基于脑电信号预发作数据段选取的癫痫发作预测

A novel algorithm for seizure prediction based on patient specific manner was proposed to improve the accuracy of epileptic prediction, including feature extraction, pre-ictal period selection, feature selection and channel selection. Time-frequency features and spatial features were extracted from each channel by 2 s windows with 1 s overlap. A continuous 10 min data was selected as a valid positive sample of the pre-ictal period from segment before seizure onset, which achieved the maximum linear separability compared with the inter-ictal period. The effective features were selected by elastic net, then the selected effective features were used to select effective channels in greedy manner. The effective features of effective channels were input into classifier. The algorithm was tested on the scalp electroencephalogram (sEEG) from the MIT Physio database and the database collected in Xuanwu Hospital. The algorithm achieved a recall of 95.76% and a false positive rate of 0.1073 h⁻¹ in MIT database, and a recall of 97.80% and a false positive rate of 0.0453 h⁻¹ in Xuanwu Hospital database. Results show that the algorithm has high sensitivity and low false positive rate.