Seizure Prediction in EEG Records Based on Spatial-Frequency Features and Preictal Period Selection

Algorithms can automatically predict seizures to reduce the occurrences of accidental injury and improve living conditions of patients. This paper proposes a novel patient-specific algorithm based on multi-channel scalp EEG recordings. 26 features for each channel are extracted from each one-second data, including 8 absolute spectral powers, 8 normalized spectral powers, 8 power spectral entropies, the shortest path length and clustering coefficient. Then, a new step to select the most discriminative five minute preictal period is proposed. The features of preictal period are combined with that of five minute non-seizure period to form a training set in order to achieve the maximum linear separability criteria. Then, the effective features of each channel are selected by Elastic Net. At the same time, greedy algorithm is used to select effective channels. The ten minute effective features obtained from effective channels are input to Logistic Regression. The algorithm is tested on 62 seizures from 12 patients in 217 hours of recordings in MIT database. Results are finally given by average of each 1 minute values of Logistic Regression. It is shown that the proposed algorithm can achieve a sensitivity of 91% and an averaged false positive rate of 0.3 per hour.