A seizure prediction method based on efficient features and BLDA

The automatic seizure prediction technology is crucial to develop a new therapy for the patients suffering from medically intractable epilepsy. This paper proposes an efficient and low-complexity method for seizure prediction. The univariate feature of line length can describe both amplitude and frequency variations of an EEG signal, and the bivariate feature of mean phase coherence (MPC) is able to quantify phase synchronization between EEG signals recorded from two different channels. The two features are combined to characterize the behavior of EEG signals. The Bayesian linear discriminant analysis (BLDA) algorithm is used as classifier to decide the feature samples. The smoothing and threshold process for the outputs of the BLDA classifier can reduce incorrect decisions further. This method has been evaluated on the publicly available EEG dataset and achieved the acceptable sensitivity of 86.36% with the false prediction rate of 0.04/h. The low computational burden of this method makes it suitable for the real-time processing in clinical practice.