Mixed supervised cross-subject seizure detection with transformer and reference learning

Automatic seizure detection aims to identify occurrences of epileptic seizures , enabling timely seizure intervention and protecting patients’ safety. In recent years, deep learning has significantly promoted the research progress in the field of seizure detection. In this paper, we propose a cross-subject seizure detection system based on a transformer encoder. A novel data fusion approach is leveraged to mitigate the imbalance issue of seizure and non-seizure data. Meanwhile, a new mapping method is employed to replace traditional feature extractors, effectively enhancing the real-time capabilities of the system. A mixed supervised and unsupervised learning approach, coupled with a specially designed loss function, is utilized to strengthen the model's ability to capture temporal and spatial features in electroencephalogram (EEG) signals. Furthermore, an innovative learning strategy named reference learning is proposed to enhance the model's generalization performance. Finally, the proposed system was evaluated on the publicly available CHB-MIT dataset using the Leave-One-Out Cross-Validation (LOOCV) strategy. The system achieved a segment-based sensitivity of 91.06 % and an event-based sensitivity of 93.59 % in the cross-subject seizure detection task. • An end-to-end transformer-based model for cross-subject seizure detection is proposed. • The mixed supervised training method is used to train the model. • Enhance model performance through an innovative reference learning strategy. • The proposed model effectively captures the temporal and spatial features of EEG.