The clinical Link. Distortion Product Otoacoustic Emission

Transcription

1 Distortion product otoacoustic emissions: Introduction Michael P. Gorga, Ph.D. Boys Town National Research Hospital Work supported by the NIH Collaborators on BTNRH OAE Projects Stephen Neely Kathy Beauchaine Darcia Dierking Tricia Dorn Cassie Garner Brenda Hoover Debra Hussain Jan Kaminski Walt Jesteadt Tiffany Johnson Doug Keefe Dawn Konrad-Martin Danielle Montoya Jo Peters Beth Prieve Joelle Redner Laura Schulte Brenda Starnes Alicia Schmiedt Lisa Stover Principles behind the use of OAEs to identify hearing loss normal cochlea behaves nonlinearly source of nonlinearity is the OHC system OHCs are physiologically vulnerable OAEs are byproducts of normal nonlinear function loss of OAEs indicates damage to the OHCs 1

2 The clinical Link OHC damage is closely linked to hearing loss, at least for losses up to about db loss of OAEs indicates OHC damage, which in turn, indicates the presence of hearing loss Stimuli for eliciting DPOAEs two primary tones, f 1 and f 2 f 2 = 1.22(f 1 ) (same as f 2 /f 1 = 1.22) primary levels may range from 0 to 85 db SPL moderate primary levels are in most common clinical use primary levels may be equal or L 1 may be greater than L 2 primary-level differences should increase as overall level decreases Distortion Product Otoacoustic Emission acoustic response measured using microphone in the sealed ear canal evoked using two-tone stimulation (f 1 and f 2 ; f 1 < f 2 ) response is due to intermodulation distortion in the basilarmembrane response many DPs are generated, but response typically is measured at the frequency equal to 2f 1 -f 2 2

3 Measuring DPOAEs many distortion products are produced but response typically is measured at a frequency equal to 2f 1 -f 2 because it is the largest one in mammals noise typically is estimated as the average level in several frequency bins above and below the 2f 1 -f 2 bin. DPOAE-to-noise ratio (SNR) is estimated as the db difference between energy at 2f 1 -f 2 and energy in adjacent noise bins DPOAE Measurements DPGrams - plots of DPOAE level as a function of either f 2 frequency or the geometric mean frequency, which is defined as the square root of (f 2 x f 1 ), while primary levels are held constant. DP Input/Output (I/O) functions - plots of DPOAE level as a function of primary level while primary frequencies are held constant. 3

4 Measurement-Based Stopping Rules stop test if noise level is reduced to some criterion level, related to the level at which system distortion occurs stop test if SNR exceeds some criteria stop test if response level exceeds some criteria stop test after some amount of test time, even if measurement-based criteria are not met. Why Measurement-Based Stopping Rules? Increases test efficiency - don t waste time measurements are terminated if high SNR is observed, meaning a response is present measurements are terminated if low noise is achieved, meaning that, if a response was present, it should have already been measured measurements are terminated after some time limit, preventing the test from continuing indefinitely. 4

5 Distortion product otoacoustic emissions in relation to hearing loss Data From Subjects with Hearing Loss DPgrams tend to mimic the audiogram. That is, there is reduced output (for fixed primary levels) at those f 2 frequencies at which hearing loss exists I/O functions also reflect the pattern of hearing loss. That is, there is less output over most of the range of levels for frequencies at which hearing is impaired. Note: high level responses may be more normal, depending on magnitude of loss 5

6 from Kemp et al. (1986), Scand. Audiol. 25 (Suppl.), Next Slide Kemp et al. (1986) Top panel: Data from normal ears Middle panel: DPgrams from three ears with hearing loss Bottom panel: Audiograms from the same three ears with hearing loss Note: Frequency is represented on a linear scale 6

7 from Martin et al. (1990), Ann. Otol. Rhinol.Laryngol. 99 (Suppl. 147), Next 3 data slides Martin et al. (1990) Top row: Audiograms (separate panels for each ear) and DPgrams (different symbols for each ear) Bottom row: DP I/O functions (different symbols for each ear) Note: The frequency scale in the audiogram goes from to 8 khz, whereas the DPgram goes from 1 to 10 khz. Both scales are logarithmic 7

8 Summary of Case Studies DPgrams follow the audiogram, showing responses for frequencies where hearing is normal and reduced or absent responses when hearing loss exists. I/O functions also are consistent with audiometric status. 8

9 Case Studies versus Group Results Case studies are informative, but what is needed are data from large samples of normal and hearing-impaired subjects Large sample studies allow for quantitative analyses from which it will be possible to describe how DPOAE tests (or any test, for that matter) perform under clinical conditions Factors affecting DPOAE test performance f 2 frequency (Gorga et al., 1993, 180 subjects) primary levels (Stover et al., 1996, 210 subjects) magnitude of hearing loss (Gorga et al., 1997, 806 subjects) Effects of f 2 Frequency From Gorga et al. (1993), JASA 93,

10 ROC Curves Hit rate as a function of false alarm rate Comparisons are made when f 2 = audiometric frequency Parameter is f 2 frequency In these coordinates, chance performance occurs when ROC curve has a slope of one (line falls on positive diagonal). Perfect performance occurs when hit rate = 100% and false alarm rate = 0%. Frequency Effects Performance at 500 Hz is near chance Performance improves as f 2 increases Best performance observed at 4 khz Slight drop in performance when f 2 = 8 khz. 10

11 Reasons for Frequency Effects noise decreases as frequency increases forward (and perhaps reverse) middle ear energy transmission is more efficient for mid to high frequencies compared to lower frequencies cochlear distortion maybe greater for higher frequencies compared to low frequencies Low-Frequency Limits of DPOAE Measurements primary source of noise in DPOAE measurements is breathing and movement noise, which are dominated by low-frequency energy noise increases as frequency decreases, thus making it harder to reliably measure a response at lower frequencies noise problems are compounded by measuring responses at 2f 1 -f 2 The Problem of Measuring Responses at 2f 1 -f 2 DPOAEs are generated at the f 2 place cochlear status at the f 2 place is being predicted from DPOAE measurements predictions are based on measurements at 2f 1 -f 2, which is about 1/2 octave below f 2, where noise levels are higher, making measurements less reliable 11

12 Stimulus and Response Representations Top: Stimuli and response. Note that 2f 1 - f 2 is below f 1 Bottom: Idealized representation of stimuli and response along the basilar membrane. Interactions Between Frequency of Interest (f 2 ) and the Frequency of Measurement (2f 1 -f 2 ) noise levels reach asymptotic low levels for f 2 s at 4 khz. Thus, it matters little that 2f 1 -f 2, the frequency of measurement, is about 1/2 octave below f 2, the frequency of interest interaction has a significant negative impact for f 2 s at and below 4 khz, because 2f 1 -f 2 falls in frequency regions for which noise levels increase 12

13 Performance at 8 khz While noise is the major source of reduced test performance for lower frequencies, noise levels are very low at 8 khz Thus, poorer performance at 8 khz cannot be due to noise Poorer performance may relate to the need to drive the loudspeakers with a greater voltage in order to get target primary levels. This high voltage may result in an increase in system distortion, resulting in poorer performance Primary Level Effects Should we use primary levels that result in the largest responses from subjects with normal hearing - that would be the highest possible level Should we use primary levels close to threshold - that would require several measurements under conditions of poor SNR Should we use primary levels that result in the greatest separation between the distributions of responses from normal and impaired ears Effects of Primary Level on DPOAE Test Performance From Stover et al. (1996), JASA 100, Next 6 data slides 13

14 I/O Functions to DPgrams I/O functions plot DPOAE amplitude vs. L 2 Obtain DPOAE I/O functions for several f 2 frequencies From each I/O function, note the DPOAE amplitude when L 2 is held constant DPgrans are plots of these amplitudes as a function of frequency 14

15 DPgrams and Audiograms in Normal ears Audiogram bottom row, middle panel DPgrams, constructed from I/O functions, for five different L 2 levels High-level bottom left panel Low-level bottom right panel Moderate levels top row of panels Results in Normal Ears Normal ear produces response for most frequencies and most L 2 levels Low-level stimuli sometimes do not produce responses in ears with normal hearing. 15

16 I/O functions in an ear with hearing loss Noise levels caused variability in I/O functions at lower f 2 frequencies Normal appearing I/O functions at 1, 1.4, and 2 khz Abnormal I/O functions at 2.8, 2, and 4 khz Probably normal I/O function at 8 khz DPgrams and Audiograms in an Ear with Hearing Loss Audiogram - bottom row, middle panel Abnormal response at high levels, but only for 4 khz (lower left panel) Abrnormal responses for moderate level primaries at frequencies for which hearing loss exists (top row of panels) Abnormal response at 8 khz for low-level stimuli, even though hearing was normal at this frequency 16

17 Test Performance Vs. L 2 Area under the ROC curve is an estimate of test performance Test performance is defined as the test s ability to correctly identify both normal and impaired ears Areas close to 0.5 Represent chance performance Areas close to 1.0 represent perfect performance Each function represents data for a different frequency 17

18 Effects of Primary Level At all f 2 frequencies, test performance increases as L 2 increases An maximum asymptotic value is achieved at moderate L 2 levels Performance decreases slightly at higher L 2 levels Reasons for Primary Level Effects normal ears may not produce a response at low levels - drives up the false positive rate impaired ears may produce a response at high levels - drives up the false negative rate moderate levels should minimize errors Laboratory vs. Clinical Observations Test performance is determined most by f 2 frequency. Do the effects of f 2 frequency, observed under laboratory conditions, hold when data are obtained under routine clinical conditions Will the optimal primary levels under laboratory conditions work well in the clinic 18

19 Large Scale Study in the Clinic 1257 ears of patients seen through the audiology clinic subjects ages covered the life span all subjects had normal tympanograms on the day of the DPOAE test audiograms were available for all subjects 65/55 db SPL primaries were used, given the results of the primary level studies Effects of f 2 Frequency under Clinical Conditions From Gorga et al. (1997), Ear & Hearing 18, ROC Curves Plots of hit rate vs. false alarm rate Triangles: SNR was used Circles: DPOAE level was used Each panel shows data for a different f 2 ROC curve areas are given inside each panel 19

20 Clinical Observations results under clinical conditions are virtually identical to those seen in the laboratory test performance is best for frequencies of 4 and 6 khz test performance is poorer as f 2 decreases and for 8 khz Summary DPOAEs accurately identify auditory status at mid and high frequencies DPOAEs are less accurate for frequencies <1.5 khz, due primarily to noise levels DPOAEs are more accurate for moderate level primaries DPOAEs will miss some ears with mild hearing loss DPOAEs will incorrectly label some ears with normal hearing as impaired 20

21 Establishing DPOAE Criteria for Use in the Clinic Gorga et al. (1996) JASA 100, Gorga et al., (1997) E&H 18, Shortcomings of Common Approaches for Selecting DPOAE Test Criteria Data from normal-hearing subjects can be used to estimate ONLY the false-positive rate (a problem not only for SNR) a priori SNR criteria, such as 3, 6 or 9 db SNRs, do NOT result in a 100% hit rate or a 0% false-alarm rate Equal SNRs do not = the Same Auditory Status DPOAE = 5 db SPL at 4 khz Noise = -4 db SPL at 4 khz SNR = 9 db, cochlea probably normal DPOAE = -15 db SPL at 4 khz Noise = -24 db SPL at 4 khz SNR = 9 db, cochlea probably abnormal SNR alone would have misdiagnosed the second case as normal hearing 21

22 Another Problem: Response Distributions Overlap There are no DPOAE criteria that perfectly separate the responses observed in normal ears from those seen in impaired ears. Errors are inevitable. The question is, which error is more important, falsepositive or false-negative errors. Which Error is More Important Prevalence may be important in choosing which error to control If the incidence of hearing loss is low in the target population, one might want to minimize the false-positive errors If the incidence is high in the target population, then one might want to minimize the false-negative errors. Cumulative Distributions (CD) CDs are the proportion of time responses occur that are some criterion value as a function of that value CDs are another way of plotting a normal distribution. The same data that result in a bell-shaped normal distribution result in a sigmoidal (S) shaped CD. CDs are just a different way of looking at the same data CDs of responses from both normal and impaired ears are needed whenever the distributions of normal and impaired responses overlap 22

23 Possible DPOAE criteria that can be used in the clinic DPOAE Level DPOAE SNR DPOAE Threshold Multivariate Scores Other, as yet, undetermined DPOAE measure, such as latency or slope of the I/O function Developing an Approach for Use in the Clinic (Gorga et al., 1996, 1997) We need CDs of the measurement variable of interest (i.e., DPOAE level or SNR). CDs from normal-hearing ears can be used to determine criteria associated with specific falsepositive rates CDs from impaired ears can be used to determine criteria associated with specific hit rates Selecting a Hit Rate from CDs One approach for selecting a criterion DPOAE value is to first decide on an acceptable hit rate or test sensitivity Draw a horizontal line at that hit rate until it intersects the distribution of responses from impaired ears Drop a vertical line to the X-axis to find the criterion value that provides that hit rate 23

24 Selecting a False-Alarm Rate Decide on an acceptable false-positive rate Draw a horizontal line at the false-positive rate until it intersects the distribution of responses from normal ears Draw a vertical line from this intersection to the X axis to find the criterion DPOAE value associated with that false-alarm rate Cumulative Distributions in Normal and Impaired Ears Note: As examples, criteria are selected that resulted in a 95% hit rate (impaired CD) and a 5% false alarm rate (normal CD) 24

25 Figure showing DPOAE Levels for Fixed False Alarm and Hit Rates DPOAE level as a function of f 2 Data from normal (left panel) and impaired (right panel) ears are shown Parameter is percentage, from 5 th to 95 th percentiles Filled symbols represent the DPOAE levels at the median (50 th ) percentile Evidence of Overlap Between Normal and Impaired Responses No criterion can be selected for which the hit rate is 100% AND the false alarm rate is 0%. All of the functions in the right panel are not below all of the functions in the left panel. Some impaired ears produce bigger responses (or lower thresholds) than some normal ears Or, stated differently, some normal ears produce smaller responses (or higher thresholds) than some impaired ears 25

26 Constructing a Useful Clinical Form (Gorga et al.,1997, 2002) Select 1 or 2 hit rates (say 90 th & 95 th %) and find the DPOAE criteria associated with them Select 1 or 2 false-alarm rates (say, 5 th & 10 th %) and find the DPOAE criteria associated with them Plot these values (DPOAE criteria vs. f 2 ) to create a form for clinical use Basis for Developing a Clinical Form Right: Schematic showing overlap between DPOAE levels from normal and impaired ears Left: top line = 95 th percentile (hit rate = 95%) from impaired distribution, second line = 90 th percentile from impaired distribution (hit rate = 90%), third line = 10 th percentile from normal distribution (false-alarm rate = 10%), bottom line = 5 th percentile from normal distribution (false-alarm rate = 5%) 26

27 Interpreting DPOAE Data, Using the Left Panel of the Previous Figure Responses above highest values from impaired CDs would be consistent with normal hearing because few impaired ears produced responses this big or bigger Responses below lowest values from normal CDs would be consistent with hearing loss because few normal ears produced responses this small or smaller Diagnosis is uncertain for responses between these values (i.e., in shaded regions), where normal and impaired responses overlap Large Sample Study figure and data from Gorga et al. (1997, E&H) and Gorga et al. (2002) in Robinette and Glattke, 2 nd Ed. Data from 1257 normal and impaired ears L 1 /L 2 = 65/55 db SPL All data collected under clinical conditions Measurement-based stopping rules were used BTNRH Clinical Form NOTE: Baljit Rehal, Au.D. was instrumental in developing this form 27

28 Caveats when using this form SNR should be 6 db IF noise floor is not below the lower limit of graph in order to use this form to help interpret responses Large DPOAE levels cannot be plotted on this form if DPOAE noise floor (i.e., SNR 0 db); such responses are uninterpretable because they could be just noise If noise is reduced below lower limits of graph AND response is not above the noise floor (SNR=0 db), the results are interpretable (i.e., consistent with hearing loss) because the reason no response was observed was because the response was so small, not because the noise was too high Diagnosis is uncertain for responses in shaded area (SNR 6 db but responses from normal and impaired ears overlap). Interpreting Responses in the Shaded Areas There should be a minimum SNR of 6 db Between 90 th (impaired distributions and 5 th (normal distributions) percentiles, responses fall in the region of uncertainty because there is overlap in responses from normal and impaired ears 28

29 Five Case Studies, in which Clinical Form was Used to Assist in Interpretation First step in interpreting results is to determine if DPOAE level was reliably measured. SNR helps to determine that. A related issue is to determine whether the noise floor was sufficiently reduced. Case #1 Frequency DPOAE Noise Signal/Noise

30 Case 1: Results Consistent with Normal Hearing Low noise levels even for lower f 2 s Large DPOAEs Positive SNR at all f 2 s Levels above 90 th percentile for impaired ears Results consistent with normal hearing because few impaired ears produce such large responses Case #2 30

31 Frequency DPOAE Noise Signal/Noise Case #2: High Noise Levels = Uninterpretable Responses Large DPOAEs High noise levels Low SNR Results are uninterpretable because large DPOAEs may be nothing more than noise Note that the levels were the same as for Case #1 31

32 Case #3 Frequency DPOAE Noise Signal/Noise

33 Case #3: Low SNR & Low Noise Levels can be Interpreted DPOAEs below the lower limits of graph Noise levels also are low Low SNR (i.e., DPOAE level was not measured reliably above the noise floor) Results are consistent with hearing loss because the reason a response was not measured was NOT due to high levels of noise, but to low level of response. Case #4 Frequency DPOAE Noise Signal/Noise

34 Case #4: DPOAEs in the region of uncertainty DPOAE levels in shaded region Noise levels well below DPOAE levels Positive SNR, meaning DPOAEs were measured reliably Results cannot be assigned to normal or impaired distribution Responses between 90 th (Impaired) and 5 th (Normal) Percentiles (Gorga et al., 2002) This is the region where overlap occurs between the responses produced by normal and impaired ears. One might conclude that these responses are completely uninterpretable If SNR 6 db, we can exploit the relation between audiometric threshold and DPOAE level to help interpret the response 34

35 Audiometric Threshold vs. DPOAE Level Audiometric threshold (db HL) as a function of DPOAE level (db SPL) Data for a different f 2 shown in each panel Solid line = 50 th percentile (median) Shaded areas = interquartile range (25 th to 75 th percentile) Audiometric Threshold vs. DPOAE Level Audiometric threshold decreases as DPOAE level increases Although variable (note that the shaded area only covers the middle 50% of the distribution at each audiometric threshold), relation exists for audiometric thresholds from -5 to about db HL No relation above 55 db HL, related to the range of levels over which outer hair cells operate 35

36 Caveats For This Form The SNR should be 6 db The shaded area covers only the 25 th to the 75 th percentile Thus, there are still cases when interpretations will be wrong However, using this form provides limited but still useful information for interpretation, especially when no other data are available to assign a response to either normal or impaired groups Interpreting Data in the Region of Uncertainty DPOAE levels reliably measured All DPOAEs are in region of uncertainty; thus, they cannot be assigned to normal or impaired distributions Given the measured levels, one would anticipate that these responses were coming from an ear with either normal hearing or mild hearing loss (see previous figure). It would be less likely that these responses were coming from an ear with moderate, severe or profound hearing loss. Case #5 36

37 Frequency DPOAE Noise Signal/Noise Case #5: Uninterpretable DPOAEs in Region of Uncertainty DPOAEs in shaded region Noise levels = DPOAE level (SNR 0) DPOAEs, therefore, are not reliable Results cannot be interpreted (cannot use either form) because measured responses may be just noise, but this cannot be known 37

38 Univariate vs. Multivariate Analyses Univariate approach compares data from one measurement (say DPOAE & noise level at one f 2 ) to predict cochlear status at the same frequency. Multivariate analyses use data from many variables (in our case, DPOAE level and noise for many f 2 s) to predict cochlear status at a single frequency. What Do Multivariate Analyses Do? Use many variables as inputs to generate a single dimensionless number select variables and coefficients that result in the greatest separation between the means of two distributions of the new dimensionless variable Select variables and coefficients that minimize the variance of each distribution One Kind of Multivariate Analysis Discriminant Analysis 38

39 Developing Multivariate Solutions Because many variables are being used, the solutions can be irregular and may not generalize to a new set of data Multivariate solutions should be validated on an independent set of data Validating Multivariate Analyses (Dorn et al., 1999) Obtained large data set randomly divided data set in half use the first half to develop (train) the algorithm do not change the training algorithm apply training algorithm to the other half of the data set 39

40 Relative Operating Characteristic Curve Areas (AROC) AROCs can be used to describe test performance for dichotomous clinical decisions (i.e., normal versus impaired hearing) AROCs range from 0.5 (chance performance) to 1.0 (perfect performance) No audiological tests have AROCs equaling 1.0. Univariate vs. Multivariate Results Multivariate solutions provide better test performance than univariate analyses greatest differences between univariate and multivariate approaches occurs in the lower frequencies Multivariate solutions do nearly as well on validation set as they did on training set. Thus, the solutions appear to be robust Validation from studies by other investigators would be useful 40

41 Problems with Validation in Original Work Both ears of many subjects were tested Ears were randomly assigned to either training or validation sets The two ears of the same subject are not independent A completely independent data set are needed to evaluate the generalizability of the multivariate solutions A Better Validation of Multivariate Analyses (Gorga et al., 2005) An entirely new set of subjects: 345 ears of 187 subjects, 2 to 86 years of age No middle-ear dysfunction Pure-tone audiograms measured for each subject DPOAE Stimuli: f 2 = 0.75 to 8 khz, half-octave steps f 2 /f 1 = 1.22 L 1 = 65 db SPL; L 2 = 55 db SPL Previously described multivariate solution, without modification, were applied to these new data. Clinical decision theory (AROCs) used to assess test performance 41

42 SNR & DPOAE Level Versus Multivariate Solution AROCs are larger for SNR compared to DPOAE level at lowest frequency, while AROCs are larger for DPOAE level at most mid and high frequencies AROCs for multivariate analyses exceed those for both DPOAE level and SNR, with the exception of 4 khz, a frequency for which areas are about the same. Conclusions Previously described multivariate solutions were robust and generalized to an entirely new set of data. Improved DPOAE test performance can be achieved by using the multivariate solutions. The improved test performance does not require any additional test time. Comparing Case Studies Pure tone audiograms DP-Level form (univariate analysis) P(N) form (multivariate analysis) NOTE: Probability of normal (P(N)) is derived from the dimensionless number generated by the multivariate analysis 42

43 Normal hearing subjects with high P(N) Hearing impaired subjects with low P(N) Hearing Impaired Subjects with Flat and Sloping Losses 43

44 Test performance is not perfect: Example of a false negative Test performance is not perfect: Example of a False Positive Summary DPOAEs do not perform perfectly Test performance can be improved by using multivariate analyses Multivariate solutions are robust and appear to generalize Even with these improvements, perfect performance is never achieved 44

45 Bad News - Good News The bad news is that errors in diagnoses are inevitable when DPOAEs are used to identify hearing loss. This is true for other tests, not just DPOAE tests. The good news is that, when auditory status is uncertain, it is more likely that we are confusing normal or mild hearing losses or mild and moderate hearing losses. It is much less likely that we are confusing normal hearing with moderate or greater losses. 45