EEG seizure classification with temporal spiking neural networks and mutual information-based feature selection

Abstract Epilepsy affects over 50 million individuals worldwide, necessitating accurate and energy-efficient seizure detection systems. While Electroencephalography (EEG) is the clinical gold standard, automated classification faces challenges from high-dimensional, noisy, and non-stationary data. Traditional deep learning solutions often possess excessive computational demands, limiting their deployment in real-time, resource-constrained devices. This work proposes TemporalSNN, a lightweight spiking neural network designed for robust EEG seizure classification, optimized for neuromorphic deployment. The methodology utilizes a biologically inspired pipeline involving comprehensive feature extraction and Mutual Information (MI)-based feature selection to reduce input dimensionality. TemporalSNN employs Leaky Integrate-and-Fire dynamics to capture temporal dependencies, with performance evaluated on the Bonn, HAUZ, and Panwar benchmark datasets. Results demonstrate that MI-based feature selection significantly enhances model robustness and accuracy. On the challenging HAUZ dataset, binary accuracy improved dramatically from $$77.2\%$$ to $$92.5\%$$ (at 75% feature reduction), while on the Bonn dataset, binary accuracy increased from $$97.1\%$$ to $$98.2\%$$ with a substantial decrease in performance variance. These findings confirm that integrating MI-driven feature reduction with spiking computation yields efficient, low-complexity, and reliable seizure detection models suitable for practical, real-time clinical applications.