Adaptive smoothing: an improved method for spectral analysis and its application to seizure EEG.

We describe the data adaptive smoothing method, an improved method for multichannel spectral analysis of seizure EEG. After Fast Fourier Transform of EEG data, spectra were computed by smoothing over adjacent frequency components. Using cross-validatory maximum likelihood criteria, unsmoothed spectral data were used to select the level of smoothing (spectral window effective bandwidth) required to minimize bias and variance errors. The statistical assumptions of this method are consistent with the statistical properties of seizure EEG. On computer simulation of seizure EEG, the smoothing level predicted by this method correlates strongly with the optimum smoothing level. The utility of the method is demonstrated by application to seizure EEG. The consistency of the method's statistical assumptions, the success in selection of the optimum smoothing level, and the variability in optimum smoothing required for seizure EEG suggest that the adaptive smoothing method is a useful method for multichannel spectral analysis.