A Multi-view Representation Learning Approach for Seizure Detection Over Multi-channel EEG Signals

Abstract

AbstractWe introduced a multi-view representation learning approach for seizure detection on EEG signals. To implement the multi-view approach for seizure detection, time and frequency domain features (views) are considered. The CHB-MIT dataset is used as an experimental dataset. For preprocessing of raw EEG data, a Butter-worth bandpass filter with a range of frequency between 0.5 and 30 Hz is used. With a window size of 1s, EEG data is split into time and frequency domain images. As EEG signals are in the time domain, time domain views are calculated. Applying the FFT on EEG signal results in a frequency domain, which is then segmented into images. From time and frequency domain images, we were able to create two different types of views on the same dataset. Both time and frequency domain images are combined for a multi-view concept. For multi-view, a CNN model is tested, and we got average accuracy of 0.9821, sensitivity of 0.5878, specificity of 0.991, and F1-score of 0.5933 using only five subjects’ data from the CHB-MIT.KeywordsElectroencephalogram (EEG)Seizure detectionConvolutional neural network (CNN)Multi-view representation learning