Spectral estimation of EEG signals using cascaded inverse filters.

While most of the studies on application of autoregressive (AR) methods to EEG signals have considered direct modelling of EEG data, this paper considers the inverse problem of passing the EEG signal through an inverse filter and shows how such inverse filters when cascaded give an improved spectral estimate of the input data. It is shown how a proper choice of model orders of such cascaded inverse filters leads to better spectral estimation of an EEG signal than by conventional AR filters. An EEG signal, when first passed through a low order inverse filter, actually results in a signal with reduced dynamic range and thus a second inverse filter with higher order gives much better spectral peaks. In fact, such cascading operation reduces the problem of ill conditioning of the autocorrelation matrix thus yielding better results. The analysis has been performed using real EEG data.