DETECTION OF SEIZURES FROM SMALL SAMPLES USING NONLINEAR DYNAMIC SYSTEM-THEORY

The electroencephalogram (EEG), like many other biological phenomena, is quite likely governed by nonlinear dynamics. Certain characteristic s of the underlying dynamics have recently been quantified by computin g the correlation dimensions (D-2) of EEG time series data. In this pa per, D-2 of the unbiased autocovariance function of the scalp EEG data was used to detect electrographic seizure activity. Digital EEG data were acquired at a sampling rate of 200 Hz per channel and organized i n continuous frames (duration 2.56 s, 512 data points). To increase th e reliability of D-2 computations with short duration data, raw EEG da ta were initially simplified using unbiased autocovariance analysis to highlight the periodic activity that is present during seizures. The D-2 computation was then performed from the unbiased autocovariance fu nction of each channel using the Grassberger-Procaccia method with The iler's box-assisted correlation algorithm. Even with short duration da ta, this preprocessing proved to be computationally robust and display ed no significant sensitivity to implementation details such as the ch oices of embedding dimension and box size. The system successfully ide ntified various types of seizures in clinical studies.