An event-based filtering and weighted enhanced deep learning epileptic seizure prediction method.

Abstract

The timely detection of impending seizures can offer physicians a critical window of opportunity to implement interventions and enable epileptic patients to take preventive and coping measures promptly. Traditional seizure prediction research has primarily focused on segment-based classification of EEG signals rather than event-based prediction, resulting in a lack of temporal continuity in prediction outcomes and limiting their practical usefulness. To address this limitation, this study introduces an innovative two-step approach for seizure prediction. First, the PSO-DAM-2DCNN model, which combines a particle swarm optimization (PSO) algorithm with a two-dimensional convolutional neural network (2DCNN) that features a dual attention mechanism (DAM) integrating spatial and channel attention modules, is utilized to conduct segment-based prediction. Subsequently, a two-layer 'k-of-n' logic filter is employed to detect seizure events effectively. The proposed method demonstrates promising performance on both the CHB-MIT and the Huashan Hospital private datasets, excelling in both segment-based performance metrics and event-based metrics such as FPR/h, FNR, and TPR.