Transfer Component Analysis to Reduce Individual Difference of EEG Characteristics for Automated Seizure Detection

Abstract

About 50 million people worldwide are suffering from epilepsy. Automated epileptic seizure detection has been widely studied so far, which brings great support to patients’ health and quality of life. However, a well-trained patient-specific seizure detection model usually shows poor performance when classifying electroencephalogram (EEG) signals of a new patient. This may due to the huge individual differences in physiological signals. More specifically, the distribution of feature space across patients differ greatly from each other. In this study, we firstly extracted highly separable features from dual-tree discrete wavelet parameters. Then we employed transfer component analysis (TCA) to construct a latent feature subspace. Features of different patients share a similar distribution when projected onto the latent subspace, so that a model trained on existing patients can be applied to a new one. Through validation on an open access scalp EEG dataset which contains EEG signals of 24 epileptic patients, the model trained in TCA feature subspace outperforms that trained in original feature space when applied to new patients excluded from the training set.