, values between 25 th and 75 th percentile are encompassed by green rectangles for epidermis and cyan rectangles for dermis samples. The data in each of the groups are coefficients from the 100 absorbance spectra of a corresponding map. In every pair of these plots for a given regression component, the dermis values are consistently larger, except for the constant term, the significance of which has been already described in this chapter. Even though the dermis layer is at least ten times thicker than the epidermis, 132 the thickness is not the only factor explaining the differences in these coefficients. Let us consider one tenth of the average collagen protein content (0.9) in dermis sample D1. The value of 0.09 will still be about twenty times larger than the maximum coefficient for collagen found for the single epidermis sample E1 (0.004). On the other hand, the values for fat are invariably negative for the dermis layer and very small, but above zero for both epidermis samples. Keratin and water content is, as it happens, thickness related. In the preceding section, the principal difference in the spatial distributions of water has been addressed. The contribution from keratin protein to the skin layers absorbance is best compared by calculating the relative (%) coefficients according to Equation III-1. Finally, the slope coefficients accounting for the nonlinear part of scattering background 170 and the effects

196 Regression Coefficient for Fat Regression Coefficient for Fat Regression Coefficient for Collagen Regression Coefficient for Collagen Regression Coefficient for Water Regression Coefficient for Water 168 (A) (B) E1 Sample Number E D1 Sample Number D2 x 10-3 (C) (D) E1 Sample Number E2 D1 Sample Number D2 x 10-3 (E) 0 (F) E1 Sample Number E D1 Sample Number D2 Figure IV-22. Box plots for (A), (B) water, (C), (D) collagen type I protein, and (E), (F) fat regression coefficients of epidermis and dermis layers, respectively. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by green (epidermis samples) and cyan (dermis samples) boxes.

197 Regression Coefficient for Slope Regression Coefficient for Slope Regression Coefficient for Constant Regression Coefficient for Constant Regression Coefficient for Keratin Regression Coefficient for Keratin 169 (A) (B) E1 Sample Number E D1 Sample Number D (C) 0.15 (D) E1 Sample Number E2 D1 Sample Number D2 10 x 10-3 (E) 0.04 (F) E1 Sample Number E2 D1 Sample Number D2 Figure IV-23. Box plots for (A), (B) keratin, (C), (D) constant, and (E), (F) slope regression coefficients of epidermis and dermis layers, respectively. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by green (epidermis samples) and cyan (dermis samples) boxes.

198 170 of temperature on the absorption of water 45, 46, 123 appear larger and more varied for the dermis samples. Dependence of optical properties of the dermis on the temperature has been demonstrated in the literature. 131 The coefficients for the constant term are mostly below zero for both layers and increase only for the folded epidermis, E2. Large negative coefficients for the constant can be explained by the difference in thickness of the regression standards and the epidermis layer. The contribution from scattering in the standard spectra (~ 1 mm thicknesses) had to be accounted for during the fitting of the thin epidermis layer data. Relative Regression Coefficients (%) for Chemical Components Because of the dramatic thickness difference between the dermis and epidermis layers, normalized regression coefficients provide useful information about the content of the tissue. The percent coefficient values calculated relative to the sum of all four chemical constituents are plotted in Figure IV-24 for (A), (B) water and (C), (D) collagen type I protein and in Figure IV-25 for (A), (B) keratin and (C), (D) fat components. The absolute and the relative coefficients are then summarized in Table IV-7 as the mean values. The reported standard deviations gauge the spread of the data. The relative contribution from water in the epidermis spectra is around 66%, which is twice higher than the 33% of water found in the dermis spectra. The validity of the observation may be questioned because the presence of the residual water from the separation procedure at the time of the microscopic mapping. Better controlled experimental conditions may help to elucidate the true nature of the larger relative water content in the epidermis. The trend for the relative coefficients for collagen protein matches the biochemical composition of epidermis and dermis: 144, 147 more than a half of the chemical-related light absorption in the dermis is explained by the presence of collagen molecules. Keratin (31%) and fat (3%) are higher in the epidermis with keratin being the second strongest absorber of the radiation after water.

199 Relative Regression Coefficient, % Relative Regression Coefficient, % Relatvie Regression Coefficient, % Relative Regression Coefficient, % 171 (A) (B) E1 Sample Number E2 24 D1 Sample Number D2 (C) (D) E1 Sample Number E2 54 D1 Sample Number D2 Figure IV-24. Box plots for (A), (B) water, and (C), (D) collagen type I protein relative regression coefficients (%) of epidermis and dermis layers, respectively. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by green (epidermis samples) and cyan (dermis samples) boxes.

200 Relative Regression Coefficient, % Relative Regression Coefficient, % Relative Regression Coefficient, % Relative Regression Coefficient, % 172 (A) (B) E1 Sample Number E2 D1 Sample Number D2 (C) (D) E1 Sample Number E2 D1 Sample Number D2 Figure IV-25. Box plots for (A), (B) fat, and (C), (D) keratin protein relative regression coefficients (%) of epidermis and dermis layers, respectively. Black circles represent mean values, red horizontal lines median values, red plus signs outliers. The values between 25% and 75% of the distribution are represented by green (epidermis samples) and cyan (dermis samples) boxes.

201 173 Table IV-7. Relative regression coefficients (%) summary for human skin layers. Collagen Water Fat Keratin Constant Slope type I Regression coefficients, mean ± standard deviation E1 E2 D1 D ± ± ± ± ± ± ± ± Epidermis Samples ± ± ± ± ± ± Dermis Samples ± ± ± ± ± ± ± ± ± ± Relative regression coefficients (%), mean ± standard deviation Epidermis Samples E1 E ± ± ± ± ± ± ± ± Dermis Samples D1 D ± ± ± ± ± ± ± ±

202 174 Correlations between Components of Skin Layers Pearson correlation coefficients were calculated for the 15 possible pairs of tissue components for the two maps of epidermis, E1 and E2, and the two dermis maps, D1 and D2. The values with their standard errors are reported in Table IV-8. The positive values for constant could not be reliably obtained for the single epidermis and dermis layers with the absorbance standards used in the six component regression. For this reason, all of the correlation coefficients, r, involving the constant term are considered insignificant and reported in a normal font. Two additional similar cases are the negative fat content in the dermis and the negative collagen type I content in the epidermis samples. As the water content in the epidermis layer was mostly residual water, the correlations involving water in the epidermis are also ignored. The correlation coefficients judged to be meaningful are reported in bold font. Among these, the underlined values held true for both samples of a respective layer. For example, for dermis a strong correlation between collagen and slope term exist in sample D2 but not in sample D1. Consequently, these r values are reported in bold font but are not underlined. The strongest negative correlations were observed in the dermis tissues between water and both proteins. The hydrated collagen and keratin aggregates displace water molecules out from the beam path causing a negative correlation between these components. The positive correlation exists in the dermis spectra between the two proteins themselves. The only strong correlation deemed significant was found in the epidermis maps that is the correlation between fat and keratin. The value of 0.43 for sample E1 is a little below 0.5 but it is the largest meaningful correlation coefficient for that sample. Interrelated localization has been previously measured with IR microspectroscopy between lipid and proteins for samples of porcine epidermis. 162 The authors demonstrated how areas with higher protein content exhibit lower lipid signal and the other way around. Of course, collecting microspectroscopic maps from more samples of human skin layers is needed to give validity to these preliminary observations.

203 175 Table IV-8. Correlation coefficients, r ± standard error, for human skin layers samples. Epidermis Dermis Components E1 E2 D1 D2 Water vs. Collagen Type I 0.30 ± ± ± ± 0.08 Water vs. Fat 0.55 ± ± ± ± 0.1 Water vs. Keratin 0.69 ± ± ± ± 0.09 Water vs. Constant 0.42 ± ± ± ± 0.09 Water vs. Slope ± ± ± ± 0.10 Collagen Type I vs. Fat 0.22 ± ± ± ± 0.09 Collagen Type I vs. Keratin 0.35 ± ± ± ± 0.07 Collagen Type I vs. Constant ± ± ± ± 0.05 Collagen Type I vs. Slope 0.28 ± ± ± ± 0.09 Fat vs. Keratin 0.43 ± ± ± ± 0.07 Fat vs. Constant ± ± ± ± 0.07 Fat vs. Slope ± ± ± ± 0.08 Keratin vs. Constant 0.46 ± ± ± ± 0.05 Keratin vs. Slope ± ± ± ± 0.08 Constant vs. Slope 0.18 ± ± ± ± 0.07

204 176 Conclusions Samples of human whole skin from two male and two female donors were microspectroscopically analyzed over in the combination region of the NIR spectrum ( cm -1 ). The levels of RMS noise for these tissues were found acceptable for determination of the major attenuators of NIR light: water, collagen type I and keratin proteins, fat, constant, and slope. The tissues were excised from different locations on the body and had different thicknesses but all of them provided evidence of spatial heterogeneity in the horizontal XY plane across the sample. The regression coefficients for each component when plotted in two dimensions according to their location of origin clearly demonstrated oval domains of high and low content from around 500 to 1000 μm in diameter. The size of the observed domains was somewhat larger than that presented in Chapter III for rat skin samples and comparable to the size of the optical fiber interface used in the initial noninvasive measurements of glucose in skin by our group. 38 In an attempt to better understand the composition of skin tissue in relation to NIR spectroscopic measurements, samples of epidermis, a fold of epidermis, and a dermis layer were obtained. Microspectroscopic data for these skin layers are consistent with previous research in that the two major components absorbing NIR radiation are fat and keratin protein in the epidermis and water and collagen protein in the dermis layer of human skin. The prevalence of the signal from the dermis tissue in the spatial maps of whole skin is also inferred from these measurements. Comparison of the coefficients for the constant offset term in the single layers of human skin and the folded epidermis sample demonstrate how scattering is mostly caused by the multilayered structure of whole skin sample as originating from interfaces between layers or media with different refractive indices. Scattering background within individual layers was found to be very low for both dermis and epidermis samples. The summarized regression coefficients for human skin will be used in Chapter V to investigate the effect of the spatial heterogeneity on noninvasive glucose predictions.

205 177 CHAPTER V LOCATION-SPECIFIC PLS CALIBRATION MODELS FOR GLUCOSE IN RAT AND HUMAN SKIN TISSUE Introduction Repositioning Errors in Noninvasive Glucose Measurements in Skin The utility of measuring glucose by noninvasive near-infrared transmission spectroscopy in an animal model was previously demonstrated by our group. 37, 112, 116, 130 In the noninvasive measurements, an optical fiber launches the incident light into one side of a fold of skin tissue and a second fiber directs the transmitted light to a detector for quantification. Glucose concentrations were obtained in those studies by multivariate methods. 37, 112, 116 Minimal variations were observed in the predicted glucose concentrations when the position of this fiber optic interface was constant throughout the measurement. Large variations in the glucose concentration predictions were observed, however, when the fiber optic interface was repositioned between each measurement as shown in Figure I-3. Standard errors for glucose concentration predictions increased 2.55-fold when the interface was repositioned between measurements. 38 Clearly, such a sensitivity of the position of the interface is a critical issue for noninvasive measurements and is a known obstacle to achieving universal calibrations for many noninvasive techniques. 10, 130, 185 Vibrational microspectroscopic mapping of rat and human whole skin was applied in this dissertation to gain a better understanding of the origin of this interface sensitivity. Distribution maps for major chemical components of the skin matrix were constructed based on skin tissue microspectroscopic data collected over the combination region of the near infrared spectrum ( cm -1 ) as described in Chapters III and IV. These maps clearly reveal chemical domains within the tissue matrix where regions on the order of µm for rat skin and μm for human skin are composed of different

206 178 amounts of water, fat, collagen protein, and keratin protein as well as possessing markedly different scattering properties. The size of these domains is significant in comparison to the size of the optical fibers used to collect the noninvasive spectra. In our initial noninvasive measurements, fibers with outer diameters of 1.8 mm were used to both launch and collect the probe radiation. The chemical information obtained from this optical geometry corresponds to the integrated composition of the skin matrix over the volume of skin probed by the propagating photons. This volume can be approximated as a cylinder with a diameter of 1.8 mm and a length of 1 mm. The microspectroscopic measurement, on the other hand, provides lateral resolution of 0.48 mm in the X and 0.36 mm in the Y direction; and involves transmitting the incident light through a single layer of skin with a thickness ranging from 0.5 to 1.5 mm. The resolution of the resulting distribution maps is governed by the step size of the stage used to position the tissue slice under the microscope objective. The superior resolution of the microspectroscopic image compared to the in vivo measurements permits identification of concentration domains within the tissue matrix. On the basis of the size of these domains relative to the optical geometry, we have speculated that chemical heterogeneity within the skin matrix is an important factor in the position sensitivity of the interface used to collect noninvasive spectra for the noninvasive measurement of glucose concentrations. 186 Principal Component Analysis (PCA) of Tissue Spectra NIR spectral data are characterized by broad and overlapping absorptions and low signal-to-noise ratios. 40, 187 These disadvantages are concerns for NIR spectroscopy of skin tissue (see Chapters II-IV for representative spectra and RMS noise values) and cause complications in identifying spectral changes solely due to glucose concentration. Principal component analysis (PCA) is a multivariate technique that offers the ability to decrease the dimensionality of the calibration data A large set of skin tissue spectra can be conveniently reduced by PCA to a smaller number of vector components, independent of each other in multidimensional space. Each of these

207 179 calculated principal components (PCs, eigernvectors, latent variables, or factors) has a scalar score associated with it. The relative magnitude of the eigenvalue defines the fraction of the total variance in the spectral data set explained by the corresponding PC. Usually, the first several factors that account for most of the tissue variance are chosen for subsequent analysis. In the present work, PCA is only used in the context of validating that the PLS calibration vector is based on net analyte signal (NAS) for glucose. For this reason, a discussion is focused on the procedure used to obtain principal components associated with the spectral background of the tissue matrix. Principal component regression (PCR) is beyond the scope of this dissertation. For the work in this chapter, principal components were obtained with the singular value decomposition (SVD) algorithm. 188 This is a common algorithm in linear algebra that extracts orthogonal vectors from a matrix of the original partially correlated data. As the first step of PCA, the simulated skin tissue spectra were arranged in a n p matrix X, where n was the number of the spectra, usually 50 per location, and p was the number of resolution elements or wavenumbers. SVD reprocessed the data matrix as shown in Equation V-1. X = U Σ V Equation V-1 where U represents a n n orthogonal matrix with the eigenvectors of XX T contained in its columns, Σ is a n p diagonal matrix of singular values of X, and each column of the orthogonal p p matrix V is the eigenvectors of X T X. The relevant number of latent variables was then selected as the number of columns of V. Because the spectral data used for multivariate analysis in this chapter is mathematically simulated with the predetermined six sources of non-analyte variance, k = 6 factors were chosen as the independent PCs of the background tissue spectra. Equation V-2 describes how the score values for these six principal components were obtained. T k = U k Σ k Equation V-2

208 180 where T k is the score n k matrix for the six selected eigenvectors, the n k matrix U k contains the first k = 6 columns of the U matrix from Equation V-1, and the diagonal k k matrix Σ k holds the first k values from the Σ matrix. The actual spectral shapes for the first six principal components were obtained by plotting the first k = 6 columns of the matrix V for the analyzed wavenumber region, cm -1. In addition, when the eigenvalues for the X T X matrix were plotted versus the number of eigenvalues to determine the optimal k, the trend sharply dropped after the first six factors. Partial Least-Squares (PLS) Regression The partial least-squares (PLS) algorithm was employed among other multivariate techniques to quantify glucose in the noninvasive transdermal measurements by our 37, 38, 112 group. Consistent with the in vivo experiments, PLS is applied to assess effects of skin heterogeneity on glucose predictions from simulated spectra. 186 PLS regression is a well-established computational tool used to extract information of interest from a complex background matrix NIR-based PLS regression models for water, glucose, 42, 110, 125, 177, 194, 195 lactate, urea, and other analytes have been reported in the literature. The three main advantages of PLS are (1) decrease in the colinearity of the data, (2) reduction of the number of independent variables to those correlated with the analyte variances, and (3) decrease or complete removal of random noise in the data set. 193 These features make PLS regression a powerful multivariate technique for analytical measurements over the NIR region of electromagnetic spectrum. 32 The algorithm is a variation of PCA and involves data reprocessing by projecting the spectra onto a new set of coordinates in a multidimensional space. This set of coordinates is obtained by maximizing correlation with the analyte information. In PLS, spectra and analyte concentration values are modeled together to obtain the latent variables maximizing covariance between the two. For the operations of linear algebra in this section, capitalized letters indicate matrices and the lower case letters are their column vectors. As the first step of the PLS

209 181 algorithm, the first loading weight w 1 of the covariance matrix between the simulated NIR spectra and the assigned glucose concentrations is obtained following Equation V-3. T X y w Equation V-3 1 T X y In this equation, X T is the transpose of the original n p mean-centered spectral data matrix, y is the n 1 vector of assigned glucose concentrations that were also meancentered. With the computed analyte-biased weight loading w 1, the first score vector, t 1, the first spectral loading, p 1, and the first concentration loading, q 1, can be determined as given by Equations V-4 through V-6. t 1 = Xw 1 Equation V-4 p 1 X T 1 T 1 t1 t t Equation V-5 q 1 t c Equation V-6 t T 1 T 1 t1 The contribution of the first loading is removed and the residuals for absorbance spectra and concentration values are calculated according to Equations V-7 and V-8. X 1 = X t 1 p 1 T Equation V-7 y 1 = y t 1 q 1 Equation V-8 Spectral residuals matrix, X 1, and concentration residuals vector, y 1, are now used in Equation V-3 in place of X and y, respectively. The second loading weight, w 2, is calculated. The algorithm expressed in Equations V-3 through V-8 is reworked until k latent variables are accumulated. The value of k is not arbitrary and should be carefully determined in the PLS regression optimization step. 19

210 182 The calibration b-vector is determined in Equation V-9 using the matrix of k loading weights, W, the p k matrix of spectral loadings, P, and the vector of k concentration loadings, q. b = W(P T W) -1 q Equation V-9 Concentrations for the unknown samples can be predicted as the product of the spectral matrix, X, and the calibration vector, b, as shown in Equation V-10 below. ĉ = Xb Equation V-10 Evaluation of Partial Least-Squares (PLS) Regression PLS regression is a powerful statistical tool that should not be abused by applying blindly without preceding optimization steps. 196 Two of the most important steps are determining the number of latent variables and optimizing the spectral range An insufficient number of calibration factors, k, eliminates important analyterelated information, thereby decreasing sensitivity and robustness of the model. On the other hand, too many latent variables can lead to overmodeling and introduce non-analyte correlations into the model. The optimal number of latent variables is determined in this chapter by leave-five-samples-out cross validation. The number of the spectra left out should reasonably represent typical variations in the simulated tissue data without compromising the degrees of freedom left for the calibration. 200 Allocating at least 5 independent samples in the calibration set for each latent variable is recommended. 201 Five spectra randomly selected at each tissue location represent 10% of the whole data set. The PLS calibration model was created with the remaining 45 spectra (maximum of 9 factors) at a given location. The procedure was repeated until all of the spectra in the data set had been sampled. The cumulative error called the cross validation standard error of prediction (CV-SEP) is then determined according to Equation V-11. The values of CV-SEP are plotted for each number of latent variables to determine the best value of k.

211 183 In the following equation, m stands for the number of times groups of 5 spectra were removed for prediction, c p_as and ĉ p are the assigned and the predicted concentrations of glucose, and n p is the number of samples left out for prediction, 5 in this application. CV-SEP m 2 1 ( cp as cˆ _ p) Equation V-11 m n p Optimization of the spectral range for the data in matrix X ensures that only the most analytically important spectral features end up in the calibration set. 202 Regions with low S/N ratios should be removed because they have detrimental effect on analyte predictions. In this work, the whole combination range between cm -1 was used for PLS regression matching the conditions of the in vivo experiments. 38 Including all of the wavenumbers of such a wide range into the calibration allowed us to model changes in broad spectral features of skin, for example, the location-dependent scattering background. Regions with poor S/N ratios because of the increased water absorbance, cm -1 and cm -1, were excluded from the data simulation. It is important to realize, that the purely statistical nature of the PLS algorithm prompts it to correlate all spectral variances in X with the concentration variance in y. The sources of the spectral variance are not taken into account by this procedure. Nonanalyte correlations with variance due to noise and/or instrumental drift have been demonstrated for PLS models computed for glucose. These seemingly functional models based on the spurious correlations reveal distinctly different shapes of the calibration b- vector than reliable models truly based on glucose information. 39, 39, However, the shapes of the b-vector cannot be directly compared with the absorptivity spectral features of glucose. Instead, comparison with the calibration net analyte signal (NAS) b-vector for glucose in skin matrix was performed. The relationship between the tissue background, the pure component glucose spectrum, and the net analyte signal (NAS) for glucose is presented in Figure V-1.

212 Figure V-1. Illustration of glucose NAS vector calculation in three-dimensional space. 184

213 185 The NAS concept was introduced by Lorber in , 204 In multidimensional vector space, the net analyte signal is the portion of the pure analyte spectrum that is orthogonal to the background spectra devoid of the analyte information. To obtain the NAS for glucose, Equation V-12 was used NAS glu = (I A -glu A + -glu)a glu Equation V-12 where a glu is the pure component spectrum of glucose, A is the tissue background matrix containing all sources of variances except glucose, A + -glu is the pseudo inverse of A -glu, and I is the identity matrix with the dimension of A -glu A + -glu. 177 Decrease in dimensionality of the background data improves NAS performance. As described earlier, PCA was used to pick out six independent eigenvectors or PCs representing more than 99.99% of the background tissue variance. The pure component spectrum of glucose, a glu, was first projected onto the principal component vector with the highest score and the remainder was projected onto the next PC. The procedure for the k th projection is described by Equation V-12: a * glu a glu a v glu T k v v k k v k Equation V-12 where a * glu is the orthogonal part of the glucose pure component spectrum remaining after projection on the k th principal component of the tissue matrix, v k, and a glu is the portion of the pure component spectrum used for projection. For the first projection, a glu is the complete pure component spectrum of glucose. Quantitation with NAS is analogous to Equation V-10 if the NAS vector is normalized first by Equation V-13: * aglu NAS glu Equation V-13 a * glu where NAS glu is now the calibration NAS vector and * aglu is the norm of the orthogonal portion of glucose spectrum left after projection onto k tissue PCs. The normalized NAS

214 186 is analogous to the PLS calibration b-vector even though the mechanism of NAS uses specific analyte information rather than purely statistical correlations. If the two vectors are similar in shape and magnitude, the selectivity of the PLS calibration vector is demonstrated. 116 In other words, the conclusion can be made that the PLS calibration is indeed modeling spectral variance in the data due to glucose and not spurious correlations. Optimized PLS calibration models were created for various locations on the tissue maps. Errors in the concentration of glucose predicted from a PLS calibration model were quantified as the standard error of calibration (SEC) for the calibration quadrant and as the standard error of prediction (SEP) for the non-calibration quadrants or prediction quadrants. These parameters and their percent ratios (SEP/SEC) can be computed according to Equations V-14 through V-16: ( c ˆ as cc ) SEC n k 1 c 2 Equation V-14 ( cas SEP n p cˆ ) p 2 Equation V-15 SEP SEP / SEC 100% Equation V-16 SEC where SEC and SEP are in units of mm, c as, c c, and c p are the assigned, calibration and prediction concentrations of glucose for each spectrum, respectively, n c and n p are the number of spectra in the calibration and prediction data sets, respectively, and k is the number of factors used in the PLS regression analysis. Generally, it is hard to obtain the same absolute values of the standard errors for the simulated data and the in vivo procedures because of all the extra potential sources of variance in the in vivo data set.

215 187 The percent ratio between the standard errors served as an indicator of the degree of degradation in sensitivity of the PLS models between locations of the tissues. Chapter Overview Herein, the impact of tissue heterogeneity is assessed by generating PLS calibration models from simulated spectra corresponding to one region of a given slice of tissue and then using this model to predict the concentration of glucose represented in spectra simulated for a different region within the same tissue slice. Sets of noninvasive near-infrared spectra are constructed mathematically as linear combinations of pure component absorption spectra for each of the major chemical components of the skin matrix (water, fat, collagen protein, and keratin protein). The ratio is computed between standard errors of prediction (SEPs) for glucose concentration predictions in the predicted locations relative to standard errors of calibration (SECs) in the calibration location. If this percent ratio is less than or equal to 100%, differences in skin matrix composition between the locations do not affect glucose predictions. If, however, the ratio is greater than 100%, then tissue heterogeneity has an impact on glucose predictions. Finally, the ratios determined from these simulated spectra are compared to the ratio of prediction errors measured during in vivo noninvasive glucose measurements. Experimental Procedures The procedure used to test the effects of skin heterogeneity is outlined in Figure V-1 that includes the basic steps of data collection and analysis. The top white block addresses processing of the in vitro spectral data from Chapters II and III, the green block explains two approaches to simulation of spectra at various locations across the skin, and the blue block at the bottom of the chart describes PLS regression procedures. Our initial approach to testing the effects of skin heterogeneity included splitting each map into four nonoverlapping quadrants and using each quadrant to predict the other three on the same map. The second, more rigorous approach incorporates 36 possible locations on the skin.

216 188 Collect microspectroscopic maps in vitro from rat and human whole skin samples. Perform least-squares fit of the skin spectra to the six-component model: Water, collagen type I protein, fat, keratin protein, constant offset, and slope. Use computed distributions of the six components to simulate average spectra for 4 quadrants of each map. Add realistic sources of variance: Use computed distributions of the six components to simulate average spectra for 36 locations of each map. + small temporal variance in each component; + glucose absorbance for random concentrations from 5 to 35 mm; + random noise at the level of 12 μau. Perform PLS calibration and prediction with the simulated spectra at various locations. Compare standard errors for a constant location (SEC) and when the location is changed (SEP) as SEP/SEC % ratios. Figure V-2. Summary of our approach to quantify the influence of skin tissue heterogeneity on noninvasive glucose prediction. 186 White block addresses processing of the in vitro spectral data from Chapters II and III, green block explains simulation of spectra at various locations across the skin, and blue block describes PLS procedures.

217 189 Simulation of Skin Tissue Spectra Average noise-free base spectra were calculated for a rectangular area on the skin corresponding to a 5 5 section of that tissue s absorbance map. Each of these base spectra corresponds to an absorbance spectrum that would be collected in vivo if the 1.8 mm diameter optical fiber probed the area in question. The base spectra were obtained by taking the averages of the 25 regression coefficients for each of the six skin components (water, collagen type I protein, fat, keratin protein, constant, and slope) and multiplying these values by the corresponding standard absorbance spectra shown in Figure II-3. The resulting weighted absorbance spectra for the six components were summed to yield the base spectrum for a given location. The regression coefficients for 25 measured points in each rectangular location were taken from the chemical distribution maps determined for actual rat and human skin samples described in Chapters II and III. The base spectrum for each quadrant was further modified to produce the set of simulated spectra needed to produce PLS calibration models for glucose. These additional modifications included: 1) adding small spectral variance associated with the temporal changes in the six major matrix components at the reference location, 2) adding glucose absorbance at the level corresponding to a specific concentration of glucose, and 3) adding realistic Gaussian distributed noise. Addition of Spectral Variance The observed temporal variations in the coefficients of these six components at the central reference location of each tissue map were added to the average base spectra calculated for all 5 5 locations on the map. As described in Chapter II, a set of reference spectra was collected in triplicate at the center point of each map array every 30 minutes. The mean and standard deviation values were computed for the six regression coefficients to provide a measure of skin changes due to instrumentation variations, dehydration, and temperature fluctuations that were deemed to be location-independent.

218 190 The relative variation for each component was then calculated and used to simulate variance in spectra for each quadrant according to the following expression: comp comp = comp i i comp center Equation V-X where (σ comp ) i is the standard deviation for a given skin tissue component in quadrant i, (σ comp / comp ) center is the relative standard deviation for this component at the central point of the map, and ( comp ) i is the regression coefficient for this component in quadrant i, calculated as the average of 25 coefficients in the quadrant. The mean ( comp ) i and standard deviation (σ comp ) i for each component in quadrant i are used to define the normal Gaussian distribution for the tissue. Next, 50 coefficients were randomly picked from these distributions for each component. The resulting regression coefficients were multiplied by the respective pure component absorption spectra and summed up to provide the simulated noninvasive tissue spectra. At this point, the spectra were noiseless and did not contain glucose information. Addition of Glucose Component Glucose spectra were added to instill glucose specific information within the sets of 50 simulated spectra. Glucose concentration spectral information was added according to the following equation: 43 A glu solution = ε w C ε C ε f C Equation V-Y w w glu w glu w glu w w glu where A glu solution is the component of the tissue absorbance spectrum representing glucose dissolved in aqueous media in the skin, ε ε w C is the absorbance due to pure water, glu glu w Cglu is the absorbance due to glucose, and εw w w Cglucose w w f is a term that accounts for changes in the measured absorbance caused by the displacement of water molecules in the optical path by the molecules of glucose. Conventional terms are used where ε w and ε glu represent molar absorptivities (AU mm -1 mm -1 ) for water and

219 191 glucose, respectively, w is the effective aqueous path length (mm) for the light propagating through the tissue slice obtained as the regression coefficient for the water component, C w respectively, and and C glu correspond to the concentrations (mm) of water and glucose, glu fw is the water displacement coefficient for glucose. Glucose concentrations had uniformly distributed random generated values between 5 and 35 mm. The water displacement coefficient was previously determined in our group to have a value of 6.24 ± 0.01 for glucose. 43 The part of a simulated absorbance spectrum associated with water absorption was taken out and replaced with the glucose solution absorbance, A glu solution. By this method, each of the 50 absorbance spectra simulated for a location of the optical fiber becomes a linear combination of standard spectra of water, collagen type I protein, fat, keratin protein, offset, slope, and glucose terms. Addition of Spectral Noise The last step in data synthesis was to introduce realistic spectral noise. The standard absorbance spectra for the six components of linear regression and the glucose component were essentially noiseless. After these values had been combined to yield the sets of 50 simulated spectra, random noise was needed to simulate the spectral quality achievable with the instrumental setup used to collect data from living rats. Root mean square (RMS) noise values for noninvasive, in vivo spectra were about 10 to 15 μau for a 1-mm sample of water and a 1.8 AU neutral density filter. 38 Based on this range, wavelength independent random noise was added onto all of the simulated spectra to result in a RMS noise level of 12 μau computed between 4400 and 4600 cm -1. PLS Models for Glucose Prediction Two approaches were adopted to investigate the effects of spatial heterogeneity on performance of PLS models as illustrated in Figures V-3 and the flowcharts in Figure V-4. In the first approach, each tissue matrix was divided into four non-overlapping quadrants. The size of each quadrant was mm 2, which was approximately the size of the fiber-optic interface used to collect noninvasive spectra. The calibration model

220 192 (A) (B) Figure V-3. Simulation of optical fiber movement between (A) 4 quadrants and (B) 36 possible locations on the skin slice.

221 193 (A) 1 Select one out of 4 quadrants 2 Simulate 50 NIR spectra 3 PLS calibration for glucose Accumulate 4 SECs and 12 SEPs per map 4 Generate one SEC Select a different quadrant and repeat PLS prediction for all other quadrants 6 Generate 3 SEPs (B) 1 Select one out of 36 quadrants 2 Simulate 50 NIR spectra 3 PLS calibration for glucose Accumulate 36 SECs and 1260 SEPs per map 4 Generate one SEC Select a different location and repeat PLS prediction for all other locationss 6 Generate 35 SEPs Figure V-4. Flowcharts illustrating the PLS regression procedures for (A) 4 quadrants and (B) 36 possible locations on the skin slice. Each cycle is reproduced 10 times.

222 194 was then generated from the set of simulated spectra for one quadrant, a standard error of calibration (SEC) was produced for that quadrant, and this model was used to predict glucose concentrations in the simulated data sets for the other three quadrants generating three standard errors of prediction (SEP). PLS calibration models were generated for every quadrant of the map as shown in Figure V-3 (A). Average standard errors per map were computed. The second approach involved exploring all possible locations of the 5 5 rectangle on a given tissue map. With the step size of a single pixel on the map, the number of possible positions of the mm 2 rectangle on the tissue is equal to 36 overlapping locations. The second approach presented in Figure V-3 (B) mimics slighter movements of optical fiber across the mm 2 tissue sample. PLS calibration was created and optimized for one chosen location out of the 36 locations on the skin map and SEC was generated for that location. This calibration was used then to predict all of the other 35 locations resulting in 35 SEPs. In both approaches, PLS models for individual locations were computed using all 50 of the spectra simulated for that location. This number of independent spectra was enough to justify the use of seven latent variables or factors in constructing the models. The optimal number of latent variables was determined by cross validation leave-10%- out procedure. A full spectral range between 4200 and 4900 cm -1 was used for building the multivariate model for the product of glucose concentration and the regressed aqueous path length, w. The predicted concentration-length values were divided by the corresponding water coefficients prior to evaluation. Evaluation of PLS Models for Glucose Concentration Predictions Selectivity of the PLS models for individual locations on the skin tissue was demonstrated by comparing the shape of the PLS calibration b-vectors with the normalized NAS calibration vector for glucose computed for the same location. The NAS calibration vectors were obtained by projecting the pure glucose analyte spectrum onto

223 195 the multidimestional space defined by principal components (PCs) representing variance in the simulated tissue matrix devoid of glucose. Errors in the concentration of glucose predicted from the resulting PLS calibration models were quantified as the standard error of calibration (SEC) for the calibration rectangular location and as the standard error of prediction (SEP) for the non-calibration locations or prediction quadrants. These parameters were computed according to Equations V-14 and V-15. The percent ratios between the average SEP from all possible non-calibration locations and SEC from the calibration quadrant were obtained and reported as SEP/SEC ratios, %. These values were then compared with the data from the in vivo rat experiment to demonstrate the effect of fiber relocation on variations in glucose predictions. Results and Discussion Effect of tissue heterogeneity on glucose prediction was studied for the skin maps from 8 male rats, 4 female rats, human female Specimens 1 and 3, and human male Specimen 5. Data for the dermis and the epidermis layers were excluded because they do not accurately represent whole skin tissue heterogeneity. Microspectroscopic data from the male Specimen 2 were regarded as an exception because of the origin of the tissue and were not used for PLS modeling. To facilitate the discussion of the PLS results, map #81 collected for female Rat 4 is used in this chapter for detailed evaluation as a representative tissue map. Spectral Simulation Study The fiber optic interface previously used to collect in vivo spectra for noninvasive glucose measurements in rat skin has a working diameter of 1.8 mm. Consequently, this interface collects radiation over a larger area than that resolved by the microspectroscopic mapping. Moreover, the SNR for the microspectroscopic measurements in Chapters II and III are two to three orders of magnitude lower compared to the SNR used for our reported noninvasive glucose measurements in living rats and

224 196 human subjects. 37, 38, 112 The low optical throughput of the microscope results in low radiant powers at the detector and high RMS noise levels for the microspectroscopic spectral data. As a result, the actual microspectroscopic data for rat and human whole skin samples cannot be used to measure glucose concentrations in these samples and cannot be used directly to assess the impact of tissue heterogeneity on the accuracy of multivariate calibration models for glucose. For this reason, simulation of skin tissue spectra was necessary to evaluate the impact of tissue heterogeneity on noninvasive glucose measurements. Skin absorbance at each possible location of the optical fiber interface across a map was approximated with 50 unique spectra. The differences between these spectra were not only in random noise but also in the combination of the six major skin components. Incorporation of the variance in these six components was necessary to make the spectra independent samples as required for PLS analysis. The principal difference between the two approaches to modeling interface relocation is the presence or absence of overlap between individual fiber locations on the skin map. To help visualize the physical meaning of our initial approach, Figure V-5 illustrates the spatial distributions of the six skin constituents for map #81 and the sizes of the rectangles corresponding to the contact area of optical fibers. More specifically, 4 non-overlapping locations or quadrants Q1 Q4 are indicated in the distribution map in Figure V-5 (F). Variations between mapped points within these quadrants are averaged by calculation of the base spectra. Essentially, the only variations in skin composition picked out by the first approach are those between the 4 base absorbance spectra for adjacent quadrants Q1 Q4. The average coefficients for water, collagen type I protein, fat, keratin, constant, and slope are listed in Table V-1. In the second approach, the size of the rectangle on the map used to obtain the average base spectra is kept the same, 5 5 points on the map. However, the locations are allowed to overlap this time until all 36 possible contact locations of the optical fiber interface are explored. The advantage of this approach is in its similarity to actual interface movements during an in vivo

225 197 (A) (B) (C) (D) (E) (F) Q3 Q1 Q4 Q2 Figure V-5. Distribution maps for sample #81 with 4 quadrants delineated for each skin component: (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant offset, and (F) sloping baseline. Location of the four quadrants is denoted in the distribution map for the slope coefficients.

226 198 Table V-1. Average regression coefficients for 4 quadrants of map #81. Quadrant Water Collagen Type I Average Coefficient Fat Keratin Offset Slope Q Q Q Q

227 199 experiment when slight shifts of the interface are possible in addition to drastic relocations. The tissue absorbance spectra for map #81 are featured in Figure V-6. Raw absorbance spectra for the map are plotted in Figure V-6 (A). The spread of the data reflects differences between all of the 100 measured points on this tissue map. If relocation of the optical fiber interface is modeled according to the first approach, the spread in the tissue absorbance spectra decreased as can be seen in Figure V-6 (B) for the 4 quadrants. This set is used as the average base spectra for PLS modeling. Figure V-6 (C) presents the average base spectra for map #81 obtained according to the second approach for 36 possible locations of the optical fiber interface. Once again, the spread of the data is less than that seen in the raw spectra for this map. Clearly, resolution of the microspectroscopic measurements is sufficient to capture the skin heterogeneity relative to the dimenstions of the optical fiber interface. In a typical noninvasive experiment, the optical interface is held stationary for collection of most of the spectra. Occasional shifts of the interface occur when an anesthetized animal or a human subject suddenly moves. Our assumption is that all of the 50 spectra for a given location are collected one after another as a group and then the fiber is shifted to a different location. Obviously, at the time when these 50 spectra were collected dehydration, changes in temperature and other temporal variations could take place. These variances are introduced in the data set by changing the amounts of the six tissue components at each location according to the distributions obtained from the central reference point on the map during the course of the microspectroscopic study. Embedding these small temporal variances as well as the glucose information and random noise into the simulated spectra allowed us to create viable PLS models for individual skin locations without incorporating the spatial heterogeneity between locations.

228 Absorbance, AU Absorbance, AU Absorbance, AU 200 (A) Wavenumber, cm (B) quadrant 1 quadrant 2 quadrant 3 quadrant Wavenumber, cm (C) Wavenumber, cm Figure V-6. Simulated absorbance spectra for map #81 with the microscopically measured absorbance spectra shown in (A), average spectra for 4 quadrants presented in (B), and the average spectra for 36 random locations on the map featured in (C).

229 201 Optimization of PLS Calibration Model Application of the PLS algorithm to extract analyte information from a complex biological matrix is often not a straightforward procedure. 19, 128, 196 Because of the statistical nature of the PLS regression, under- or overfitting of the data can easily take place if the calibration model has not been properly optimized. It is important to account for all of the variance in the data set correlated with the analyte information but exclude possible chance correlations and correlations with random spectral noise. The number of latent variables used in the PLS models in this dissertations was determined by cross validation procedure as demonstrated in Figure V-7 for 4 adjacent quadrants of map #81. In this figure, cross validation standard errors of prediction (CV-SEPs) are plotted as obtained by a leave-five-out procedure. The number of factors that generated the lowest CV-SEP was 7 for all of the quadrants on the map. The decrease of CV-SEP from 7 to 8 factors is not statistically significant. This number makes a lot of sense for the simulated data sets because there are 7 sources of nonrandom variance present in the simulated spectra: variance in the six skin components of the linear regression and variance in the glucose component used for augmenting the spectra. Cross validation optimizes a PLS model from the sensitivity point of view. The predictive ability of the model still can potentially originate from chance correlations within the data set. 102 To prove that the PLS model utilizes actual glucose analyte information, investigation of calibration b-vectors shapes was performed. Representative PLS calibration b-vectors for 4 quadrants of map #81 are presented in Figure V-8 (A) and (B) for models with 6 and 7 latent variables, respectively. Notably, the same set of 50 simulated spectra was used to obtain these vectors with the only difference being the number of calibration factors. The model with 6 latent variables insufficiently accounts for the variance between locations as suggested by the markedly different shapes of the calibration vectors for all four quadrants in Figure V-8 (A). Quadrants Q2 and Q3 produce calibration vectors that are comparable in shape and magnitude. The other two

230 CV-SEP, mm quadrant 1 quadrant 2 quadrant 3 quadrant Latent Variables Figure V-7. Cross validation standard error of prediction (CV-SEP) values for 4 quadrants of sample #81.

231 PLS Calibration b-vector, mm*mm/au PLS Calibration b-vector, mm*mm/au quadrant 1 quadrant 2 quadrant 3 quadrant 4 (A) Wavenumber, cm (B) quadrant 1 quadrant 2 quadrant 3 quadrant Wavenumber, cm Figure V-8. B-vectors from the 4 quadrants of map #81 obtained with (A) 6 and (B) 7 latent variables.

232 204 quandrants, Q1 and Q4, are altogether different in their shape both from Q2, Q3, and each other. Similar water and collagen type I content for quadrants Q2 and Q3, previously reported as the average values in Table V-1, is noted. Apart from this observation, the exact relationship between skin composition at a given location and the shape of the corresponding calibration vector has not been yet elucidated. The calibration b-vectors for 7-factor models in Figure V-8 (B) appear identical apart from the spectral noise supporting the assumption that these models have overcome most of the impact of interquartile variance for map #81. Similarities, however, exist between the calibration vectors obtained with 6 and 7 latent variables. All of the vectors possess a combination of sharp bands between 4500 and 4200 cm and broader bands at around 4650 and 4600 cm -1. Generally, features for -1 PLS calibration vectors in complex mixtures cannot be directly compared with the pure analyte absorption spectral bands. To demonstrate that it is the glucose-specific information that is used for prediction by these PLS models, the NAS calibration vectors were computed for the 4 quadrants in map #81. As explained in the introduction to this chapter, NAS comprises the information in the pure component spectrum of glucose that is unique compared to the skin tissue background. 203 First of all, the background tissue matrix spectra for the 4 quadrants of map #81 devoid of glucose were subjected to single value decomposition (SVD) procedure to reduce the dimensionality of the background spectral set. The resulting uncorrelated principal components are presented in Figure V-9 (A) (D) for quadrants Q1 Q4, respectively. The principal components are arranged in the order of decreasing singular value that reflects the fraction of variance in the background tissue matrix explained by this PC. For all 4 quadrants in map #81, the first six PC vectors, represented by colored solid curves in Figure V-9, account for more than % of the variance in the background spectra. The shapes of these principal components contain spectral features resembling the six standard constituents of skin tissue. For example, the first PC, that explains ~99.65% of total background variance, has

233 Spectral Loadings Spectral Loadings Spectral Loadings Spectral Loadings 205 (A) PC 1 PC 2 PC 3 PC 4 PC 5 PC 6 PC 7 (B) PC 1 PC 2 PC 3 PC 4 PC 5 PC 6 PC Wavenumber, cm Wavenumber, cm (C) (D) PC 1 PC 2 PC 3 PC 4 PC 5 PC 6 PC 7 PC 1 PC 2 PC 3 PC 4 PC 5 PC 6 PC Wavenumber, cm Wavenumber, cm Figure V-9. PC vectors for simulated skin tissue spectra for 4 quadrants of map #81. Principal components for quadrants 1-4 are shown in plots (A)-(D), respectively.

234 206 the shape of a flat scattering background. The second PC looks very much like the spectrum of fat used in the linear regression. Other vectors include some kind of combination of protein and fat spectral features. The shape of the seventh PC, plotted in grey color, is visibly dominated by noise and is not included in the calculation of the NAS calibration vectors in our studies. If the seventh PC is included in the vector projection, the calibration vector will not change its shape but will contain more noise. Calibration b-vectors averaged for the 4 quadrants are compared in Figure V-10 for the PLS and NAS regression methods. The absorptivity spectrum of glucose is also included in this plot. As expected, the PLS vector obtained by means of purely statistical correlation is almost identical to the NAS calibration vector originating from the pure component spectrum of glucose. This finding validates our PLS models as containing chemical information derived from glucose. When compared directly with the pure component glucose absorptivity spectrum, the NAS and PLS calibration vectors appear correlated with the former though the correlation is not perfect. In the region from 4900 to 4700 cm -1, all three spectra are close in shape and all three exhibit peaks around 4700 and 4400 cm -1. The relative magnitudes of spectral features are very different however. The glucose peak at 4300 cm -1 is shifted in the b-vectors to a lower wavenumber. The calibration vectors represent glucose information in multicomponent tissue matrices of the 4 quadrants whereas the pure component spectrum of glucose was obtained from an aqueous solution and can partially overlap with the spectra of the tissue background in the multidimensional vector space. In this sense, comparing the PLS b-vector to an analyte pure component spectrum is typically not valuable. Similarity between the NAS and PLS b-vectors is a much better gauge for selectivity of the PLS calibration. Representative Results of PLS Regression The impact of chemical heterogeneity was assessed between quadrants of map #81 by applying the PLS calibration vector determined from simulated spectra associated with one quadrant to simulated spectra for a different quadrant. The concentration

235 207 Glucose Calibration Vector, mm*mm/au NAS b-vector PLS b-vector Absorptivity Wavenumber, cm Glucose Absorptivity, 10-5 mm*mm/au Figure V-10. Comparison of NAS and PLS b-vectors (averaged for 4 quadrants of map #81) with an absorptivity spectrum of glucose.

236 208 correlation plots are presented in Figures V-11 for quadrants Q1 and Q2 and in Figure V- 12 for quadrants Q3 and Q4. In these figures, predicted glucose concentrations from the calibration quadrant are plotted as red circles and the red solid lines represent a perfect correlation. The blue circles mark the predicted values for glucose obtained from an adjacent quadrant on map #81. The linear fit to these predicted concentrations is added to the plots as the blue solid line. For clarity of presentation, glucose concentration values below 10 mm and higher than 25 mm are not shown but the standard errors and the linear fits were computed for the complete data sets. If the PLS calibration has a perfect performance, calibration and prediction values will all align on the 100% correlation line without any offset or spread. Obviously, this is not the case for any of the 4 quadrants of map #81. Standard errors of calibration and standard errors of prediction for these locations originate both from the random spread of the data around the linear fit and from the observed offsets. In Figure V-12 (B) such an offset is negative and pronounced. The plot illustrates that when the glucose concentrations for quadrant Q1 are predicted with the calibration set of quadrant Q4, the values are consistently lower than the assigned values. Another example is plot (D) in Figure V-11. A positive offset signifies predicted values that are consistently higher than the concentrations assigned in the course of data simulation. Here, examples include plots (A) and (C) in Figure V-12.Similar offsets have been described in the literature for glucose PLS predictions in aqueous solutions of Intralipid. 127 A simple offset removal strategy was implemented that helped to lower the SEP values: the y-intercepts of the linear fits (shown in blue) was subtracted from all of the predicted glucose concentrations and the SEs of prediction were recalculated for the adjusted data. Standard errors of prediction (SEPs) as well as the standard errors of calibration (SECs) and the SEP/SEC percent ratios are summarized for map #81 in Table V-3. The averages for SEP values and percent ratios calculated among all three predicted quadrants are also included. The y-intercepts of the linear fit to the predicted data are included as

237 Predicted Concentrations, mm Predicted Concentrations, mm Predicted Concentrations, mm Predicted Concentrations, mm Predicted Concentrations, mm Predicted Concentrations, mm prediction of q2 linear fit calibration for q1 100% correlation (A) prediction of q1 linear fit calibration for q2 100% correlation (B) Assigned concentrations, mm Assigned Concentrations, mm prediction of q3 linear fit calibration for q1 100% correlation (C) prediction of q3 linear fit calibration for q2 100% correlation (D) Assigned Concentrations, mm Assigned Concentrations, mm y = 0.997*x prediction of q4 linear fit calibration for q1 100% correlation (E) prediction of q4 linear fit calibration for q2 100% correlation (F) Assigned Concentrations, mm y = *x Assigned Concentrations, mm Figure V-11. Concentration correlation plots for quadrants 1 (A, C, E) and 2 (B, D, F) of map #81. Data are shown for concentrations between 10 and 25 mm to increase the visibility of the offset. The concentrations used for calibration are presented by red circles, the predicted concentrations by blue circles. Red solid line demonstrates 100% correlation and the blue line is the linear fit to the predicted concentration values from 5 to 35 mm of glucose.

238 Predicted Concentrations, mm Predicted Concentrations, mm Predicted Concentrations, mm Predicted Concentrations, mm Predicted Concentrations, mm Predicted Concentrations, mm prediction of q1 linear fit calibration for q3 100% correlation (A) prediction of q1 linear fit calibration for q4 100% correlation (B) Assigned Concentrations, mm Assigned Concentrations, mm prediction of q2 linear fit calibration for q3 100% correlation (C) prediction of q2 linear calibration for q4 100% correlation (D) Assigned Concentrations, mm Assigned Concentrations, mm prediction of q4 linear fit calibration for q3 100% correlation (E) prediction of q3 linear fit calibration for q4 100% correlation (F) Assigned Concentrations, mm Assigned Concentrations, mm Figure V-12. Concentration correlation plots for quadrants 3 (A, C, E) and 4 (B, D, F) of map #81. Data are shown for concentrations between 10 and 25 mm to increase the visibility of the offset. The concentrations used for calibration are presented by red circles, the predicted concentrations by blue circles. Red solid line demonstrates 100% correlation and the blue line is the linear fit to the predicted concentration values from 5 to 35 mm of glucose.

239 211 SEP offset values. SEP and SEP/SEC ratios after the removal of the offset in the correlation plots are improved from the original values. These adjusted errors appear in bold font in Table V-3. In this study, the SEC values correspond to prediction errors expected if the fiber interface remains stationary at each quadrant during all measurements, while the SEP values represent prediction errors when the calibration vector is generated with the interface in one position and the predictions are performed at a different location along the tissue matrix. All of these SE values are at least an order of magnitude smaller than in the reallife noninvasive experiment on rats. 37, 112 With this in mind, SEP/SEC ratios are used for comparison between the simulated data in this chapter and the results of the in vivo experiment. Once again, ratios of 100% would be obtained in a situation when the SEP and SEC values are equal and there is no impact from the matrix of the noncalibration quadrants. SE ratios close to 100% are expected when the two, calibration and prediction, quadrants possess similar compositions. Remarkably, the SEP/SEC percent ratios in Table V-3 are close to the 255% value obtained in the anesthetized rat study. 38 The SEPs are about 2 to 4 times greater than the corresponding SECs for all quadrants of map #81. This decrease in the predictive ability of the PLS model is primarily due to the spatial variances across the location on the skin map. The other components included in the simulation of the skin spectra (i. e. glucose concentrations, small temporal variance in six skin constituents, and random noise) were the same for all 4 quadrants. In part, this relative increase in SEPs is due to the uniform offsets seen in Figures V-10 and V-11. When these offsets were removed, the adjusted SE percent ratios went down to between 123 and 164% for map #81, significantly mitigating the effects of spatial heterogeneity on glucose predictions. The ratios never dropped below 100%, which indicates that the spread around the linear fit is always greater for the predicted data than for the calibration data set. Apparently, the effects of tissue heterogeneity also manifest themselves in the spread of the concentrations and cannot be completely

240 212 Table V-2. Results of PLS regression for 4 quadrants of map #81 with 7 latent variables. Calibration Quadrant SEC, mm Predicted SEPs for Quadrants Q2 Q3 Q4 Mean Values Q SEP/SEC Ratio, % SEP Offset, mm Adjusted SEP, mm Adjusted SEP/SEC Ratio, % Q Q1 Q3 Q SEP/SEC Ratio, % SEP Offset, mm Adjusted SEP, mm Adjusted SEP/SEC Ratio, % Q Q1 Q2 Q SEP/SEC Ratio, % SEP Offset, mm Adjusted SEP, mm Adjusted SEP/SEC, % Q Q1 Q2 Q SEP/SEC Ratio, % SEP Offset, mm Adjusted SEP, mm Adjusted SEP/SEC, % Note: Adjusted SEP and SEP/SEC Ratio, % values are calculated after the offset in predicted concentrations had been removed.

241 213 removed by simple offset correction. Moreover, this simplistic correction is not desirable in the simulation study because it would not match well with the in vivo experimentation. In the noninvasive experiment, the optical fiber interface was repositioned after collection of a single spectrum. Every time this was done, the interface could be positioned at a different site on the tissue with varying pressure or even returned to the original location. Because there was only one data point collected between relocations it is impossible to determine if the spread of the data around the black curve of the reference glucose concentrations in Figure I-3 was due to an offset between the predictions at the initial and the newly occupied location or due to random scatter. The offset removal procedure would compromise the integrity of the simulated PLS study by mechanistically adjusting the predicted concentrations after the regression has been performed without accounting for the origins of the offset as part of the calibration process. For this reason, the adjustment was not used with the rest of the rat and human simulated skin data. Effects of Rat Skin Tissue Heterogeneity on Glucose PLS Predictions Statistical data for the unadjusted SEP/SEC percent ratios are presented for male rats 1-8 in Figure V-13. Skin tissue from each animal was characterized by 4 maps in Chapter II. Each of these maps was used for simulation of the tissue spectra for 10 runs. In every run, the average base spectra for fiber locations were unchanged and only the temporal variations and glucose concentrations were represented by different values. However, the mean and the standard deviations defining the temporal variation in the six skin components and the range of glucose concentration values were fixed. With every new run, the MATLAB software randomly selected values according to these pre-fixed distributions. Each boxplot in Figure V-13 (A) contains 4(maps) 4(Qs) 10(runs)=160 values averaged for a given quadrant for the first approach involving 4 locations of the interface. Each boxplot in (B) comprises 4(maps) 36 (locations) 10(runs) = 1440 SE ratios. For all of the boxplots in this chapter, green circles represent mean values, red horizontal lines are median values, red plus signs are outliers, and blue boxes represent

242 SEP/SEC Ratio, % SEP/SEC Ratio, % (A) * 7 8 Male Rat Number (B) * 6 7* 8 Male Rat Number Figure V-13. SE ratios for male rats obtained for (A) 4 quandrants and (B) 36 random locations on the tissue. Outlier values for the rats marked with a star (*) that are greater than (A) 2500% and (B) 10000% are not included in these plots for clarity of presentation and can be found in Table V-4. Mean values are shown as green circles. The slashed black line is added for comparison with the in vivo SEP/SEC ratio of 255%. 38

243 215 Table V-3. Summary of standard errors (SEC, SEP) and standard error ratios (SEP/SEC Ratio, %) for male rats. Rat 1 Rat 2 Rat 3 Rat 4 Rat 5 Rat 6 Rat 7 Rat 8 SEC, mm 4 Quadrants ± ± ± ± ± ± ± ± Locations ± ± ± ± ± ± ± ± SEP, mm 4 Quadrants 0.24 ± ± ± ± ± ± ± ± Locations 0.21 ± ± ± ± ± ± ± ± 0.19 SEP/SEC Ratio, % 4 Quadrants 310 ± ± ± ± ± ± ± ± Locations 280 ± ± ± ± ± ± ± ± 220

244 216 the values between 25% and 75% of the distribution. The dashed black line indicates the ratio of 255% from the in vivo experiment. Table V-3 that follows contains the mean and standard deviation values for unadjusted SECs, SEPs, and SEP/SEC ratios for male rats obtained with both approaches. Even though the absolute values of standard errors for the regions when the fiber is stationary and when it is moved are much lower than for the in vivo case, the mean ratios between the errors for both approaches are between 143 % and 540 %. This range encompasses the ratio of 255% obtained for a living animal. 38 This can be seen both in Figure V-13 and Table V-3. For all of the 8 male rats, 255% value is located between the 25% and 75% distribution of the SE ratios. The results imply that such level of PLS model degradation can be explained by spatial heterogeneity of male rat skin tissue. Two approaches used to model the movements of the optical fiber interface across the tissue produced very similar SEC and SEP values, as expected. Calibration models were created for a stationary fiber with addition of a small amount of temporal variance in each tissue component. These variances are very well accounted for by the computation of the PLS calibration vector with 7 latent variables so that the mean standard errors obtained without fiber movements are between and mm, which is about 7 times lower than the prediction errors obtained for an anesthetized animal. 37 The SEP values reflect the changes in predictive performance when the information about tissue matrix at a new location is introduce. These features may not have been sufficiently modeled by the calibration at the original location and thus adversely affect the prediction of glucose at the new location. The mean standard errors increase to between and 0.51 mm for male rats. SEP/SEC percent ratios represent the relative increase in the standard errors. The differences in results obtained with two approaches to fiber movement modeling emerge as the number and magnitude of the outliers shown as red plus-signs in Figure V-13. Outliers in this figure are the values found outside the whiskers extending to

245 217 Table V-4. List of standard error ratios (SEP/SEC Ratio, %) outliers for rat skin samples, not included in boxplots. Rat # 6 (male) 2 (female) 4 (female) 5 (male) 7 (male) 4 (female) Outlier SEP/SEC Ratio, % 4 quadrants > 2500%: 8.45E+3, 4.85E+3, 3.55E+3, 6.54E+3, 4.74E+3, 3.83E+3 > 2500%: 2.85E+3 > 2500%: 2.69E+3 36 locations > 10000%: 1.27E+4, 1.20E+4, 1.06E+4 > 10000%: 3.62E+4, 3.05E+4, 2.73E+4, 3.80E+4, 3.02E+4, 3.00E+4, 3.87E+4, 2.54E+4, 3.64E+4, 3.66E+4 > 4000%: 4.06E+4, 3.28E+4, 2.52E+4, 2.23E+4, 2.43E+4, 9.05E+3, 2.94E+4, 3.77E+4, 2.13E+4, 3.31E+4

246 interquartile ranges for the SEP/SEC data arranged in the order of increasing magnitude. For clarity of presentation, the percent ratio values higher than 2500% for 4 quadrants procedure plotted in (A) and 1000% for the 36 locations method in (B) are not included in the figure and can be found instead in Table V-4. Modeling of all possible 36 locations on the tissue generates a much greater number of extreme outliers with values up to % whereas the maximum value for SEP/SEC ratio computed for 4 quadrants is 4.5 times lower (8450 %). At the same time, the mean SE ratios are in most cases lower for the 36 locations procedure as seen in Table V-3. A possible reason for this behavior is that when average tissue matrix is simulated for only 4 nonoverlapping locations there is less chance to encounter a drastically dissimilar tissue composition than when 36 locations are probed. Our hypothesis is that an outlier SE ratio value occurs when a new location on the skin is in some way different from the other locations. If this is the case, the PLS calibration model created for this unusual location will poorly predict glucose values in the noncalibration locations. As mentioned earlier, nine times higher SE ratios are observed for the second approach. The relative impact of the outliers is then slightly lowered because most of the 36 locations on the tissue have similar compositions. On the other hand, the outlying SE ratios have greater values because it is possible to capture these unusual locations with a smaller step size of the simulated interface movement. Analogous results were obtained for whole skin from 4 female rats. The values for SEP/SEC percent ratios for both approaches are plotted in Figure V-14. Outliers higher than 2500% and 4000% for the 4 quadrants and 36 locations methods, respectively, are not presented in this figure but are included instead in Table V-4. It should be noted that the cut off percentages used as the basis for the outlier removal are purely arbitrary. They were chosen to be as large as possible to demonstrate most of the outliers but also to ensure visibility of the mean values and the 255% line. Table V-5 summarizes the mean and the standard deviation values for the standard errors of prediction, calibration, and

247 SEP/SEC Ratio, % SEP/SEC Ratio, % (A) * 3 4* Female Rat Number (B) * Female Rat Number Figure V-14. SE ratios for female rats obtained for (A) 4 quandrants and (B) 36 random locations on the tissue. Outlier values for the rats marked with a star (*) that are greater than (A) 2500% and (B) 4000% are not included in these plots for clarity of presentation and can be found in Table V-4. Mean values are shown as green circles. The dashed black line is added for comparison with the in vivo SEP/SEC ratio of 255%. 38

248 220 Table V-5. Summary of standard errors (SEC, SEP) and standard error ratios (SEP/SEC Ratio, %) for female rats. Rat 1 Rat 2 Rat 3 Rat 4 SEC, mm 4 Quadrants ± ± ± ± Locations ± ± ± ± SEP, mm 4 Quadrants ± ± ± ± Locations ± ± ± ± 1.6 SEP/SEC Ratio, % 4 Quadrants 200 ± ± ± ± Locations 170 ± ± ± ± 2400

249 221 the SE percent ratios for female rats. SEC values for female rats are slightly larger than those for male animals (around 0.12 mm) but are several times lower still than the calibration errors obtained in a typical noninvasive experiment. 37, 205 The exact reason for this increase is not clear but may be related to the change in aqueous path length as the female skin samples were thinner. Because the PLS models in this dissertation assumed glucose is dissolved solely in the water layer of the skin, smaller aqueous path lengths could lower the sensitivity of the multivariate model. The effect of lower thicknesses of female rat skin samples on PLS predictions at the noncalibration locations is not as pronounced because tissue matrix heterogeneity is present at similar relative levels for male and female rat skin. Mean SEP values for female rat skin range from to 0.53 mm. SEP/SEC percent ratios for female and male rats are also very close for both algorithms. The average SE ratio obtained for female rats is 221 %, which is very similar to the 255 % change observed in the in vivo experiment. The number of outliers obtained for the 36 locations method is considerably lowered for female rats compared to male rats. Outliers are present for Rat 4 only: one value greater that 2500 % for 4 quadrants and 10 values for 36 locations. We associate this once again with the lower thickness of rat skin and finer features of spatial heterogeneity. More of these smaller domains can fit into a single contact area for the optical fiber interface as simulated in this study for female compared to male rats. Consequently, the chance of locating a tissue matrix with an extreme composition is similar for 4 quadrants and 36 locations. In fact, most outliers for the 36 locations method on the female skin maps are below 2500 %, the cut off value for the 4 quadrants approach. Effects of Human Skin Tissue Heterogeneity on Glucose PLS Predictions Identical procedures, spectral data simulation with two algorithms and PLS prediction, were performed for three specimens of human skin: Specimens 1 and 3 for female donors and Specimen 5 samples obtained from a male donor. In keeping with the

250 SEP/SEC Ratio, % SEP/SEC Ratio, % (A) x 104 1* 3* 5* Specimen Number (B) * 3 5* Specimen Number Figure V-15. SE ratios for human skin obtained for (A) 4 quandrants and (B) 36 random locations on the tissue. Outlier values for the specimens marked with a star (*) that are greater than (A) 5000% and (B) 50000% are not included in these plots for clarity of presentation and can be found in Table V-7. Mean values are shown as green circles. The dashed black line is added for comparison with the in vivo SEP/SEC ratio of 255%. 38

251 223 Table V-6. Summary of standard errors (SEC, SEP) and standard error ratios (SEP/SEC Ratio, %) for human skin samples. Specimen 1 (female) Specimen 3 (female) Specimen 5 (male) 4 Quadrants 36 Locations ± ± SEC, mm ± ± SEP, mm ± ± Quadrants 36 Locations 11 ± ± ± ± ± ± 9.6 SEP/SEC Ratio, % 4 Quadrants ± ± ± Locations 3600 ± ± ± 41000

252 224 rat samples, four maps were spatially characterized per human donor. When simulated skin tissue data doped with glucose was used for location-specific PLS prediction, SE ratios were obtained and plotted in Figure V-15. The numerical summary for the mean SEC, SEP and SEP/SEC values are provided in Table V-6. The standard errors of calibration ranged from to mm for human skin by both procedures. These values are smaller than the SECs for female and even male rats in accordance with the increased tissue thickness and, more importantly, increased aqueous path length for glucose. The absolute differences between locations in regression coefficients for the six skin components are also considerably larger then computed for rat samples. Larger spatial differences drive the SEP values up, as seen in Table V-6. A special case was noted for the first quadrant of map #83 of female Specimen 1. The SE ratio data from all 10 runs of the PLS regression with this quadrant fall outside the cut off value for human skin, 5000 %. These values represent extreme outliers that affect the mean SEP/SEC percent ratio for the whole specimen and raise it to %. Both the mean and the 10 outliers are reported in bold font in Table V-7. The first quadrant of map #83 contains negligible amounts of fat and the highest amounts of keratin protein among all locations on the map. The positive sloping baseline is also very steep at this quadrant. The distribution of these components is more uniform across other locations on the skin slice. Because the domains of spatial heterogeneity in human skin samples, including map #83, are closer to the size of optical fiber than those found in rat skin spatial heterogeneity has a stronger effect on the glucose predictions. For map #83, the unusual levels of keratin and large slope coincide with the position of quadrant Q1 but are not found at the other locations. The contribution from this quadrant is 0.25 by the first approach (4 quadrants) and only 0.03 by the second (36 locations). Not surprisingly, the contribution from this special case is mitigated for the 36 locations method both for SEP and SEP/SEC % values and are highlighted in bold font in Table V-6.

253 225 Table V-7. List of standard error ratios (SEP/SEC Ratio, %) outliers for human skin samples, not included in boxplots. Specimen # 1 (female) 3 (female) 5 (male) Outlier SEP/SEC Ratio, % 4 quadrants > 5000%: 3.31E+5, 4.47E+5, 2.98E+5, 3.81E+5, 4.80E+5, 5.40E+5, 5.10E+5, 4.10E+5, 4.32E+5, 3.23E+5 2.6E+4 (mean value) > 5000%: 5.07E+3, 2.36E+4, 6.26E+3, 1.45E+4, 8.65E+3, 1.53E+4, 1.36E+4 > 5000%: 5.12E+3, 7.42E+3, 1.00E+4, 9.62E+3, 9.57E+3, 8.85E+3, 1.33E+4, 1.89E+4, 6.56E+3 36 locations 1 (female) 5 (male) > 50000%: 1.80E+5, 3.86E+5, 2.32E+5, 3.39E+5, 2.96E+5, 1.84E+5, 1.59E+5, 1.04E+5, 2.01E+5, 1.21E+5, 2.67E+5, 3.88E+5, 4.06E+5, 2.70E+5, 3.26E+5, 8.49E+4, 8.50E+4, 1.30E+5, 2.25E+5, 2.12E+5 > 50000%: 4.37E+5, 4.00E+5, 4.04E+5, 6.04E+5, 4.44E+5, 1.61E E+5, 2.26E+5, 1.14E+5, 2.63E+5, 4.45E+5, 5.00E+5, 3.86E+5, 4.57E+5, 4.25E+5, 5.60E+4, 1.68E+5, 1.38E+5, 1.87E+5, 1.42E+5, 2.14E+5

254 226 On the whole, the SEP/SEC ratios used in this dissertation as the marker of skin functional heterogeneity are on the order of several thousand percent for human skin samples. Consistently with the effect of larger domains of six components described previously for male and female rats, the number of SE ratio outliers is increased in human skin for the 36 locations method. Conclusions Effects of repositioning the optical fiber interface used to collect noninvasive spectra of living skin tissue were explored in this chapter by using PLS regression models based on microspectroscopic near-infrared analysis of both rat and human skin tissue samples. Simulation of fiber movement was performed with two approaches: separating each tissue map into 4 nonoverlapping quadrants and exploring all 36 possible locations for the fiber approximated in size by a 5 5 block of the spectral map. Thickness of the tissue and the size of spatial domains of the skin constituents differ for male and female animals as well as for humans. These chemical heterogeneities create increases of various magnitudes in the prediction errors for PLS calibration models when simulated measurements are made through regions of skin not used to generate the calibration model for glucose. The PLS calibration models from synthesized data were generated in the same way as for the in vivo experiment. However, the rank of the model was limited to a maximum of seven factors in order to not exceed the number of tissue components. The fractional coefficient for water served as an internal standard for the calibrations. Our findings suggest that chemical heterogeneity within the skin tissue matrix can easily explain the previously measured increase in glucose prediction errors of 255% when the fiber interface was repositioned before each measurement. Occasionally, SE percent ratio between two very different locations on the skin increases dramatically. Still, the basic structure of the calibration vectors is similar for most locations. The similarity in these

255 227 calibration vectors suggests that the accuracy of such glucose concentration predictions depends strongly on the exact features of these vectors and subtle differences can have a profound impact on measurement performance. The simulation study in this chapter demonstrates how complexity of the skin matrix in rats and humans can impose constraints on non-invasive near-infrared measurements of glucose with an optical fiber interface.

256 228 CHAPTER VI IMPROVEMENTS TO THE SIX COMPONENT SKIN TISSUE MODEL ACHIEVED WITH RAT AND HUMAN FAT SPECTRAL STANDARDS Introduction Lipids in Skin Lipids are essential components of skin tissue that help maintain its barrier and transfer functions. 146, 168, 183, 206 Lipid molecules are present in the epidermis layer where they are located between corneocytes and keratinocyte cells. The conventional model for stratum corneum consists of the flattened corneocytes arranged in stacks inside the lipid matrix. 207 Modifications to these protein-lipid junctions regulate permeability of the skin to extraneous species, including harmful antigens, carcinogens, and toxins from the environment. 208 The stratum corneum (SC) lipid-protein barrier prevents transepidermal water loss. Epidermal lipid content depends on species, diet, gender, and skin diseases. 133, 136, 137, 167, 209, 210 Conditions like psoriasis, ichthyoses, and eczema alter skin lipid profiles. 210 In the dermis skin layer, lipids are much less abundant and concentrated in hair follicles and sebaceous glands. 151, 211 Major components of the dermal extracellular matrix (ECM) are hydrated filaments of proteins, collagen being the major dermis protein. 144, 212 The lower amount of dermal fat is consistent with the different functions of the epidermis and dermis layers of skin. 143, 147 The subcutaneous layer is mostly composed of fat cells and has far less protein components than either the dermis or epidermis layers. 17, 132, 144 The functions of the subcutaneous fat layer are protection from physical damage and maintaining of body temperature. The subcutaneous layer of rat and human skin is not considered in detail in the presented work because it was removed from the skin samples in Chapters III and IV. The in vivo glucose measurement on rats was performed through the first two layers of skin. 38, 39, 112 Examples of epidermal lipid molecules are shown in Figure VI-1. These represent the major classes of ceramides

257 229 (A) (B) ` (C) (D) (E) Figure VI-1. Chemical structures of major classes of skin lipids: (A) fatty acid (palmitic acid shown), (B) ceramide, (C) cholesterol, and (D) sphingomyelin. 1 Structure of n-hexadecane is shown in (E).

258 230 (~50% of epidermal lipids), cholesterol (25%), sphingolipids, and free fatty acids. Generally, lipids in skin contain are presented by polar and nonpolar molecules. For this reason, complete extraction of skin lipid usually involves several extraction steps or the use of solvent mixtures. 158, 213, 214 The relative amounts of polar and nonpolar lipids have been found to vary depending on the species. 167 In an HPTLC study, the highest total content of lipids was found in rat epidermis followed by bovine udder, dog, and pig samples. 158 From the point of view of vibrational spectroscopy, lipid molecules contain a number of characteristic absorption features both in the fundamental infrared 172, 181 and the near infrared regions. 55, 172, 181 Vibrational microspectroscopic measurements of lipids in skin and epithelia have gained importance in the last 15 years for drug and cosmetics permeability research and studies of atherosclerosis. 57, 215 In the latter application, the ratio of protein NIR signal and the lipid-associated absorption indicates the stability of the plaque. Lipid signal is also included as an important parameter in the vibrational spectroscopy of cancerous tissue. 56 In relation to the goal of noninvasive glucose detection pursued by our group, lipid absorption in the combination region of the near infrared ( cm -1 ) is a major interference. The features arising from combinations of C-H stretching and bending fundamental modes are very pronounced in NIR spectra of lipids and overlap with the analytically important glucose absorptions at 4300 and 4400 cm -1. In all of the structures in Figure VI-1, saturated hydrocarbon chain fragments are present. The structure of n-hexadecane is added for comparison in Figure VI-1 (E). This alkane contains the backbone of 16 carbon atoms all of which are bonded with at least two hydrogen atoms. The combination spectrum of n-hexadecane collected is provided in Figure VI-2. The black dotted line separates the lipid absorption due to the combination of C-H modes. The two strongest peaks are measured at 4330 and 4260 cm -1 for n- hexadecane. Similar spectral shape is observed for the fat standard plotted in Figure II-4.

259 Absorbance, AU Wavenumber,cm Figure VI-2. Near infrared spectrum of n-hexadecane collected with IR microscope in the combination region ( cm -1 ). The black dotted line separates the region of cm -1 where C-H vibrational peaks are prominent.

260 232 Because of these fat absorptivity features, glucose measurement sites on rats, the upper shoulders, and humans, the skin on the knuckle of a hand, were selected to minimize the subcutaneous fat content in the noninvasive spectra. Unfortunately, there is currently no way to avoid the presence of epidermal fat in these noninvasive spectra. These features are observed as two slight peaks at 4330 and 4250 cm -1 in rat and human skin spectra presented in Chapters III and IV. Residual Spectral Features of the Skin Tissue Model In this dissertation, the six component fit described by Equation II-6 is used to model rat and human whole skin absorbance in the combination NIR region. In particular, tissue absorbance data between 4900 and 4200 cm -1 are least-square fitted with the linear combination of spectral standards of water, collagen type I protein, bovine fat, keratin protein, constant offset, and sloping baseline. The product for a single measured point on the skin slice is the set of six regression coefficients, β, that signify the abundance of each component in the sampled tissue. The details of the procedure and the multivariate analysis of the numerical results have been described in Chapters II-V. With the bank of the regression coefficients for major skin components, simulation of a huge number of combination NIR skin spectra at independent locations becomes possible. However, small differences exist between the simulated skin spectra and the absorbance of skin samples measured ex vivo. 129, 186 These differences are reflected in the shapes of spectral residuals, ε, also computed according to the least squares model in Equation VI-6. The magnitude of these residuals is about 1 2% of the total measured tissue absorbance. The average residual spectra for a rat and a human skin map are plotted in blue and red, respectively, in Figure VI-3. The six modeling standards are noise-free. Obviously, all of the spectral noise in the actual skin spectra should end up in the residuals and become randomly distributed around zero. However, a number of nonrandom features are observed in the residuals in Figure VI-3. These have not been eliminated by averaging of the 100 locations for each

261 Absorbance, AU rat skin residual human skin residual Wavenumber, cm Figure VI-3. Average spectral residuals after the six-component fit for samples of (A) rat skin and (B) human skin. Bovine fat of 1 mm thickness is used as the fat standard.

262 234 map. Sharp positive and negative peaks are notable in the cm -1 region that indicate suboptimal modeling of the methylene groups associated with lipids in the skin matrix. The shape and magnitude of the residuals vary depending on the mapped location. Examples of the six-component fit deficiencies are provided in Figure VI-4 for rat (A) and human (B) skin. In this figure, the actual measured whole skin spectra are represented by solid blue and black lines; the best fits are added as the green and the red dashed line spectra. These examples reveal pattern characteristic for most of the samples analyzed for this thesis: discrepancy between the actual and the regressed absorbance is relatively larger around the two lipid-related peaks are 4,330 and 4,255 cm -1 and the magnitude of the discrepancy increases with an increase in fat content in the skin tissue sample. The solid black and green curves (higher fat coefficient) are visibly separated whereas the dotted red and the solid blue curves (lower fat content) match well throughout the combination region. This trend suggests that the bovine fat pure component may not be an ideal match for the fats found in rat and human skin. The use of native fat samples is proposed instead. Chapter Overview This chapter describes efforts to minimize the nonrandom features in the spectral residuals and improve the quality of the six-component regression fit for rat and human whole skin spectra. Following the reports of species-dependent profile of epidermal lipids we collected pure component standards of fatty tissue from rat and human sources alternative to the original bovine fat spectrum. These were substituted for the bovine fat standard and the resulting residuals were compared. Both shape and the magnitude of the nonrandom features in the residuals were taken into account in this comparison. Principal component analysis (PCA) was used to decompose the location-related variances in the spectral residuals into independent (orthogonal) loadings. Inspection of these loadings was performed to elucidate the effects of fat standard substitution.

263 Absorbance, AU Absorbance, AU 235 (A) rat skin spectrum with low fat rat skin spectrum with high fat regressed spectrum with low fat regressed spectrum with high fat Wavenumber, cm (B) human skin spectrum with low fat human skin spectrum with high fat regressed spectrum with low fat regressed spectrum with high fat Wavenumber, cm Figure VI-4. Six component linear fit for (A) rat and (B) human skin spectra with high fat, represented by blue solid line, and low fat content, represented by black solid line. The corresponding regressed spectra are shown as the red dashed curve for the low and as the solid green curve for the high fat coefficient.

264 236 Experimental Procedures Preparation of Fat Tissue Samples Abdominal fat was collected from the male Rat #6 immediately after sacrifice and at the same time as the collection of skin between the shoulders for microspectroscopic mapping. The area on the upper back of the rat did not contain a visible layer of fat under the skin. For this reason, fat tissue from the abdominal region was used instead. Since the animal had been bled beforehand according to the procedure described in Chapter III, the blood content in the fat layer was quite low. Abdominal fat was blotted dry with Kimwipes and snap frozen in liquid nitrogen. The frozen samples were stored in the freezer compartment at -18 C. Two days before the measurement, a penny-size sample of this rat fat was thawed overnight in the refrigerator at 3 C and was subsequently placed in the desiccator with fresh desiccant. Periodically, the weight of the sample was measured with Mettler A160 electronic balance (Mettler-Toledo, Inc., Columbus, OH). When the sample of fat reached a constant mass, the desiccation was stopped and the sample of dried fat tissue was measured microspectroscopically. For the pure component spectrum of human fat, a sample of subcutaneous fat was removed from a whole skin sample from Specimen 5 (male donor, 19 y. o.). Specimen 5 tissue was generously shared by Dr. Bickenbach from the Department of Biochemistry, University of Iowa. The tissue was received with the subcutaneous layer still attached. After washing with PBS, the fat was carefully separated from the dermis and epidermis layers, blotted dry, and refrigerated in the desiccator at 3 C until its weight stabilized. Instrumentation and Spectral Data Collection Measurements were performed with an IR Plan Advantage microscope (SpectraTech, Inc., Shelton, CT) coupled with a Nicolet Magna 560 Fourier transform infrared spectrometer (Nicolet Instrument Corp., Madison, WI). The specifics of the instrumental set up are discussed in Chapter II. The microscope featured an RGB CCD

265 237 vision camera module (model XC-711, Sony Electronics Inc., San Diego, CA) that allowed choosing the location for the measurement. Procedures for the microspectroscopic measurements were identical for rat and human fat samples. A set of air reference spectra was obtained in triplicate from an empty compression cell described in Chapter III. After a sample was sufficiently desiccated, it was removed from the freezer, wiped dry with a Kimwipe, and placed between the sapphire slides inside the compression cell. The cell was equipped with Teflon spacers, 0.47 mm thick for rat fat and 0.93 mm thick for human fat. After 30 minutes at room temperature, the tissue measurement with the microscope was performed in triplicate in the combination region ( cm -1 ) with a nominal point spacing of 3.8 cm -1 and a resolution of 7.6 cm -1. First, the location with the least amount of visible inhomogeneities, such as air bubbles, was selected. Then, 2048 and 512 interferograms were accumulated for rat and human samples, respectively. The S/N ratio for the microscope was low at the time of the rat fat measurements. To account for this poor signal quality, more co-added scans were used. The structure of the two dried fat samples was very different. After desiccation, the sample of rat abdominal fat appeared as a transparent liquid core surrounded by the rigid dry membrane. When this sample was compressed between the sapphire slides, the membrane split and the rat fat pure component spectrum was obtained from the liquid fraction only. The human subcutaneous fat sample contained multiple spherical fat cells when viewed under the microscope. Structure of the human fat did not change upon compression and it appeared opaque even after desiccation. These differences can be attributed to various sizes of the fat samples and the fact that the human sample was measured on the same day and had never been frozen. Freezing and subsequent thawing might have promoted breakage of the fat cell membranes due to crystal formation. Also, the amount of human fat collected was about twice greater than the rat fat. This greater amount allowed switching to the thicker Teflon spacer of 0.93 mm and, hence, longer

266 238 path length for the human fat standard. The original fat pure component spectrum was collected as a transmission spectrum through a 1 mm thick sample of bovine fat. Data Processing All spectral data were collected with the OMNIC (Nicolet Instrument Corp., Madison, WI) software and spectral processing was done in MATLAB (Version 7.0, The MathWorks, Inc., Natick, MA). Tissue absorbance spectra were calculated by taking the negative log of the ratio of the single beam spectra for fat and air. The fat and air single beam data were averaged before calculating the absorbance of the pure component to minimize random noise in the resulting spectral standards. For the murine sample, the absorbance was multiplied by 2 in accordance with the Beer s law to approach the path length of the original bovine standard. The absorbance for the human fat was used as the new standard without further modifications. The differential absorbances were obtained for bovine and native fat spectra as given in Equation VI-1. A dif = A b A n Equation VI-1 The values of the differential absorbance are represented by the A dif spectrum, A b and A n correspond to the bovine fat standard absorbance as plotted in Figure II-3 and the new native (rat or human) fat absorbance standard. The features of the A dif spectrum highlight dissimilarities between the old bovine standard and the native fat pure components. The new fat standards were substituted for bovine fat term in the linear regression model for skin tissue presented in Equation II-6. Two representative skin maps were chosen for testing of the native fat models: map #49 collected for male Rat 6 tissue and map #88 collected for male Specimen 2 skin. The former was chosen because of its appreciable fat content. The latter also produced fat coefficients larger than zero, but more importantly, Specimen 2 was the only human skin sample that did not contain subcutaneous fat when received, which helped lower the chances of contamination of the whole skin spectra with fat from the subcutis. Once again, our goal was to improve the

267 239 regression fit of skin dermis and epidermis layers only as these are the two layers 37, 38, 205 primarily probed by the in vivo measurements. In addition to the sets of regression coefficients, the linear fit of the two representative maps yielded sets of unmodeled spectral residuals, ε, 100 per map. These were split into two complementary regions: 4400 to 4200 cm -1 and 4900 to 4400 cm -1. The sums of squared residuals (SSRES) were computed for both regions and for the whole range of cm -1 as indicated in Equation VI-2: n 2 SSRES A i A i Equation VI-2 i 1 _ fit where A i and A i_fit are the measured and the fitted absorbances at location i on the tissue, n is the total number of locations, 100 for each map. SSRES values represent the absolute magnitude of the residuals irrespective of sign. The lower SSRES indicate an overall better quality of the fit. The parameter was computed for the original fit of maps #49 and #88 and the fit after fat standard substitution. Correlation coefficients, r, were obtained for the SSRES and the respective fat component according to Equation III-2. The expression for the standard error of the Pearson coefficient is provided in Equation III-3. PCA on Residual Spectra Singular value decomposition (SVD) was used to demonstrate the changes in the spectral residuals after substitution of the native fat spectrum for the original bovine fat standard. The principles of the singular value decomposition can be found in the introduction section of Chapter V. Briefly, this statistical technique decreases the dimensionality of a data set in a multidimensitonal vector space by extracting the independent eigenvectors or principal components that describe the most of the variance between the spectra, or vectors. The shapes of the latent variables and their eigenvalues provide clues to the sources of variance in the data set. The procedure was performed for the residuals of both maps between 4900 and 4200 cm -1 obtained with different fat standards.

268 240 Results and Discussion Native Fat Standards The initial strategy to improve the quality of the six component fit was to substitute a spectrum of skin lipids for the bovine fat spectrum. To that end, lipid extraction from dried human skin was performed according to Folch s procedure. 213 The amount of fat removed from the skin with the 2:1 v/v mixture of chloroform and methanol constituted 3% of the total sample weight. This value was in agreement with literature. 171 However, two obstacles prevented collection of quality near infrared spectra of this fat extract. First of all, a large fraction of skin lipids are nonpolar and are not soluble in water, 214 which compromises the quality of spectra collected for lipid/water mixtures because of scattering. If only the spectrum of the aqueous phase were collected the skin lipids would not be fully represented in the new standard. On the other hand, organic solvents, methanol and chloroform, could not be used because they exhibit strong absorption features in the combination region. In particular, the spectrum of chloroform contains an intense peak around 4250 cm -1 that can introduce error into the spectrum of native fat. Second, direct microspectroscopic measurements of the skin lipid product were attempted. With those, success was only marginal: we were able to create a very thin human fat mirror between two sapphire windows but the amount of skin lipid produced from a single 1 cm 2 slice of human tissue was too small to provide a sufficient path length for the measurement. The microspectroscopic data were dominated by noise and scattering due to different interfaces between the sapphire and fat and had no analytical value. In the future, combining fat extracts from several skin slices might provide a better skin lipid standard spectrum. An alternative method was developed, that used a more abundant source of native fat. Samples of abdominal rat fat and subcutaneous human fat were collected and processed as previously described. Comparison between these and the original bovine fat

269 Absorbance, AU Absorbance, AU 241 (A) bovine fat standard rat fat standard Wavenumber, cm (B) average residual for rat skin sample #49 difference between bovine and rat fat spectra Wavenumber, cm Figure VI-5. Comparison of bovine fat spectrum and the rat fat spectrum. In (A), pure component spectra are shown; in (B), the average residual for rat skin sample #49 is compared with the differential absorbance spectrum between rat and bovine fat spectra.

270 Absorbance, AU Absorbance, AU bovine fat standard human fat standard (A) Wavenumber, cm (B) average residual for human skin sample #88 difference between bovine and human fat spectra Wavenumber, cm Figure VI-6. Comparison of bovine fat spectrum and the human fat spectra. In (A), pure component spectra are shown; in (B), the average residual for human skin sample #88 is compared with the differential absorbance spectrum between human and bovine fat spectra.

271 243 spectrum is seen in Figure VI-5 for the rat fat and in Figure VI-6 for the human fat standard. Even though the exact composition of these tissues may not be the same as the skin lipid fraction, the motivation was that the spectra of fat coming from the same species would still model rat and human skin tissue better than the bovine fat spectrum. In Figures VI-5 (A) and VI-6 (A), the blue solid line corresponds to the original bovine fat spectrum, collected with an FTIR spectrometer, and the red and green solid lines represent the rat and human fat spectra, respectively. Differences in the position, width, and magnitudes of the absorption features are evident. Firstly, all three types of fat demonstrate two strong absorption peaks in the cm -1, called C-H region in this thesis. The exact position of these peaks varies for different species. These peaks are red-shifted in bovine fat absorbance spectra relative to both human and rat fat. Secondly, the magnitude of the absorption is different for the three fat standards. This can be explained by differences in the optical path length. 101, 111 Moreover, the density of fat in the bovine fat sample could be externally controlled whereas in the microspectroscopic experiment no such control was attempted. Dissimilarity in the human and bovine fat standards is seen in Figure VI-6 (A): the scattering background for the human fat appears elevated, likely because of an effect of the heterogeneous structure of fat tissue as seen in the CCD camera images for the human fat sample. The rat fat in Figure VI-5 (A) is comparatively similar to the bovine standard in both magnitude of absorbance and the scattering background. The differential absorbance spectra for rat and human fat calculated according to Equation VI-1 can be seen as the red and green spectra presented in plots (B) in Figures VI-5 and VI-6, respectively. If these spectra were identical, such differential absorbance profiles would be a horizontal line located at zero with possibly a negligible amount of randomly distributed noise. Instead, the differential absorbance spectra exhibited several nonrandom features. As expected from the shapes in plots (A), most of the difference between bovine fat and native fat spectra is concentrated around the C-H absorption

272 244 peaks or the fat-related peaks. In both figures, the blue dotted lines are added for the corresponding average spectral residual obtained with the bovine fat as part of the six component model for rat and human skin. The residual absorbance values were multiplied by 10 to provide a comparable magnitude. These residuals were previously plotted together with their original scale in Figure VI-3. Curiously, the negative dips in the average residual spectra in the region of cm -1 correlate in part with the spectral shifts in the absorbance peaks for both species. When the bovine fat is used, the six components least squares fit accounts for the relatively strong fat absorbance peaks by producing the negative dips in the residuals. With this in mind, the strong positive differential absorbance peaks for both native fat samples provide hope that these new standards can be a better match, thereby producing lower residuals and superior fits. Six Component Regression Model with Bovine and Native Fat Standards The new native fat standards were tested in the six component regression. Results of these fits are shown in Figure VI-7 for the rat skin sample #49 and in Figure VI-8 for the human skin of map #88. Plots (A) and (B) represent results for when the bovine fat standard is used for the 100 mapped locations. Plots, (C) and (D), represent the same results when the new native fat standards were used. In plots (B), sharp C-H peaks dominate the residual profile especially for the human skin sample #88. The magnitude of the residuals in the higher wavenumber region of cm -1 is small in comparison. The residuals for the map #49 of rat skin appear more balanced but nevertheless fat-related features are still the major contributors. When native fat standards are substituted, a visible decrease of the fat-related spectral residuals around 4,330 and 4,255 cm -1 is notable in plots (D) of Figures VI-7 and VI-8. Now the C-H absorbance peaks are on the same or even lower scale compared with the cm -1 broad band residuals. The spread in the residual absorbance values at the 4,330 and 4,255 cm -1 peaks has also decreased. In other words, the residuals obtained for the native fat models are overall smaller and more uniform across both the spectral region across all

273 Regression Coefficient Absorbance, AU Regression Coefficient Absorbance, AU 245 (A) (B) water collageni fat keratin constant slope Mapping Location Number Wavenumber, cm (C) (D) water collageni rat fat keratin constant slope Mapping Location Number Wavenumber, cm Figure VI-7. Regression coefficients and spectral residuals for rat sample #49 obtained with (A), (B) bovine fat standard and (C), (D) rat fat standard.

274 Regression Coefficient Absorbance, AU Regression Coefficient Absorbance, AU 246 (A) (B) 0.4 water collageni fat keratin constant slope Mapping Location Number Wavenumber, cm (C) (D) 0.4 water collageni human fat keratin constant slope Mapping Location Number Wavenumber, cm Figure VI-8. Regression coefficients and spectral residuals for human sample #88 obtained with (A), (B) bovine fat standard and (C), (D) human fat standard.

275 locations on the maps. Plots (A) and (C) in Figures VI-7 and VI-8 demonstrate the regression coefficients for the 100 point of the skin maps. In (A), the coefficients were obtained with the original six component model including the bovine fat standard. It is these results that were used for simulation of skin tissue absorbance spectra and the multivariate regression in Chapter V. As noted in the (C) plots, coefficients obtained with the native fat standards are mostly unchanged for water, keratin protein, and slope. The magnitude of fat absorbance is smaller for rat and human fat, so naturally the regression coefficients for the native fat are universally larger. In Figure VI-8 (A) and (C), the constant offset term values are affected by the substitution of the fat standard. This is expected because the human fat spectrum in Figure VI-6 (A) carries a substantial scattering background compared with the bovine standard. Part of the constant offset is modeled by this new fat term and the resulting regression coefficient for the offset term is lower. In addition to the changes in the fat and constant terms, the coefficients for collagen type I protein change the most for both species as seen in the shape of the green trace in plots (A) and (C). These changes are not entirely surprising. Substantial overlap can be found between absorption features in the lipid spectra and the pure component spectrum of collagen in the cm -1 range (Figure II-4). The origins of the overlap are not completely clear. A small percentage of lipid substances can potentially be present in the commercially available collagen samples. 216 On the other hand, the protein side chains contain numerous C-H bonds that may contribute to these overlapping absorption peaks. On the whole, the results of the six component fit with the native fat standards are very encouraging. For one, the residuals, whose sum of squares are listed in Table V-1, decrease in magnitude and are less weighted towards a particular region of a spectrum. Also, the regression coefficients for water experience only minor changes with the substitution. The water coefficients can be used in place of the aqueous path length in the simulation of skin tissue containing glucose 186 or, for that matter, any water soluble

276 248 Table VI-1. Sum of squares of residuals (SSRES) for rat and human skin samples obtained with the original bovine standard and the new native fat standards. Spectral Region, cm -1 SSRES (mean ± std dev), AU 2 Rat Map #49 Bovine Fat Rat Fat ± ± ± ± ± ± Human Map #88 Bovine Fat Human Fat ± ± ± ± ± ±

277 249 metabolite. The stability of these values with the new fat standard proves that the aqueous path length in skin can be determined with the six component fit independently of the fat component. The change in the collagen protein content may prove significant for tissue heterogeneity studies but the exact effects of this change are yet to be evaluated. Data in Table V-1 provide numerical evidence of the improvements to the six component fit with the native fat substitutions. For rat and human tissue, the overall SSRES value decrease by 14 and 57 %, correspondingly. Furthermore, the SSRES in the region of C-H vibrations ( cm -1 ) decreased by 18% for rat species and by as much as 70 % for the human skin sample. The residuals between 4900 and 4400 cm -1 were not improved that much (6 % and 12 %), respectively, which is not surprising considering that this spectral region contains small absorption features for fat molecules. Any adjustments in these weak absorbance peaks though beneficial will not greatly affect the overall residual. For rat skin sample, the residuals before and after are somewhat weighted towards the cm -1 region. A little more than a half of the total SSRES is due to these features. For the human map, however, the substitution equally affects both regions. Lipid-related residual peaks having become lower in magnitude, the residual nonrandom features in the cm -1 range starts to contribute more than 50% of the total SSRES. These SSRES values are consistent with the balanced residual spectra across wavenumbers in Figure VI-8 (D). If the perfect fit of these spectral maps were accomplished, the residual spectra would only contain random noise features uniformly distributed around zero. The magnitude of this noise would be characteristic of the quality of the microspectroscopic measurement and no important analytical information about skin composition at each location would be lost. Such a perfect fit has not been realized to date. However, the new native fat standards provide lower nonrandom residual features, especially for the human skin sample. Shift of the peak positions in the cm -1 range is lower for these spectra compared with the original bovine standard.

278 250 Correlation between Fat Content and SSRES Values In the introductory section of this chapter, the accuracy of the six component linear fit containing the original bovine fat standard was demonstrated to vary between locations on the same tissue slice for rat and human skin. The phenomenon can be visualized by comparison of the two-dimensional spatial distribution maps for fat coefficient and SSRES. Such maps are presented in Figure VI-9 for the bovine fat coefficient in (A) and (B) plots and for the native fat coefficient in (C) and (D) plots for rat skin map #49 and human skin map #88, respectively. The plots use higher density of orange color to denote the locations with low fat content. Bright yellow and white spots produced the higher coefficients for the corresponding fat standard. The solid black contours separate areas for which the change in fat coefficient is at least 10 % of the whole range of fat values in the mm 2 tissue sample. The maps in Figure VI-9 are yet another example of the differences in the typical sizes of heterogeneous domain found for rat and human skin samples. For human skin, the dimensions of the island of high fat content in plots (B) and (D) are on the order of 1 mm. Similar features in rat skin have dimensions a little under 0.5 mm as gauged by visual inspection of plots (A) and (C). Irrespective of the size of these tissue domains, the distribution of the native fat standards matches the original distribution really well. The coefficients for these new fat terms are greater for the reason discussed earlier but the spatial arrangement remains unchanged. Figures VI-10 and VI-11 illustrate the spatial variance in the quality of the models fit. In both figures, SSRES for the whole combination spectral region are plotted in (A) and (B); for the C-H absorption region between 4400 and 4200 cm -1 in (C) and (D); and for the absorption between 4900 and 4400 cm -1 - in (E) and (F). Bovine fat standard was used for (A), (C), and (E) plots. Native fat standards were used for plots (B), (D), and (F). Black contours denote 10% change in the SSRES.

279 251 (A) (B) (C) (D) Figure VI-9. Spatial distribution plots for bovine fat regression coefficients of (A) rat map# 49 and (B) human map #88. The distributions for native fat standards are shown for the same maps in (C) and (D), respectively. Black contours denote 10% change in regression coefficient.

280 252 (A) (B) (C) (D) (E) (F) Figure VI-10. Spatial distribution plots for sum of squares of spectral residuals (SSRES) for rat skin sample #49. SSRES for the whole combination spectral region, cm -1, are plotted in (A) and (B); for the C-H absorption region between 4400 and 4200 cm -1 in (C) and (D); and for the protein related absorption between 4900 and 4400 cm -1 in (E) and (F). Bovine fat standard was used for (A), (C), and (E) plots. Rat fat standard is used for for (B), (D), and (F). Black contours denote 10% change in the SSRES.

281 253 (A) (B) (C) (D) (E) (F) Figure VI-11. Spatial distribution plots for sum of squares of spectral residuals (SSRES) for human skin sample #88. SSRES for the whole combination spectral region, cm -1, are plotted in (A) and (B); for the C-H absorption region between 4400 and 4200 cm -1 in (C) and (D); and for the protein related absorption between 4900 and 4400 cm -1 in (E) and (F). Bovine fat standard was used for (A), (C), and (E) plots. Human fat standard is used for (B), (D), and (F). Black contours denote 10% change in the SSRES.

282 254 Consistent with the size of the spatial domains for regression coefficients obtained for rat and human skin, the SSRES are localized into smaller features for rat samples, both when the bovine standard was used and when the native rat fat was substituted. The distribution of SSRES for the bovine fat model is somewhat correlated with the fat coefficient map in Figure VI-9 (A), especially for the cm -1 region as shown in Figure VI-10 (C). This correlation is lost for the SSRES in the cm -1 region presented in Figure VI-10 (E). The series of plots on the right hand side of Figure VI-10, that is (B), (D), and (F) plots, demonstrates the effects of substitution of the fat standard on the spatial variance of SSRES values. The most significant result corroborated by all of the three maps is that the correlation of SSRES with the native fat distribution is diminished in all of the spectral regions. In other word, the fat present in rat skin is modeled better with the native abdominal fat spectrum than by the bovine fat standard. The trend described for the rat skin is even more visible in the distribution maps for SSRES for the human skin sample in Figure VI-11. Sample #88 contains two large domains of elevated fat coefficient in the upper left hand corner of the maps (B), for bovine fat, and (D), for the native human fat, in Figure VI-9. For the bovine fat standard, the location and the shape of these domains are conserved in two of the three the SSRES distribution plots, (A) for the whole range and (C) for the cm -1 region. In Figure VI-11 (E) though, SSRES appear to be inversely correlated with the bovine fat content but the magnitude of the residuals in is much lower so this inverse correlation does not counterbalance the positive correlation of the cm -1 SSRES. The plots (B), (D), and (F) on the right hand side in Figure VI-11 show considerable improvement achieved with the native fat distribution: the magnitude of residuals decreases and neither of the three maps appear positively correlated with the human fat map in Figure VI-9 (D). Conversely, the negative correlation is present which suggests in part that the lipid-related residuals are no longer the major contributors. The SSRES values from cm -1, inversely correlated to the fat content from the

283 255 start, now play a more significant role. The broad band features at the higher wavenumbers are attributed to the skin proteins, collagen type I and keratin. 49 For both rat and human maps, inverse correlation between fat and keratin protein was measured. This may explain the relationship between the fat coefficient and the SSRES at the wavenumbers higher than 4400 cm -1. More detailed study of the correlations between the magnitude of the spectral residuals and the regression coefficients of other skin components is needed to support this suggestion. Table V-2 summarizes the Pearson correlation coefficients, R, for SSRES and the corresponding fat coefficients. The values are accompanied by the corresponding standard errors. The upper part of the table contains correlation coefficients obtained with the bovine fat standard. The R-values for the native fat spectra follow in the bottom part. Comparison of the two parts demonstrates the complete removal of the positive correlation between the SSRES and the fat content both for rat and human samples. The effect is the greatest for the human sample #88. This may be because the map has a domain of high fat content. PCA of the Spectral Residuals Principal component analysis (PCA) for rat and human skin samples was used to related the effects of the fat standard substitution on the overall variances in the spectral shapes of the residuals. After the SVD procedure, the percent variance explained by each loading was calculated and the first six unweighted loading vectors for the residuals obtained with the bovine fat and the new native fats were plotted in Figure VI-12 for the rat skin sample #49 and in Figure VI-13 for the human skin sample #88. Plots (A), (C), (E) were obtained for the residuals of the six component model with the bovine fat standard and plots (B), (D), (F) with the native fat standards. All our previous observations are supported by these plots. The first principal components (PCs), plotted as dashed black curves, for the residuals obtained with bovine fat account for % for rat skin and % of variance for the human skin sample.

284 256 Table VI-2. Pearson coefficients, r, for correlation between bovine and native fat content and sum of squares of residuals (SSRES). Pearson correlation coefficient (bovine fat standard), Spectral Region, cm -1 r ± std error Rat Map #49 Human Map # ± ± ± ± ± ± 0.05 Spectral Region, cm -1 Pearson correlation coefficient (native fat standards), r ± std error Rat Map #49 Human Map # ± ± ± ± ± ± 0.04

285 PCA Loading PCA Loading PCA Loading PCA Loading PCA Loading PCA Loading 257 (A) (B) PC 1 (70.75 %) PC 2 (8.05 %) PC 1 (71.89 %) PC 2 (5.00 %) Wavenumber, cm Wavenumber, cm (C) (D) PC 3 (2.29 %) PC 4 (1.56 %) PC 3 (2.30 %) PC 4 (2.02 %) Wavenumber, cm Wavenumber, cm (E) (F) PC 5 (1.28 %) PC 6 (0.77 %) PC 5 (1.30 %) PC 6 (0.82 %) Wavenumber, cm Wavenumber, cm Figure VI-12. PCA loadings for rat skin sample #49 with 1-6 latent variables obtained with (A), (C), (E) the bovine fat standard and with (B), (D), (F) the rat fat standard.

286 PCA Loading PCA Loading PCA Loading PCA Loading PCA Loading PCA Loading 258 (A) (B) PC 1 (81.85 %) PC 2 (17.43 %) PC 1 (95.33 %) PC 2 (2.85 %) Wavenumber, cm Wavenumber, cm (C) (D) PC 3 (0.48 %) PC 4 (0.11 %) PC 3 (1.18 %) PC 4 (0.29%) Wavenumber, cm Wavenumber, cm (E) (F) PC 5 (0.05 %) PC 6 (0.03 %) PC 5 (0.16 %) PC 6 (0.08 %) Wavenumber, cm Wavenumber,cm Figure VI-13. PCA loadings for human skin sample #88 with 1-6 latent variables obtained with (A), (C), (E) the bovine fat standard and with (B), (D), (F) the human fat standard.

287 259 These contain intense sharp features around 4330 and 4250 cm -1. When the native fat standards are substituted, the lipid-related peaks in the first PCs decrease in magnitude so that they are no longer the dominant features. At the same time, the percent of the variance in the data set explained by these more balanced loadings extends to over 95 % for human skin and increases only slightly to % for the rat sample. The remaining PCs invariably contain some form of nonrandom features in the cm -1 region but the exact shape of these and the percent of the variance explained vary with the change of the fat standards. Incorporation of the spectral noise becomes noticeable in the fourth PC both for bovine fat and the rat fat residuals in Figure VI-12. However, this loading explains 1.56 % of variance in the residuals obtained with the bovine fat and more than 2 % for the native fat case. Results are similar for the human skin sample. The sixth PC, where some random noise becomes obvious, represents only 0.03 % of the variance for the bovine fat set and 0.08 % for the human fat set. The contribution of the random noise to the residuals is increased with the use of native fat for both species. Conclusions Nonrandom features are consistently present in residuals obtained after the fit of rat and human skin absorbance spectra with the six component linear regression model. When the spectrum of bovine fat is used as the fat term in the model, these residuals are heavily weighted towards the cm -1 region associated with lipid absorption features. Two new fat standards, rat abdominal fat and human subcutaneous fat, were collected and their spectra were substituted into the model. Improvements in the quality of the fit were observed: the new set of residuals appeared more balanced throughout the entire combination spectral region and SSRES values decrease overall for both rat and human skin samples. With the native fat, PCA demonstrates an increase in the contribution of random noise to the residuals.

288 260 CHAPTER VII FUTURE WORK Sensor repositioning error was observed during the noninvasive measurements of glucose performed with optical fibers through a fold of skin , 112, 205 Nonuniform spatial distribution of the skin components was proposed as a major source of this error. In this thesis, the effects of skin tissue lateral heterogeneity on noninvasive glucose prediction in rats and humans are explored ex vivo with NIR microspectroscopy. 129, 186 To appreciate differences and similarities in the spatial distributions of water, fat, proteins, and scattering properties between different species and sexes, near infrared microspectroscopic data were fitted with a six component linear regression model. For most samples, 98 % of the nonrandom spectral information was fitted by this simple model containing the major absorbers of NIR light. The resulting regression coefficients obtained from rat and human skin samples can be used to simulate a desired number of realistic skin spectra. An example of such a simulation and subsequent PLS regression for different locations across skin slices is detailed in Chapter V. The results of the simulation study have created many questions still to be tackled if a location-insensitive glucose prediction is to be realized. Directions for future work can be divided into three parts: tissue modeling and sampling size advancements, hardware and protocol improvements, and improvements in the multivariate calibration model. Presently, the skin tissue model incorporates only four major chemical components experimentally determined to provide a good fit in the combination region of 147, 184 the near infrared spectrum. Exact skin composition is much more complex. Candidates for future incorporation into the model include hyaluronis acid, elastin protein, melanins, glycoproteins, proteoglycans, and other types of collagen protein. 217 Even for the existing spectral standards, the samples used to obtain them may not be the best choices for rat and human skin. The advantages of using native fat instead of the original bovine fat are demonstrated in Chapter VI. A possible next step could be the

289 261 substitution of a better keratin standard. Human fingernail is reported to be composed primarily of keratin protein but may also contain smaller amounts of water and lipids. Moreover, keratin in nail (known as hard keratin) varies in its cystine content compared to soft keratin in skin. 108, 109, 150 The biological significance of multiple keratins is not yet known but the difference in disulfide bridge bonding may affect the NIR spectrum of this component. Alongside structural differences, molecular interactions can also affect the components absorptivity causing deviations from Beer s law. Ideally, precise knowledge of biological analyte/matrix interactions is needed to correctly interpret NIR spectra of mixtures. 40 The effect of hydration on the spectra of protein fibers in skin, for example, could be manifested in the absorption spectra. 218 Literature data suggest that hydrogen bonding between protein and water molecules may modify the spectrum of protein in the 47, 51, skin matrix relative to the spectrum of the same compound collected in its dry form. 108 Apart from water, in stratum corneum, proteins were demonstrated to be largely bound to lipids. 207 At least more than 1 % of lipid components are reported for some commercial collagen samples. 219 These details of the matrix composition have not yet been reflected in the regression model for skin. Providing access to an expanded database of properly characterized NIR spectra of biologically relevant compounds is an auxiliary goal of the study. Such a database will greatly benefit the scientific community. 70 Broadening the study to include skin from subjects with diabetes, common skin conditions, as well as, in general, collecting more human skin data is needed. 174, 175 It has not been trivial to obtain freshly harvested human skin samples from the back of a hand. If such samples were obtained, however, their heterogeneity data would be particularly valuable because it would reflect the composition of the site chosen for our noninvasive measurements on humans. Advancements in instrumentation include a brighter light source (currently, a conventional tungsten halogen lamp is used) for microspectroscopic measurements. A brighter source will increase the S/N ratio and provide the means to characterize thicker

290 262 samples. In other words, the skin tissue may not need to be squeezed so hard to achieve a measurable light throughput and the skin structural features will be less distorted. In terms of the in vivo experimental setup, the implication of the results included in this dissertation is that optical fiber movements should be minimized during measurements. A better holder for the sampling fibers or the sapphire rods is needed. Increase in the diameter of the fibers will help to average out the spatial differences in skin matrix as seen in comparison of human versus rat tissue. Currently, 3 mm sapphire rods are used in the interface to provide signal over a wider area of the skin tissue. With the distribution maps reported in this thesis, it is clear that the data obtained with a thicker rod should be less sensitive to the effects of tissue heterogeneity. The ultimate goal of the study is to improve the multivariate calibration component of glucose prediction to attain universal calibration. 220 Efforts have been made to correlate the change in standard error to a particular skin component. It was determined that there is no single constituent in the skin tissue the difference in which can explain the PLS model s degradation. More in depth analysis of the source of the heterogeneity-related prediction error is necessary.

291 263 APPENDIX The appendix contains illustrative figures not included in the main text. Figure A-1 provides the plots of relative change in the correlation coefficient for male rat skin sample #27 and female rat skin sample #80. Figures A-2 and A-3 include the Gaussian fits to the distributions of the six components for male and female rats, respectively. These distributions were used for the ANOVA described in Chapter III of this thesis. Figure A-4 presents the spectral map for the second sample of human dermis layer, D2. The spatial distributions of all six components are very similar to D1 sample that can be found in Chapter IV.

292 Coefficient Ratio Coefficient Ratio (A) water collagen I fat keratin constant slope Mapping Location Number (B) water collagen I fat keratin constant slope Mapping Location Number Figure A-1. Plots of ratios of regression coefficients for map (A) # 27 and (B) # 70.

293 Density Density Density Density Density Density 265 (A) (B) Rat 1 Rat 2 Rat 3 Rat 4 Rat 5 Rat 6 Rat 7 Rat Rat 1 Rat 2 Rat 3 Rat 4 Rat 5 Rat 6 Rat 7 Rat Regression coefficient for water Regression coefficient for collagen type I (C) (D) Rat 1 Rat 2 Rat 3 Rat 4 Rat 5 Rat 6 Rat 7 Rat Rat 1 Rat 2 Rat 3 Rat 4 Rat 5 Rat 6 Rat 7 Rat Regression coefficient for fat Regression coefficients for keratin (E) (F) Rat 1 Rat 2 Rat 3 Rat 4 Rat 5 Rat 6 Rat 7 Rat Rat 1 Rat 2 Rat 3 Rat 4 Rat 5 Rat 6 Rat 7 Rat Regression coefficient for constant Regression coefficient for slope Figure A-2. Gaussian distribution fits for A) water, B) collagen type I protein, C) fat, D) keratin protein, (E) constant, and (F) slope regression coefficients of male rats.

294 Density Density Density Density Density Density 266 (A) (B) Rat 1 Rat 2 Rat 3 Rat Rat 1 Rat 2 Rat 3 Rat Regression coeffficient for water Regression coefficient for collagen type I (C) (D) Rat 1 Rat 2 Rat 3 Rat Rat 1 Rat 2 Rat 3 Rat Regression coefficient for fat (E) Regression coefficient for keratin (F) 5 4 Rat 1 Rat 2 Rat 3 Rat Rat 1 Rat 2 Rat 3 Rat Regression coefficient for constant Regression coefficient for slope Figure A-3. Gaussian distribution fits for A) water, B) collagen type I protein, C) fat, D) keratin protein, (E) constant, and (F) slope regression coefficients of female rats.

295 267 (A) (B) (C) (D) (E) (F) Figure A-4. Spatial distribution maps for the sample of a single layer of dermis (map #D2), where regression coefficients for (A) water, (B) collagen type I protein, (C) fat, (D) keratin protein, (E) constant, and (F) slope are plotted across the mapped sample. Black contours denote 10% change in the regression coefficient. Two white circles represent the size of the optical fiber interface used for the in vivo glucose measurement. 38