Motif-informed hypergraph spectral learning for recording-level EEG seizure classification

Abstract Accurate classification of seizure and non-seizure EEG remains challenging because of the complex and non-stationary temporal characteristics of epileptic activity. This study proposes a framework that integrates temporal motif discovery with hypergraph-based representation learning for EEG seizure classification. EEG segments are divided into overlapping subsequences, and recurring temporal patterns are identified through K-means clustering to construct a global motif vocabulary. Each original EEG recording is represented by a motif-occurrence histogram, and a motif-informed hypergraph is constructed in which recordings form vertices and shared motif patterns define hyperedges. Spectral decomposition of the normalized hypergraph Laplacian produces compact representations that are classified using Random Forest and a one-dimensional convolutional neural network. Random Forest achieved 96% accuracy on the primary recording-level evaluation, compared with 94% for the CNN. Repeated stratified recording-level validation further showed stable Random Forest performance, with a mean accuracy of 95.3 ± 1.4% across 10 splits. The results demonstrate that motif-informed hypergraph spectral learning effectively captures recurring temporal structure together with higher-order relationships among EEG recordings, providing a structured and discriminative representation for seizure-class classification. As the evaluation was performed on short, pre-segmented single-channel benchmark recordings, the findings should be interpreted as benchmark recording-level performance rather than as clinical or patient-level validation.