MTL-SSU: A Multi-Task Self-Supervised Learning Framework for Epileptic Seizure Prediction

Predicting epileptic seizures is crucial for reducing patient suffering and informing clinical treatments. However, training supervised models for seizure prediction requires extensive labeled data, which is labor-intensive and costly. We propose a novel Multi-Task Self-Supervised Learning framework with U-Net architecture (MTL-SSU) for EEG-based seizure prediction. Unlike conventional self-supervised methods, MTL-SSU incorporates domain-specific knowledge for epileptic seizure analysis. Leveraging U-Net, the approach performs two key tasks: mimic segmentation and channel discrimination. These tasks enable the model to capture effective feature representations from unlabeled EEG signals. After pretraining, the U-Net encoder and a linear classifier are fine-tuned for seizure prediction. Employing a rigorous k-fold cross-validation strategy on the CHB-MIT database, MTL-SSU achieves 93.55% accuracy and 0.9775 AUC in patient-specific epilepsy prediction. Notably, even in semi-supervised learning scenarios with limited labeled data, MTL-SSU exhibits exceptional performance, highlighting its significant potential for clinical application.