Chaos-Modulated Spiking Neural Networks for Seizure Detection: A Dynamical Systems Approach to Clinical EEG

Abstract

Automated seizure detection from EEG remains challenging due to signal complexity and the need for patient-independent generalization. While deep learning offers high accuracy, it lacks interpretability regarding neural dynamics. Spiking Neural Networks (SNNs) provide a biomimetic alternative but often struggle with limited dynamical states. We propose a Chaos-Modulated Spiking Neural Network (Chaos-SNN) operating at the 'edge of chaos.' By incorporating controlled chaotic modulation into a recurrent spiking reservoir, we increase dynamical sensitivity while maintaining bounded stability and energy efficiency. The model was evaluated on controlled tasks and the clinical CHB-MIT dataset using leave-one-patient-out (LOPO) cross-validation. Our findings indicate that chaotic modulation reorganizes spike-train structures, creating interpretable distinctions between ictal and interictal patterns. While chaos did not significantly increase absolute accuracy in clinical settings, it substantially reduced cross-patient variability and stabilized internal dynamics. We conclude that chaos serves as a dynamical prior that balances sensitivity and stability, offering a robust, energy-efficient neuromorphic solution for clinical EEG monitoring.