A High Cross-Individual Accuracy EEG-Based Seizure Detection Algorithm Based on Multiple Source Domain Adaption

Electroencephalography (EEG) has been proven to be very effective in seizure detection. However, individual variability has severely limited its practical use. Due to differences in brain structure and skin, EEG signals can vary greatly from one individual to another, which leads to a low cross-individual seizure detection accuracy. To solve the problem, two methods are proposed in this work. Firstly, by treating cross-individual tasks as transfer learning scenarios, template matching based on multiple source domain adaption neural network method is proposed. The method selectively eliminate differences between multiple source domains and target domain to improve the accuracy. Secondly, considering the fact that seizure data is much less than non-seizure data, adaptive calibration data select based on average Pearson correlation coefficient with principal component analysis method is proposed. The source domain seizure data is adaptive selected to alternate the calibration seizure data. With the method, the accuracy further improved. The proposed methods are validated on CHB-MIT EEG data set to achieve state-of-the-art performance with 85.21% sensitivity, 93.76% specificity, and 92.50% accuracy.