An explainable spatio-temporal graph neural network for seizure localization using multi-domain analytics

Detecting epileptic seizures from scalp electroencephalographic (EEG) is made difficult by the nonstationary nature of brain signals and the complex spatio-temporal interactions between neural regions in different locations. This paper introduces a spatio-temporal channel-attention graph neural network (ST-CA-GNN) for automatic seizure detection based on its ability to represent dynamic functional connectivity across all EEG channels. EEG signals are modelled as time-dependent graphs, where each node of the network corresponds to an EEG channel and edges capture phase-based synchronisation patterns between brain regions. Then each node is represented by multi-domain features encompassing time-domain statistics, frequency-domain features, complexity features and dynamical characteristics derived from dynamic mode decomposition. By jointly learning spatial dependencies and temporal dynamics ST-CA-GNN distinguishes ictal and interictal brain states from multi channel EEG recordings. The attention mechanism adaptively weights channel contributions to identify EEG regions most relevant to seizure activity. The model interpretability is improved through post hoc explainability analysis that can quantify channel wise importance and crucially provide clinically meaningful information pertaining to seizure related network dynamics. The proposed method is evaluated on the CHB-MIT and Siena EEG datasets and results are robust across subjects. Statistical analysis confirms that ST-CA-GNN significantly outperforms baseline models across all evaluation metrics. These findings suggest that the proposed attention-based spatiotemporal graph learning framework is an effective approach that leverages spatial and temporal EEG relationships to identify seizure-associated scalp EEG channels and provide supportive information for seizure localization.