A Hybrid Deep Learning Framework with Supervised Contrastive Learning for Robust Seizure Detection in Long-Term EEG

Accurate and timely automatic detection of epileptic seizures is crucial for reducing the workload for visually inspecting long-term electroencephalogram (EEG) and improving diagnostic efficiency. However, this task remains a critical challenge owing to the complex and non-stationary dynamics of ictal EEG activities and the severe class imbalance in long-term EEG data. Therefore, we develop an end-to-end hybrid deep learning framework named CNN-SwT, aiming to detect seizures by integrating Convolutional Neural Network (CNN) and Swin Transformer (SwT). CNN-SwT leverages CNN to capture the spatiotemporal local features from multi-channel EEGs and SwT to further derive their global dependencies. Furthermore, supervised contrastive learning is adopted in place of the traditional cross-entropy loss to address the extreme data imbalance inherent in seizure detection. Through extensive evaluation over the publicly available CHB-MIT long-term EEG database and our SH-SDU database, CNN-SwT achieved an event-based sensitivity of 100% with an FDR of 0.36/h and an event-based sensitivity of 94.44% with an FDR of 0.78/h, respectively. These outstanding cross-database results underscore the robustness of the CNN-SwT model and its potential to advance clinical seizure detection.