Evaluating the Impact of Uncertainty Learning Techniques on EEG-Based Seizure Prediction

Epileptic seizure prediction using electroencephalogram (EEG) data is critical for improving patient care, but challenges arise from the inherent complexity and variability of EEG signals. This study investigates the efficacy of data uncertainty learning (DUL) techniques in enhancing the prediction accuracy of a Vision Transformer (ViT) model. By systematically integrating advanced uncertainty learning methods-including Bayesian Neural Networks (BNNs), Monte Carlo Dropout, Cost-Sensitive Learning, and others-we demonstrate significant improvements in seizure prediction performance across two prominent datasets: CHB-MIT and AES. On the CHB-MIT dataset, the Bayesian Neural Network approach achieved a remarkable sensitivity of 89.2%, with a low false positive rate of 0.055 per hour and an area under the curve (AUC) of 0.950. Similarly, on the AES dataset, the BNN method yielded a sensitivity of 84.3%, an FPR of 0.047 per hour, and an AUC of 0.912. These results underscore the potential of uncertainty learning techniques to address data variability and enhance the reliability of seizure prediction models. By quantifying and managing uncertainty in the embedding space, our approach provides a robust framework for developing more accurate, clinically applicable biosignal prediction systems, ultimately offering improved intervention strategies for individuals with epilepsy.