Hybrid LPF-LSTM Model for Enhanced Epileptic Seizure Detection in EEG Signals

Abstract

Accurate prediction and detection of epileptic seizures using electroencephalogram (EEG) signals are crucial for advancing clinical diagnostics and improving patient outcomes. This letter proposes a distinctive hybrid framework that combines a linear prediction filter with a long short-term memory network, designed to address challenges in noise reduction and temporal pattern recognition in EEG signals. The detection performance, particularly specificity, is enhanced by applying dynamic thresholding based on residual energy analysis. The proposed method, with key aspects of the validation framework, enhances cross-patient generalization by validating the model on the CHB-MIT Scalp EEG Database across four distinct age groups: infants, children, adolescents, and young adults. The hybrid approach achieved 98.4% accuracy, 97.8% sensitivity, 96.2% specificity, and 0.98 area under the curve, outperforming traditional approaches by 3%–5%.