Wearable Respiratory Monitoring: Algorithms and System Design

Transcription

1 An IEEE Distinguished Lecture, 2 A/Prof. Wee Ser ewser@ntu.edu.sg School of EEE Nanyang Technological University (NTU) Singapore

2 An IEEE Distinguished Lecture, 2 Acknowledgement A/Prof. Wee Ser ewser@ntu.edu.sg IEEE Circuits and Systems Chapter, Santa Clara (Dr. Ping Chen) School of EEE Nanyang Technological University (NTU) Singapore for inviting me to give this IEEE DL Talk

3 The IEEE IEEE Canada IEEE-USA sections worldwide

4 IEEE Circuits and Systems Society IEEE CASS Membership (38 Societies in IEEE) > 4, 8,53 Number of Chapters, CASS is the 9 th largest IEEE Society (after CS, ComSoc, MTTS, PES, SPS, SSCS, EDS, and IA) CASS is a technical society of IEEE

5 Nanyang Technological University (NTU) 955: Established as the st Chinesemedium university outside China 98: Nanyang Technological Institute set up (English) with 3 Engineering Schools 99: NTU inaugurated, incorporating National Institute of Education Now (2): 3rd NTI/NTU President College of Engineering Nanyang Business School College of Science College of Humanities, Arts and Social Sciences School of Medicine Student population > 3,

6 x 4 Outline of Presentation I. Introduction A Clinical Perspective II. III. Motivations and Problem Formulation Wearable Wheeze Monitoring NTU a) Overall System Design b) Wearable Sensors and Platform Suite c) Wearable Signal Processing Algorithms IV. Lab Tests, Pre-Clinical Trials and Results V. Conclusions and Future Work 3

7 x 4 Acknowledgement Research Nanyang Technological University A/Prof. W Ser, Dr. J Zhang, J Yu, Dr. T Zhang Research National University Hospital A/Prof. Dr. Daniel YT Goh and Team 2

8 x 4 I Introduction - A Clinical Perspective 4

9 x 4 Respiratory Disorders Chronic respiratory diseases include disorders that affect any part of the respiratory system, not only the lung but also the upper airway (nose, mouth, pharynx, larynx, and trachea), chest wall and diaphragm, as well as the neuromuscular system that provides the power for breathing 5

10 x 4 Our Lungs and Airways Major features of lungs: bronchi, bronchioles and alveoli OSA disorder is characterized by repetitive collapse of the upper airway 6

11 x 4 Examples of Respiratory Disorders Chronic Obstructive Pulmonary Disease (COPD) Irritation of the lungs, can lead to asthma, emphysema, chronic bronchitis, or a mix of these Asthma Periodic constriction of bronchi & bronchioles more difficult to breathe Chronic Bronchitis Irritant in bronchi and bronchioles - stimulates secretion of mucus air passages clogged persistent cough 7

12 x 4 Basic Types of Respiratory Sounds Normal Breath Sounds Normal vesicular sounds ( Hz) Normal tracheal sounds (from < to >3 Hz) 8 Adventitious Sounds Continuous sounds (e.g. Wheezes) Discontinuous sounds (e.g. Crackles) Upper Airway Sounds Snore 8

13 Wearable Respiratory Monitoring: x 4 Normal vs Adventitious Lung Sounds x 5 Short duration frequcy Normal Breath Sound time Long duration x x 4 frequcy Crackles time 9 frequcy 5 Wheezes time

14 x 4 Wheezes Continuous, sinusoidal-like, musical sound Most obvious during exhaling (breathing out) but may occur during inhaling too One of the most common noises in respiratory diseases (indicative of disorders) Causes: Asthma, Chronic Bronchitis, COPD (Chronic Obstructive Pulmonary Disease), Allergy, foreign object, forced expiration, etc.

15 x 4 Waveform and Spectrum of Wheezes.8.6 Waveform.4 frequency amplitude frequency x 4 Spectrum frequcy Time-Frequency time

16 x 4 Definition of Wheeze Wheeze ATS Signal Duration > 25ms Fundamental Frequency > 4Hz CORSA Signal Duration >ms Fundamental Frequency > Hz 2

17 x 4 II Motivations & Problem Formulation 3

18 x 4 Why Monitor Respiratory Disorders? Common and important cause of illness and death Can affect people of American all ages, both Lung genders Association: About 335, Americans die of lung 3 million people suffer disease from every Asthma year and 2 million people with COPD (27), world wide Lung disease is No. 3 killer in America, responsible for in 7 deaths in 7 in UK affected by chronic lung disease Lung disease and other breathing Responsible for over % problems of hospitalizations are No. killer of babies & over 6% of deaths in Canada younger (including than year lung oldcancer) > 3 million Americans are living with Chronic Obstructive Pulmonary Disease Source: University of Maryland Medical Center web site 4

19 x 4 Market / Commercialization Potential U$ (billions) 6 Source: GlobalData Report

20 x 4 Wheezing is a common noise in respiratory disorders Prevalence of Childhood Asthma: * % - 2% (varies with countries) Adult Asthma: * 2% Adult COPD: * 5% All ages USA 23 million Asthmatics Over 5 million cases of COPD 5

21 x 4 Wheeze Monitoring Values Accurate detection and assessment of wheezes prompt diagnosis & appropriate treatment of Asthma, COPD, etc, Improved symptom (wheeze) control reduced hospitalization healthcare savings How (Current Clinical Methods) Auscultation using a stethoscope Based on history - patient recall or asthma diary or audio or audio-visual recording by patients 7

22 x 4 Problem of Wheeze Monitoring Next! Auscultation with stethoscope is subjective, not for long-duration monitoring History of occurrence is important but patients descriptions are often subjective with limited accuracy Wheezes often occur at night; Intermittent wheezes are not easily monitored at night 8

23 x 4 The System Needed A screening system that records and analyzes wheezes over an extended duration of time A simple and robust system suitable for both clinical and home use A system that allows the patient to use at anywhere and anytime - hence no skin contact, no external wirings that restrict movements Non-Skin Contact Wearable System 9

24 x 4 III Wearable Wheeze Monitoring Overall System Design NTU 2

25 x 4 Goal Study and develop a wearable solution for the detection, recording, and analysis of wheezes, anywhere and over a long duration, for the purpose of wheeze conditions monitoring (including treatment response) 22

26 Overview of our Solutions x 4 Wearable wheeze detection and recording system with Signal Analysis Tools for physicians Wearable System Signal Analysis Tools PI: A/Prof. W Ser (NTU) and A/Prof. Daniel YT Goh (NUH) Funded by A*STAR & MOE Signal Analysis Tools 23

27 x 4 Our Wearable System Design Respiratory sounds (lung, tracheal) are recorded by sound sensors (stethoscopes or air-conductance microphones) Sensor outputs are band-passed, amplified, sampled, and segmented. A low-complexity algorithm is applied to detect presence of wheezes Segments with wheezes are recorded & their severity estimated; respiratory & heart rates are estimated too 24

28 x 4 Our Signal Analysis Tools (SAT) Respiratory Signal General Analysis Signal Manipulation Processing Data (wheezing sounds) recorded in wearable suite are further processed Respiratory by Signal the SAT parameters Energy Signal Analysis copy, cut, paste Amplitude Power & periodicity Spectral Signal Analysis of amplification wheeze, Wheeze length in a breathing The SAT has several Pitch functions Analysis Labeling cycle, etc. Physiological parameters Format Analysis Filtering Signal manipulation Zero-Crossing Segmentation Analysis General signal inspiration processing or expiration, etc.), changes in Classification Types of wheeze (monophonic or polyphonic, severity after treatment, Wheeze location, etc. Respiratory signal analysis The SAT can also be used for training purposes 25

29 x 4 Problems and Challenges I Sound based design sensitive to external audio noise How sensors are placed on wearable suite important Need filter and noise reduction algorithms Wearable power and packaging constraint Need to operate at low sampling rates Need low complexity processing algorithms Need low-power platform (processor) Only small dry batteries allowed No external wiring 26

30 x 4 Problems and Challenges II Unsupervised wearing and monitoring Need to design for ease of wearing Need to use multiple sensors Cannot have air-gap between sensors and body Different sizes of wearers Need flexible length of suite fastening Need a few sizes 27

31 x 4 III Wearable Wheeze Monitoring Wearable Sensors & Platform Suite NTU 28

32 Our Inward- & Outward- Looking Sensing Designs x 4 Medical Knowledge Outward-looking Microphone Array Wearable Suite + Sound Sensors Inward-looking Distributed Sensors 29

33 Wearable Wheeze Detection System Version #: Microphone Array, Sampling (K, 2-bit) x 4 DSP based System Notebook for Playing Sounds Speaker LED Indicators TMS32C673 DSP board Experimental Set Up 4-channels Microphone Array (amplifier and filter included) 3

34 Wearable Wheeze Detection System (Version #2: PDA based, 2-microphone design) x 4 PDA based System NI CF64 DAQ 5-pin Compact Flash connector Microphones (head phone) Signal conditioning Wheeze and Snore Indicators PDA (HP ipaq hx27 series) Waveform display 3

35 Acquisition System (Version #3: Embedded Processor based System) Wearable Wheeze x 4 Signal Conditioning NI Daq Signal conditioning Font View of Wearable Suite 4-channel modified stethoscopes or microphones Notebook for experimental control 32

36 x 4 Sensors and Placements of Sensors Inward-looking system Stethoscope is modified by replacing long acoustic tube with a microphone and wire Microphone based system uses special acoustic coupler design 4 distributed sensors: tracheal, left lung, heart, and right lung Outward-looking system 4-channel microphone array design End-fire array configuration 33

37 x 4 Bandpass Filter, ADC, Platform The bandpass filter is designed to have a bandwidth of up to KHz (without DC) The ADC has a sampling rate of 2 KHz and a resolution of bits The platform uses one MSP43 chip for each channel (for 4 channels) The selected sounds are stored in a micro SD card 34

38 Data Processing & Recording x 4 Recording Cycle ADC Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) ADC 2 Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) ADC 3 Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Used for Processing (Wheeze Detection) Used for Processing (Wheeze Detection) ADC 4 Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) Data Block (3 sec) 35

39 x 4 III Wearable Wheeze Monitoring Wearable Signal Processing Algorithms NTU 36

40 x 4 Types of Wheeze Detection Algorithms Time-domain methods (waveform analysis, Kurtosis, zero-crossing rate ) Frequency-domain spectral analysis methods Time-frequency analysis (peak detection) Pattern recognition approach etc. 37

41 x 4 Spectral Analysis Methods f to f 2 : Wheeze frequencies P ave : Average power in f to f 2 A f : Scaling factor (= 5) Wheeze Detection Criteria: > f and > A f x P ave Results: (5 Asthmatic and 2 normal teenagers): < 2% for false +ve and false -ve T. R. Fenton, H. Pasterkamp, A. Tal and V. Chernick, Automated Spectral Characterization of Wheezing in Asthmatic Children, IEEE Trans. on Biomedical Engineering, vol. BME-32, No., pp5-55,

42 x 4 Time-Frequency Analysis Methods Search for peaks in successive spectrum & decide using threshold relating to amplitude, frequency, duration and number of wheezes Peaks Extraction Results (37 asthmatic, 23 normal) at high, medium and low air flow rates Time-Frequency Plot A. Homs-Corbera, R. Jane etc., Time-frequency detection and analysis of wheezes during forced exhalation, IEEE Trans. on Biomedical Eng., vol. 5, No., pp82-86,

43 x 4 Pattern Recognition Methods Typically 3 stages: Feature extraction Dimensionality reduction Pattern classification. Commonly used Features: MFCC, LPC, AR model, Wavelet coefficients, etc. Dimensionality Reduction Algorithms: SVD, PCA, etc. Classification Methods: k-nn Classifier, DTW, VQ, NN, HMM, etc. 4

44 x 4 Pattern Recognition Methods (Example) Use Wavelet Packet Decomposition for feature extraction Use Learning Vector Quantization for classification Results { Sensitivity = TP / (TP + FN); Positive predictivity = TP / (TP + FP) } L. Pesu, P. Helistö, E. Ademovic, J.-C. Pesquet, A. Saarinen and A.R.A. Sovijärvi, Classification of respiratory sounds based on wavelet packet decomposition and learning vector quantization, Technology and Health Care, vol. 6, pp , Jun998. 4

45 x 4 Problems of Existing Algorithms Not accurate enough or computationally too expensive for wearable solutions Our Approach Review wheeze signal characteristics and attempt to derive simpler but effective methods that over-detect/record 42

46 x 4 Re-visiting Wheeze Signals Amplitude amplitude Amplitude Frequency (Hz) Normal Time Breath (s) Time (s) Wheeze Time (s) Time (s) Energy concentrates at well below Hz Contrasting energy distribution across frequency Typically, wheeze occurs only in part of a breathing cycle 43 frequency (Hz) Frequency (Hz)

47 x 4 Our Wheeze Detection Algorithm Design goal: or 2 parameters, low complexity, robust, and suitable for wearable systems Received signals: st () = wt () + nt () w(t): Wheeze signal n(t): Non-wheeze signals (e.g. breath sound, noise) 44

48 x 4 Our Wheeze Detection Algorithm (Cont d) Short-Time Fourier Transform + * 2 (, ) () ( ) j π ft Sτ f = sth t τ e dt h(t) is the window function, τ is the time shift, and * is the complex conjugate operator. Peak Detection by Masking: For each STFT, determine moving thresholds and use them to select peaks (dominant values) S p ( n, S a ( n, f ) = S( n, f ) S f ) = a L ( n, L n= f ) S( n, 45 if if f ) S( n, S( n, f ) S f ) S a a ( n, ( n, f f ) ) >

49 x 4 Our Wheeze Detection Algorithm (Cont d) Entropy based Feature Extraction: Assuming there are N dominant components, C, C 2,, C N, in S p (n, f), calculate cn pn =, n=, 2,, N N c n= Entropy can be calculated as Smoothed Entropy (by averaging filtering) Features can be extracted as, for example, E = max( E ) / min( E E n ratio diff = max( E 46 N E p log ( p ) = a a n= ) min( E n b n E a a = M a E t M t= ) )

50 x 4 Our Wheeze Detection Algorithm (Cont d) Wheeze detection: Set T ratio and T diff as the thresholds for E ratio and E diff Presence of wheeze can be decided by (for example) Wheeze is present No Wheeze Unable to determine if if E ratio > T ratio and E diff > Tdiff E ratio < T ratio and E diff < Tdiff Otherwise 47

51 x 4 Our Wheeze Severity Estimation Algorithm Values of E ratio or E diff are computed once every 3 seconds Severity Score = Number of positive detections in minute 48

52 x 4 Our Respiratory Rate Estimation Algorithm Rate estimated from second peak of autocorrelation 49

53 x 4 IV Lab Tests, Pre-Clinical Trials, Results 5

54 x 4 Results of Wheeze Detection (Downloaded Data) Sensitivity 2 8 Entropy Difference Detection Accuracy (%) true Positive false Negative true Negative false Positive Specificity normal wheeze Entropy Ratio SNR (db) At SNR = 2dB, 8% of Sensitivity & Specificity have been achieved (9 wheezy and 6 normal ) Sensitivity = TP/(TP+FN) x %; Specificity = TN/(TN+FP) x % 5

55 x 4 Results of Wheeze Detection (Pre-Clinical Trial) Data Collection at National University Hospital, Singapore More than 8 have been collected Performance: Sensitivity > 85%; Specificity > 7% 52

56 x 4 Results of Severity Estimation Algorithm (Pre-Trial) Samples Severity Scores Doctor s Assessment No Wheeze 2 Low 3 35 Yes 4 5 Low 5 Low 6 2 Low 7 5 Low 8 95 Yes 53

57 x 4 Results of Respiratory Rate Estimation (Pre-Trial) Estimated RR (bpm) Actual RR (bpm) 54

58 x 4 V Conclusions and Future Work 55

59 x 4 Conclusions & Future Work Wearable sound based sensing (+ PC signal analysis) can be effective for automatic wheeze detection Sensitivity higher than 85% and specificity above 7% have been achieved using our algorithm/system for data collected from some real patients (more data are needed to validate the results) Future work includes extensive clinical trials and studies on other types of respiratory disorders 56

60 x 4 Some of our Research Publications W Ser, JM Zhang, JF Yu, and TT Zhang, A Wearable Multiple Audio-Sensor based Design for Respiration Monitoring, patent filed 2. W Ser, ZL Yu, JM Zhang, and JF Yu, A Wearable System Design with Wheeze Signal Detection, 5 th International Workshop on Wearable and Implantable Body Sensor Networks, 28. W Ser, TT Zhang, JF Yu, and JM Zhang, Detection of Wheeze by Distributed Microphone Array Analysis of Breathe and Lung Sounds, The 6 th International Workshop on Wearable and Implantable Body Sensor Networks, 29. JM Zhang, W Ser, JF Yu, and TT Zhang, A Novel Wheeze Detection Method for Wearable Monitoring Systems, The International Symposium on Intelligent Ubiquitous Computing and Education, 29. TT Zhang, W Ser, Daniel YT Goh, JM Zhang, JF Yu, C Chua, IM Louis, Sound Based Heart Rate Monitoring for Wearable Systems, The 7 th International Workshop on Wearable and Implantable Body Sensor Networks, 2. Daniel YT Goh, W Ser, JM Zhang, JF Yu, C Chua, IM Louis, TT Zhang, Development of a Wearable System to Record and Analyse Respiratory Sounds: Preliminary Data from Wheeze Analysis, The 26 th International Paediatric Association Congress, 2. 57

61 x 4 Some other References T. R. Fenton, H. Pasterkamp, A. Tal and V. Chernick, Automated Spectral Characterization of Wheezing in Asthmatic Children, IEEE Trans. on Biomedical Engineering, vol. BME-32, No., pp5-55, 985. Y. Shabtai-Musih, J.B. Grotberg and N. Gavriely, Spectral content of forced expiratory wheezes during air, He, and SF 6 breathing in normal humans, J. Appl. Physiol., vol. 72, pp , 992. R. Jané, D. Salvatella, J. A. Fiz, and J. Morera, Spectral analysis of respiratory sounds to assess bronchodilator effect in asthmatic patients, Proc. of the 2 th IEEE EMBS, Hong Kong, 998. M. Waris, P. Helistö, S. Haltsonen, A. Saarinen2, A.R.A. Sovijärvi, A new method for automatic wheeze detection, Technology and Health Care, vol. 6, pp. 33-4, Jun 998. L. Pesu, P. Helistö, E. Ademovic, J.-C. Pesquet, A. Saarinen and A.R.A. Sovijärvi, Classification of respiratory sounds based on wavelet packet decomposition and learning vector quantization, Technology and Health Care, vol. 6, pp , Jun998. A. Homs-Corbera, R. Jane etc., Algorithm for time frequency detection and analysis of wheezes, IEEE EMBS, pp , New York, USA, Sept

62 x 4 Some other References (Cont d) S. Lehrer, Understanding Lung Sounds, third edition., W.B. Saunders Company, Pennsylvania, USA, 22. M. Bahoura and C. Pelletier, New parameters for respiratory sound classification, Proc. IEEE CCECE, pp , Montreal, Canada, May 23. Y. P. Kahya, E. Bayatli ect., Comparison of different feature sets for respiratory sound classifiers, Proc. of the 25 th IEEE EMBS, pp , Cancun, Mexico, Sept. 23. A. Kandaswamy, C. Sathish Kumar, Rin. PL Ramanathan, S. Jayaraman, and N. Malmurugan, Neural classification of lung sounds using wavelet coefficients, Computers in Biology and Medicine, vol. 34, pp , 24. Patents and product information of KarmelSonic Products LVQ-PAK Software, Department of Information and Computer Science, Helsinki University of Technology, Finland. School of Medicine, University of California at Davis, USA, Review of Lung Sounds. [Online] 59

63 Thank You Questions or Comments? 6