Methods of TEOAE signal analysis


Traditionally the TEOAE responses are analyzed in the frequency domain via the popular Fast-Fourier Transform. The FFT decomposes the original TEOAE signal into a family of sinusoids ( the higher the frequencies of the TEOAE response, the larger the number of sinusoids in the output of the FFT).It should be noted that this sort of approach is ONLY an approximation of the true characteristics of the TEOAE signal, because we mainly assume that the auditory periphery is behaving in a linear manner. Nevertheless the FFT approach, pioneered by the ILO-xx family of devices, has been very popular the last 10 years and the majority of papers in the literature refer to FFT analyses.

      The FFT approach is very useful in an primary description of the TEOAE response which is needed in clinical protocols like neonatal screening. But when it is necessary to understand better the dynamics and the characteristics of the TEOAE response, the FFT approach is not appropriate. In the paragraph above we have mentioned that the FFT method works sufficiently, assuming that the auditory periphery is Linear. The reader might be confused at this point , given the names of the current TEOAE protocols which are known as Linear (L) or Non-Linear (NL) . It should be clarified that these names are not related in any way to non-linear methods. The NL protocol refers to collected TEOAE data where we assume that the differential clicks applied to the cochlea ( 3 positive and 1 negative ) generate an averaged TEOAE response which is dominated by non-linear components ( i.e responses from saturated cochlear generators). This assumption works well for adult subjects but not for well-babies or premature infants.


      Considering the evidence shown by a number of papers in the literature, suggesting that the TEOAE response contains not only TEOAE components, but distortion products and spontaneous emissions (see Brownell et al 1999; Kalluri and Shera 2001), this section is dedicated to new methods of TEOAE analysis which can provide new insights on the structure of TEOAEs and the relationship between the TEOAEs , DPOAEs and SOAEs. Although the signal processing algorithms are quite numerous, only a very small fraction of them has been applied to TEOAEs. The basic approaches are :

    1. Wavelet decomposition of TEOAEs(known also as Time-scale analysis). The TEOAE signal is decomposed not with sinusoids but with a function we call a wavelet. This approach refers to one of the most successful chapters of biomedical signal processing and it is very efficient if we would like to reproduced the TEOAE response but with significantly lower levels of noise (de-noising). For more information the reader can consult the following MEDLINE-enabled papers:
      1: Yao J, Zhang YT.
      Bionic wavelet transform: a new time-frequency method based on an auditory model.
      IEEE Trans Biomed Eng. 2001 Aug;48(8):856-63.

      2: Tognola G, Grandori F, Ravazzani P.
      Time-frequency analysis of neonatal click-evoked otoacoustic emissions.
      Scand Audiol Suppl. 2001;(52):135-7.

      3: Janusauskas A, Marozas V, Engdahl B, Hoffman HJ, Svensson O, Sornmo L.
      Otoacoustic emissions and improved pass/fail separation using wavelet analysis and time windowing.
      Med Biol Eng Comput. 2001 Jan;39(1):134-9.

      4: Morand N, Khalfa S, Ravazzani P, Tognola G, Grandori F, Durrant JD, Collet L, Veuillet E.
      Frequency and temporal analysis of contralateral acoustic stimulation on evoked otoacoustic emissions in humans.
      Hear Res. 2000 Jul;145(1-2):52-8.

      5: Tognola G, Grandori F, Avan P, Ravazzani P, Bonfils P.
      Frequency-specific information from click evoked otoacoustic emissions in noise-induced hearing loss.
      Audiology. 1999 Sep-Oct;38(5):243-50.

      6: Zheng L, Zhang YT, Yang FS, Ye DT.
      Synthesis and decomposition of transient-evoked otoacoustic emissions based on an active auditory model.
      IEEE Trans Biomed Eng. 1999 Sep;46(9):1098-106.

      7: Tognola G, Grandori F, Ravazzani P.
      Wavelet analysis of click-evoked otoacoustic emissions.
      IEEE Trans Biomed Eng. 1998 Jun;45(6):686-97.

      8: Blinowska KJ, Durka PJ.
      Introduction to wavelet analysis.
      Br J Audiol. 1997 Dec;31(6):449-59.

      9: Tognola G, Grandori F, Ravazzani P.
      Time-frequency distributions of click-evoked otoacoustic emissions.
      Hear Res. 1997 Apr;106(1-2):112-22.

      10: Rahko T, Kumpulainen P, Ihalainen H, Ojala E, Aumala O.
      A new analysis method for the evaluation of transient evoked otoacoustic emissions.
      Acta Otolaryngol Suppl. 1997;529:66-8.

      11: Wit HP, van Dijk P, Avan P.
      Wavelet analysis of real ear and synthesized click evoked otoacoustic emissions.
      Hear Res. 1994 Mar;73(2):141-7.

      12: Pasanen EG, Travis JD, Thornhill RJ.
      Wavelet-type analysis of transient-evoked otoacoustic emissions.
      Biomed Sci Instrum. 1994;30:75-80.

    2. Wigner-Ville-based decomposition of TEOAEs (known also as time-frequency analysis) and other adaptive kernels. In this approach the TEOAE signal is decomposed into a Time and Frequency matrix and interesting information can be gathered about the interaction of various TEOAE, DPOAE and SOAE components. The basics for the TF analysis are presented in the editorial of December 2001. For additional information the reader can consult :


1:  Hatzopoulos S, Cheng J, Grzanka A, Martini A.  
Time-frequency analyses of TEOAE recordings from normal and SNHL patients.
Audiology. 2000 Jan-Feb;39(1):1-12.

2:  Hatzopoulos S, Tsakanikos M, Grzanka A, Ratynska J, Martini A.  
Comparison of neonatal transient evoked otoacoustic emission responses recorded with linear and QuickScreen protocols.
Audiology. 2000 Mar-Apr;39(2):70-9.

3:  Hatzopoulos S, Cheng J, Grzanka A, Morlet T, Martini A.  
Optimization of TEOAE recording protocols: a linear protocol derived from parameters of a time-frequency analysis: a pilot study on neonatal subjects.
Scand Audiol. 2000;29(1):21-7.

  1. Higher-order kernels of the Volterra series: The TEOAE responses are analyzed to obtain the linear part and estimates of the slices of the 2nd and 3rd order Volterra kernels. Having this information one then can proceed in the estimation of other useful relationships such as the non-linear temporal interactions between the estimated kernels. For additional information the reader can consult :
    1: Thornton AR, Shin K, Gottesman E, Hine J.
    Temporal non-linearities of the cochlear amplifier revealed by maximum length sequence stimulation.
    Clin Neurophysiol. 2001 May;112(5):768-77.

    2: Thornton AR.
    Maximum length sequences and Volterra series in the analysis of transient evoked otoacoustic emissions.
    Br J Audiol. 1997 Dec;31(6):493-8.

  2. Recurrence Quantification Analysis (RQA):The RQA method aims to a direct and quantitative description of the amount of deterministic structure of the TEOAE response and it was shown to be an efficient and relatively simple tool in the non-linear analysis of many physiological signals. The basic idea behind RQA is the identification of recurrence of local data points in a reconstructed phase-space. The targeted system is analyzed by reconstructing the space of the true signal dynamics, using a coordinate system of surrogate variables, created by a combination of the measured signal and time-lagged copies of itself. For more information the reader can consult:
    1:  Zimatore G, Giuliani A, Parlapiano C, Crisanti G and Colosimo A .

    Revealing deterministic structures in click-evoked otoacoustic emissions
    J Appl Physiol 2000 Apr;88(4)1431-7.

  3. Classification Methods: This category refers to techniques which decompose the TEOAE or the DPOAE response into a set of unique parameters (i.e. correlation, response, noise level, S/N ratios at various ands etc). For the classification it is customary to use a training sets of data which in theory represent well the properties of the signals under classification ( i.e. from normal and hearing impaired ears). Once the training has been concluded another set of data (called the testing set) is classified. A positive aspect of these classification techniques is that every time the classification is correct, the classification system learns more and becomes more precise. Although this scenario might behelpfull for a few discrete categories, for example OAEs from normal ear and ears with otosclerosis, when the distributions of the categories under classification become overlapping the classification results are erroneous. For these cases it has been suggested to use very large training sets, a situation which is not very realistic in normal clinical set-ups.
            Two classification methodologies have been presented in the literature.

    • Spectral distriminant analysis: According to this method the classification is conducted not via parameters extracted from the time-waveform of the signal, but on data derived from the FFT of the TEOAE response. Since the FFT is very noise-sensitive the classification results are also noise dependent ( i.e. good TEOAE recordings are needed). The use of the FFT parameters provide an excellent way to resolve the issue of overlapping distributions, but the approach requires very large data sets. The user might find more information in the following papers.

      1:   Hatzopoulos S, Prosser S, Mazzoli M, Rosignoli M, Martini A.  
      Clinical applicability of transient evoked otoacoustic emissions: identification and classification of hearing loss.
      Audiol Neurootol. 1998 Nov-Dec;3(6):402-18.
      2:   Hatzopoulos S, Mazzoli M, Martini A.  
      Identification of hearing loss using TEOAE descriptors: theoretical foundations and preliminary results.
      Audiology. 1995 Sep-Oct;34(5):248-59.

    • Neural Networks: the developments in the area of NN are really impressive the last ten years and it issurprisingg that the method has not been used more extensively in the area of OAE characterization and classification. Usually the classification is conducted on time domain variables (TEOAEs) or S/N values of the DPOAEs, via the standard training and testing sets. More information can be found at :

      1:  Buller G, Lutman ME.
      Automatic classification of transiently evoked otoacoustic emissions using an artificial neural network.
      Br J Audiol. 1998 Aug;32(4):235-47.

      2: Kimberley BP, Kimberley BM, Roth L.
      A neural network approach to the prediction of pure tone thresholds with distortion product emissions.
      Ear Nose Throat J. 1994 Nov;73(11):812-3, 817-23.


Extra sources of information in the Portal


The reader might consult the following white papers and lectures