FAST ACQUISITION OF OTOACOUSTIC EMISSIONS BY MEANS OF PRINCIPAL COMPONENT ANALYSIS

Transcription

1 FAST ACQUISITION OF OTOACOUSTIC EMISSIONS BY MEANS OF PRINCIPAL COMPONENT ANALYSIS P. Ravazzani 1, G. Tognola 1, M. Parazzini 1,2, F. Grandori 1 1 Centro di Ingegneria Bioedica CNR, Milan, Italy 2 Dipartiento di Bioingegneria, Polytechnic of Milan, Italy Abstract - Transient-Evoked Otoacoustic Eissions (TEOAE) are nonstationary acoustic signals coing fro the inner ear after acoustic stiulation by clicks and they are applied as tools in newborn hearing screening progras to allow the early identification of hearing loss and the consequent diagnosis and intervention. In any screening progra, the duration of each test is a crucial paraeter. For TEOAE, it is strongly influenced by the acquisition procedure, typically based on classical synchronous averaging technique over 260 sweeps, with an average acquisition tie of about 2-3 inutes. This paper present an application of Principal Coponent Analysis (PCA) to rapidly-acquired TEOAE (averaged over only 10, 60 or 100 sweeps) for the detection of this type of cochlear response. The PCA approach is shown to be able to enhance the signal-tonoise ratio (SNR) and, in turn, to allow a correct detection of the responses. Results of the application of this approach in coparison with responses recorded, fro the sae ears, with classical technique will be shown. The reduction of the acquisition tie to about one forth with respect to its typical value and with approxiately the sae final signal-to-noise ratio will be discussed. Keywords - Principal Coponent Analysis; Fast Acquisition; Otoacoustic Eissions I. INTRODUCTION Transient-Evoked Otoacoustic Eissions (TEOAE) are nonstationary acoustic signals coing fro the cochlea in response to short transient stiuli, such as clicks and tonebursts [1]. These responses can be recorded fro the ear canal of all noral adults, children and neonates, and are increasingly used as a clinical test to assess the integrity of the peripheral organ [2]. In particular in newborns, TEOAE can be recorded fro all noral ears shortly after birth, and, for this reason, newborn hearing screening protocols based on the easure of TEOAE are now largely ipleented [3]. TEOAE are known to be characterized by an extreely high inter-subject variability and intra-subject repeatability. In particular, the inter-subject variability akes the identification of the true eission a crucial task when otoacoustic eissions are utilized in hearing screening progras. Their identification is coonly perfored with the help of soe siple statistical easure such as the correlation between replicate trials, and, partially, by visual inspection [4]. The rapid diffusion of TEOAE-based hearing screening progras has increased the iportance of the search of any ethod capable of increasing the objectivity and sensitivity of eission identification. In particular, the huan eye ore readily detects the presence or absence of an otoacoustic eission when it siultaneously sees ore than one response collected at different stiulus levels, coparing soe utual features within a set of traces. In clinical practice, this procedure has not yet found widespread acceptance due to the long acquisition tie needed for the synchronous averaging technique. This paper ais to present a ethod, based on a Principal Coponent Analysis (PCA) approach, to increase the signalto-noise ratio (SNR) in a set of click-evoked otoacoustic eissions recorded averaging only a few sweeps. The ain goal is that of reducing the acquisition tie with respect to the classical procedure typically based on synchronous averaging of a larger nuber of sweeps. The PCA Approach II. METHODOLOGY Principal coponent analysis (PCA) involves a atheatical procedure that transfors a nuber of (possibly) correlated variables into a (saller) nuber of uncorrelated variables called principal coponents. The first principal coponent accounts for as uch of the variability in the data as possible, and each succeeding coponent accounts for as uch of the reaining variability as possible. PCA is a well-established tool in the interpretation of analytical data. It depicts separate and close observations fro a sall-uncorrelated variable set. When the data is constituted by a set of wavefors, the analysis of the relations between the wavefors in the set could be perfored to deterine soe coon characteristics. To this purpose, one should describe the with a iniu nuber of paraeters. One typical approach is based on the linear representation of the set. The general for of a linear representation of a set of N data wavefors s (t) where n ranges fro 1 to N, is given by: n M s ( t) = c f ( t) (1) n = 1 where f (t) is the th -basic wavefor and c n is the weighting coefficient, which represents the contribution of the th -basic wavefor to the n th -data wavefor. Each datawavefor can be coputed by the knowledge of the M basic wavefors and M weighting coefficients. The M basic wavefors with the weighting coefficients can be estiated using the PCA approach [5]. Let us consider the atrices S [N rows x T coluns] of the s(t) data wavefors, C [N x M] of the weighting coefficients and F [M x T] of the basic wavefors. Thus, relation (1) can be written as:

2 Report Docuentation Page Report Date 25 Oct 2001 Report Type N/A Dates Covered (fro... to) - Title and Subtitle Fast Acquisition of Otoacoustic Eissions by Means of Principal Coponent Analysis Contract Nuber Grant Nuber Progra Eleent Nuber Author(s) Project Nuber Task Nuber Work Unit Nuber Perforing Organization Nae(s) and Address(es) Centro di Ingegneria Bioedica CNR Milan, Italy Sponsoring/Monitoring Agency Nae(s) and Address(es) US Ary Research, Developent & Standardization Group (UK) PSC 803 Box 15 FPO AE Perforing Organization Report Nuber Sponsor/Monitor s Acrony(s) Sponsor/Monitor s Report Nuber(s) Distribution/Availability Stateent Approved for public release, distribution unliited Suppleentary Notes Papers fro 23rd Annual International Conference of the IEEE Engineering in Medicine and Biology Society, October 25-28, 2001, held in Istanbul, Turkey. See also ADM for entire conference on cd-ro., The original docuent contains color iages. Abstract Subject Ters Report Classification Classification of Abstract Classification of this page Liitation of Abstract UU Nuber of Pages 4

3 S = CF (2) The basic wavefors F are typically constrained to be an orthonoral set. Considering now the correlation atrix R coputed on the data wavefors S, the corresponding eigenvalue diagonal atrix, ë, and the associate eigenvector atrix U, one could easily deonstrate [5] that: C = U λ (3) F = 1 U ' S λ (4) where U' is the transpose of atrix U. Now one should consider that a coonly used easure of the size of a signal wavefor is its power P n. Considering Eq. (1) and using the orthogonality relation for the basic wavefors, one should find: 2 P n c nt = M = 1 A easure of the iportance of a basic wavefor is the power it contributes to the total power of the entire set of wavefors. This contribution, considering also Eq. (3), can be denoted by P N N 2 cnt = n= 1 n= 1 = λ u = λ The purpose of developing the theory of linear expansions is that it perits finding expansions that hopefully will perfectly reconstruct S utilizing significantly fewer basic wavefors F. Fro Eq. (6) one can easily infer that the eigenvalue ë represents the contribution of the th basic wavefor to the total power in the set of data wavefors. It then becoes obvious that the ost iportant basic wavefors to use in representing the experiental wavefors are those corresponding to the largest eigenvalues. Thus the eigenvalues can be used to infer the diensionality of the signal space. The PCA Approach For Fast Acquisition of TEOAE The PCA approach described above is applied here to the fast acquisition of a set of TEOAE recorded fro the sae ear, during the sae test session at different stiulus levels. On a practical ground a set of otoacoustic eissions, which were acquired at different stiulus intensity levels is considered as atrix S. Each set was coposed by 11 responses fro the sae ear at decreasing stiulus levels (see below) and copiled into a 2D-array resulting in a atrix of 11 rows and 512 coluns (see the exaples in Fig. 1). Responses to different stiulus intensities were arranged in rows and the post-stiulus tie changes in coluns. Following Eq. (1), the set of otoacoustic eissions S can be considered as the result of the product between a atrix F of n (5) (6) basic wavefors and a atrix C of weighted coefficients. Hence, considering, as stated before, that the highest eigenvalues are responsible of the ost of the power of the responses and, on the contrary, the lowest eigenvalues can be related to noise, one could set at zero the latter ones, sieving the basic wavefors responsible for the actual otoacoustic response fro other coponents due to noise. A set of signals can be reconstructed, constituted by the original set of eissions S without the contribution of noise coponents, hence increasing the SNR. In order to evaluate the perforances of the PCA approach as a whole and identify the optial setting of the paraeters, the Siilitude was used, defined as the value of the correlation function at zero-lag between a reference set (in the following Golden Standard (GS)) and each rapidlyacquired set, before and after PCA approach. The reference set coprises the responses recorded fro the sae ear by eans of the classical averaging procedure (average over 260 sweeps). Siilitude was coputed in a tie window fro 6 to 18 s. TEOAE Response Acquisition TEOAE were collected fro 15 noral ears of 9 adults with an ILO88 apparatus (Otodynaics Ltd.), with standard adult ILO probes. The recordings were processed on-line with the ILO88 default procedure (saple frequency saples/s, resulting in 512 sapled points, windowing fro 2.5 to 20 s). TEOAE were recorded using the derived nonlinear ode for acoustic artifact reduction [6]. Clicks were used as stiuli with intensity ranging fro 53 to 83 db SPL, step of 3 db, for a total nuber of 11 recordings per ear. For each response, two replicate recordings (A and B) were collected at the sae recording conditions. For each test session, four types of responses were recorded: one reference set (Golden Standard GS), consisting of an average of 260 sweeps (typical in clinical practice) and three different rapidly-acquired set of responses, averaged over 10, 60 and 100 sweeps, respectively. All responses were collected during one single recording session, and if possible, without reoving the probe fro the ear canal, to aintain alost the sae recording conditions. The applied protocol could be suarized as follows: 1. Recording of the four set of 11 TEAOE responses fro the sae ear at decreasing stiulus levels. 2. Coputation of the correlation atrix of S. The correlation atrix was coputed by the cross covariance between the n th and the k th wavefors, in order to reflect differences in signal intensity. 3. Coputation of the corresponding eigenvalue diagonal atrix, ë, and the associate eigenvector atrix U. 4. Zeroing all the eigenvalues but the highest two. 5. Coputation of C and F by Eq. (3) and Eq. (4), to obtain, estiation of S. 6. Coparison of with the GS set and easuring the perforances of the PCA approach by the Siilitude.

4 III. RESULTS Fig. 1 shows an exaple (Ear #9 - left) of a rapidly acquired set of TEOAE responses (averaged over 60 sweeps) to stiuli at intensities fro 53 to 83 db SPL before (on the left) and after (on the right) the application of the PCA approach. The iniu diension of each set, i.e., the iniu nuber of TEOAE responses that should be collected in order to obtain a satisfactory application of PCA, is a crucial paraeter for the practical application of the procedure. Also in this case, the Siilitude between the GS and the rapidly acquired set of signals was observed. PCA was applied to three reduced set of three, four and five responses (fro 71 to 83 db SPL) collected fro each ear. The average Siilitude is shown in Fig. 3. In any case, the axiu difference of the Siilitude is of 6 PP whereas, at the highest stiuli, the Siilitude reaches approxiately the 90% and there is practically no difference in applying PCA over a set of 3, 4 or 5 signals Siilitude [%] Fig. 1. An exaple of a set of TEOAE (Ear #9, responses averaged over 60 sweeps) before (on the left) and after (on the right) PCA processing. The responses were recorded t different stiulus intensities, fro 53 (botto trace) to 83 db SPL (top trace). For each trace, A and B replicates are shown superiposed. Fig. 2 shows the average Siilitude for all 15 recordings over stiulus intensity coputed for all three rapidlyacquired set of recordings (10, 60 and 100 sweeps). The average Siilitude for the 60-sweep and 100 sweep set ranges fro 49% at the lowest to 90% at the highest stiulus levels, whereas the 10 sweep set, with a very low signal-to-noise ratio before PCA processing, shows figures ranging fro 5% to 53% at the lowest and highest stiulus intensities respectively. Siilitude [%] Stiulus Intensity [db SPL] 10 Sweeps 60 Sweeps 100 Sweeps Fig. 2. Siilitude as a function of the stiulus intensity of the rapidly acquired set of responses averaged over 10, 60 and 100 sweeps with respect to the GS set. The figures are relative to the average over all ears Stiulus Intensity [db SPL] Fig. 3. Siilitude as a function of the stiulus intensity of the rapidlyacquired set of responses averaged over 60 sweeps with respect to the GS set changing the nuber of TEOAE responses in the set. The figures are relative to the average over all ears. IV. DISCUSSION AND CONCLUSION 5 Signals 4 Signals 3 Signals The reduction of the acquisition tie of TEAOE is crucial for the use of otoacoustic eission both in clinical practice and as tool for newborn hearing screening progras (see, e.g., [7]). This paper presents an innovative ethod to collect a set of otoacoustic responses with a substantial reduction in acquisition tie. The ethod is based on copiling a rapidly acquired set of responses (averaged over a nuber of sweeps lower than the classical 260, typical in clinical practice) and applying a PCA approach in order to increase the SNR. As an exaple, one can record in about the sae acquisition tie either one otoacoustic response classically recorded averaging 260 sweeps or four responses at different stiulus intensity levels averaged over only 60 sweeps. To provide a benchark of the PCA approach perforances, Siilitude was introduced. The coparison between these etrics estiated in all 15 ensebles of recordings before and after PCA processing, with respect to the Golden Standard provided evidence that the use of PCA produces no loss of inforation in the set of data and that the coparison between the wavefors of the PCA processed

5 and those of the GS confired that signal orphology and latency were fairly well preserved. However, the effectiveness of this procedure strongly depends on the SNR of the original rapidly acquired set of data. Averaging only 10 sweeps usually yielded a very low SNR (Fig. 2), with no appreciable increase in Siilitude after PCA processing. Therefore, it is advisable to apply the described approach only to set of data with sufficiently high SNR, which for the type of signals addressed by this study is typically achieved with an average of at least 60 sweeps. Finally, the results of this study suggest that the rapidacquisition of only three responses at high stiulus levels could be sufficient to obtain satisfactory perforances fro the PCA approach. In that sense, the PCA approach allows the collection of three responses at different stiulus levels in less than the tie of the classical "single shot" (i.e., only one response at stiulus level of 80 db SPL), used in typical TEOAE-based newborn hearing screening progra. ACKNOWLEDGMENT This work was partially coordinated within the fraework of the European Theatic Network AHEAD II (Fifth Fraework Prograe 5FP, Quality of Life, Contract Nuber QLG , European Coission, ). REFERENCES [1] D.T. Kep, Stiulated acoustic eissions fro within the huan auditory syste, J. Acoust. Soc. A., vol. 64, pp , [2] R. Probst, B.L. Lonsbury-Martin, and G.K. Martin, A review of otoacoustic eissions, J. Acoust. Soc. A., vol. 89, pp , [3] B.R. Vohr, A.B. Maxon, "Screening infants for hearing ipairent," The Journal of Pediatrics; vol. 128, pp , [4] M.E. Lutan, Reliable identification of click-evoked otoacoustic eissions using signal-processing techniques, Br. J. Audiol., vol. 27, pp , [5] E.M. Glaser and D.S. Ruchkin, "Principal of neurobiological signal analysis," New York, Acadey Press, [6] P. Bray, and D.T. Kep, An advanced cochlear echo suitable for infant screening, Br. J. Audiol., vol. 21, pp , [7] P. Ravazzani, G. Tognola, F. Grandori, J. Ruohonen, "A two-diensional filter to facilitate detection of transientevoked otoacoustic eissions," IEEE Trans. Bioed. Eng., vol. 45, pp , 1998.