Cross-Subject Seizure Detection by Joint-Probability-Discrepancy-Based Domain Adaptation

Detection of epileptic seizure from offline electroencephalogram (EEG) is of great significance in clinical diagnosis. Traditional epileptic seizure detection methods are usually based on the basic assumption that the training and testing data are sampled from datasets with the same distribution. However, in the context of epilepsy diagnosis, the EEG data vary from subject to subject, and the generalization performance of a classifier trained on data of multiple subjects typically degrades when applied to new subjects. To address this issue, we propose a cross-subject transfer learning framework for epileptic seizure detection to improve the classification performance on new subjects with unlabeled EEG samples (target domain) by transferring useful information from multiple subjects with labeled EEGs (source domain). In detail, first, an adversarial strategy is used to identify a set of source-domain EEG samples that are most suitable for transfer learning and thus are selected as training samples for the follow-up domain adaptation. Second, a novel domain adaptation method, the joint-probability-discrepancy-based domain adaptation (JPDDA), is proposed to predict the labels associated with the target-domain samples. Specifically, joint probability distribution discrepancy that measures the transferability between domains and discriminability between classes is proposed to learn a domain-invariant classifier jointly with structural risk and manifold consistency. Third, the epileptic seizure detection framework based on JPDDA is validated on the Children's Hospital, Zhejiang University School of Medicine (CHZU) dataset. Experimental results show that the proposed JPDDA can achieve high cross-subject detection accuracy, which reveals the good transferability of JPDDA.