Novel seizure detection algorithm based on multi-dimension feature selection

Abstract

In machine learning based seizure detection research studies, the number of features directly affects the performance of models. In order to decrease the amount of features under the premise of guaranteeing the performance of the model, a seizure detection algorithm based on multi-dimension feature selection is proposed. Firstly, the wavelet packet decomposition (WPD) method is applied to EEG signal, and different features are extracted from original EEG signals and sub-band signals. Random forest (RF) is then applied to analyze feature importance, in order to find redundant features. The importance of features is accumulated according to three dimensions: channel, sub-band signal and feature individual. Select the dimension with least importance and delete its relevant features. Repeat to find the dimension with least importance and delete its relative features until the least dimension importance exceeds the manual threshold. To evaluate the proposed method, 10-fold cross-validation is performed on the training set. The obtained accuracy (AC), specificity (SP) and sensitivity (SE) are 98.03%, 99.04% and 97.02%, respectively. Event-based real-time seizure detection is performed on the test set, and the obtained AC, SP, SE and FPRE are 99.76%, 99.84%, 87.23% and 0.245 times/h, respectively.