An attention-guided graph spiking approach for seizure localization and detection in healthcare

Epilepsy is a chronic neurological disorder with recurrent seizures, posing significant challenges for timely diagnosis and treatment. The manual identification of seizures from long-term EEG is labour-intensive, time-consuming, and heavily dependent on expertise, which underscores the need for automated seizure detection systems. This study proposes a patient-specific hybrid Graph Neural Network–Spiking Neural Network (GNN–SNN) framework, integrating attention-driven channel importance estimation with graph-based spiking dynamics for interpretable seizure detection. The EEG channels are represented as graph nodes with attention layers modeling their spatial relationships, and Adaptive Leaky Integrate-and-Fire neurons represent biologically inspired temporal dynamics. A key feature of the framework is its capacity to measure channel-wise contributions through attention weights, which yields human-interpretable information about which EEG channels contribute most to seizure detection. Each model is trained and optimized independently for every patient to learn their unique spatiotemporal EEG patterns, preserving the patient-specific design while maintaining a uniform architectural pattern for all subjects. The proposed model achieves robust performance on the CHB-MIT dataset, with 98.94 % accuracy, 95.01 % sensitivity, and 99.23 % specificity, while improving interpretability for EEG-based seizure detection. Although this model emphasizes dominant EEG channels, the model is not suitable for clinical localization and would require validation by intracranial EEG (iEEG) for translational applications. The integration of graph attention mechanisms with spiking computation provides better seizure detection with physiologically interpretable insights into EEG channel contributions. Future work will focus on validating these interpretability results on clinical gold standards, generalizing to patient-independent scenarios, and scaling up the framework for energy-efficient, real-time seizure monitoring. • Introduce a patient-specific framework integrating attention and graph spiking for seizure detection in healthcare. • Capture spatial relations with graph attention and temporal dynamics with spiking neural computation. • Provide interpretable channel contributions to support seizure onset zone localization in patients. • Achieve strong performance with high accuracy, sensitivity, and specificity on seizure data. • Offer a non-invasive method that aids presurgical planning and advances personalized epilepsy care.