Methods of topographical time-frequency analysis of EEG in coarse and fine time scale

Two complementary methods of EEG analysis are discussed. The first of them operates on a time scale of several (or more) seconds and describes the covariance structure of the signals. The second offers time-frequency resolution close to the theoretical limit and makes possible accurate description of EEG transients. Both methods were applied in analysis of overnight sleep EEG recorded from 21 derivations in 8 healthy subjects. The first method was based on a vector autoregressive AR model. All channels of the AR process were evaluated simultaneously (not pair-wise), and ordinary, multiple, and partial coherencies and Directed Transfer Functions were calculated for all combinations of channels, sharp decrease of partial coherencies with distance was found, in agreement with intracortical studies. Increases of coherence were observed for sleep stages 2, 3 and 4. Use of the Directed Transfer Function made possible identification of the main centers from which EEG activity is propagated during sleep and wakefulness. The Matching Pursuit (MP) algorithm was applied to identification of sleep spindles. MP is a recently introduced adaptive time-frequency method of signal analysis. An iterative algorithm fits the local signal structures with waveforms from a large and redundant dictionary. Sleep spindles were described in terms of natural parameters, i.e. position in time and frequency, width in time, and amplitude, with accuracy close to the theoretical limit. Comparison of automatic detection with visual analysis showed high concordance which decreased with threshold amplitude. Time, frequency and spatial characteristics of spindles were also evaluated.