Epilepsy Detection and Prediction Using a High-Performance Ensemble Learning Algorithm Based on Feature Fusion

Abstract

Epilepsy detection and prediction remain among the most challenging data analysis tasks for chronic brain disorders. Manual detection is time-consuming, and the inability to predict seizures artificially has driven the development of numerous automatic algorithms. However, most electroencephalogram (EEG) classification methods rely on a single feature, resulting in poor accuracy, and typically address either detection or prediction rather than both. In this study, we propose an integrated model based on feature fusion and selection. First, EEG signals were segmented into interictal, preictal, and ictal phases. Next, we applied the Discrete Wavelet Transform (DWT) for signal decomposition and extracted Standard Deviation (STD), Root Mean Square (RMS), Approximate Entropy (ApEn), and Fuzzy Entropy (FuzzyEn) features from each subband. A Random Forest (RF) algorithm then selected the ten most important features, and Ensemble (ENS) learning algorithms were employed for both detection and prediction. Validation on the CHB-MIT dataset showed that, for detection, the average accuracy, sensitivity, precision, F1-score, specificity, and Matthews correlation coefficient across 24 patients were 99.07%, 98.71%, 99.40%, 99.05%, 99.40%, and 0.98, respectively. For prediction, the corresponding metrics were 95.41%, 96.74%, 94.50%, 95.55%, 94.02%, and 0.94. The proposed model thus enables high-precision automatic detection and prediction of epileptic EEG signals, allowing timely preventive measures to mitigate seizure impact.