Parametric modeling of somatosensory evoked potentials.

In this paper, we examine methods of characterizing somatosensory evoked potentials (SEP's) in both the time and frequency domains. We have found that the truncated impulse response (TIR) method produced an accurate time domain model of the SEP signals at model orders greatly reduced from the original state space matrix. The TIR method was valuable for smoothing signals that were slightly corrupted by noise. In this case, the simulated data sequence was close to the original data sequence in the mean squared error sense. For signals that were greatly corrupted by noise, the TIR method was not able to perform as well. Therefore, the TIR method was not a feature extraction method but was valuable for data simulation. In the frequency domain, we have used the autoregressive moving average model (ARMA) to parameterize the SEP signal. An overdetermined set of Yule-Walker equations was created to determine the autoregressive (AR) parameters of the original data with the model order established by the singular value decomposition. From these AR parameters, a residual time series was generated which was used to find the moving average parameters. The resulting ARMA model was used to produce a simulated data sequence. The frequency domain characteristics of the simulated sequence and the corresponding power spectral density of the ARMA filter were very close to the periodogram of the original data sequence. Accurate parameterization was achieved for the SEP waveforms at low filter lengths.