Comparison of basic linear filters in extracting auditory Evoked Potentials

Creative Commons License

Aydin S.

TURKISH JOURNAL OF ELECTRICAL ENGINEERING AND COMPUTER SCIENCES, vol.16, pp.111-123, 2008 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 16
  • Publication Date: 2008
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED)
  • Page Numbers: pp.111-123
  • Keywords: adaptive filtering, Wiener filtering, auditory evoked potential, EEG, STEP-SIZE
  • Hacettepe University Affiliated: No


In the present study, the performances of two well-known linear filtering techniques are compared for extraction of auditory Evoked Potential (EP) from a relatively small number of sweeps. Both experimental and simulated data are filtered by the two algorithms into two groups. Group A consists of Wiener filtering (WF) applications, where conventional WF and Coherence Weighted WF (CWWF)) have been assessed in combination with the Subspace Method (SM). Group B consists of the well-known adaptive filtering algorithms Least Mean Square (LMS), Recursive Least Square (RLS), and one-step Kalman filtering (KF). Both groups are tested with respect to signal-to-noise ratio (SNR) enhancement by comparing to the traditional ensemble averaging (EA). We observed that KF is the best method among them. The application of the SM before filtering improves the performance of the LMS and the assessments of WF where the CWWF works better than the conventional WF in that case.

Conclusion and Discussion: The results show that, the SM can remove a large amount of EEG noise. In addition, the characteristic of the EEG noise remaining on the projections renders white noise. Thus, the SMCWWF was found to be better than the SMWF for all data sets. Thus CWWF is the better alternative to the WF in spite of its higher computational complexity. In the Group B, the RLS and KF were better compared to EA. The RLS is the best filter in simulations, whereas the KF provides the highest performance in experimental studies. When we analyze the KF after 128 sweeps, it shows characteristic of a low-pass filter, which has a narrower bandwidth, compared to the RLS filter, indicating the possibility of better performance. The LMS filter performance depends on 1) the number of sweeps, 2) the step size parameter, 3) the filter length, and 4) the input SNR of single sweeps. It was, in general, found unsuccessful for low input SNR cases. The selection of step size parameter was assumed to be the crucial factor in the performance. To obtain a better performance with the LMS filtering, various methods were proposed which explore an optimum stepsize at each iteration [8]. In another study, the optimum value was determined methodologically considering the filter length, input signal variance and the desired signal [2]. In the present study, these approaches are not attempted; instead the performance of the SMLMS algorithm is tested. The SMLMS algorithm appeared to be relatively less sensitive to the step size and showed better performance compared to the EA in both experimental and simulation trials. However, its performance in the pseudo-simulations proves again unsatisfactory. In conclusion, we leave the following matters for future work: 1) the use of optimal step size, and, 2) exploring further properties of the SMLMS algorithm for better performance. In conclusion, most of the basic linear estimation techniques show definitely better performance than EA in extracting the EPs. The KF effectively reduce the experimental time (to one-fourth of that required by EA). The SM proves to be a useful pre-EEG filter.