Unsupervised Subtyping of Focal Epilepsy IEDs via Graph-Based MEG Networks.

OBJECTIVE: Interictal epileptiform discharges (IEDs) are clinically important biomarkers in focal epilepsy, yet their source-level network organization and subtype structure remain insufficiently characterized. This study aimed to identify reproducible IED network subtypes from source-level magnetoencephalography (MEG) connectivity and to examine whether these subtypes capture clinically meaningful information related to MEGIEDlocalization. METHODS: Fifty-five patients with focal epilepsy were retrospectively included. After manual IED identification, source-level MEG signals were reconstructed and converted into 148 × 148 phase-locking value connectivity matrices. These matrices provided the basis for graph-based representation learning, in which DeepWalk, Node2Vec, Graph Convolutional Network (GCN), Graph Attention Network (GAT), and Multi-View Graph Representation Learning (MVGRL) models were paired with K-means or spectral clustering to identify latent interictal epileptiform discharge (IED) network subtypes. The selected subtype structure was then examined in a downstream clinical prediction task. Specifically, a subtype enhanced model encoded each patient as a bag of IEDconnectivity matrices, summarized IED-level embeddings through attention based multiple-instance pooling, and integrated subtype composition through soft prototype routing for MEG IED localization classification. RESULTS: A total of 8,509 IEDs were identified. MVGRL combined with K-means achieved the best clustering performance and revealed four IED network subtypes. These subtypes differed in global network efficiency, nodal organization, and associations with MRI-derived cortical thickness. Overall network efficiency followed the order Subtype 4, Subtype 1, Subtype 2, and Subtype 3. Matrix-level validation showed that the four subtypes could be classified from graph theoretical features with an overall SVM accuracy of 93.90%. In patient-level validation, the subtype-enhanced model achieved 58.8% balanced accuracy for temporal-versus-other MEG IED localization classification, outperforming both the graph-feature random forest baseline and the deep multiple-instance baseline without subtype routing. CONCLUSION: This study demonstrates that leveraging the spatiotemporal dynamics of IEDs through unsupervised clustering enhances the understanding of shared/unique network features across focal epilepsy types. SIGNIFICANCE: These findings provide an imaging-based framework for characterizing the heterogeneity of interictal epileptiform network states in focal epilepsy. Further validation against independently established clinical seizure-onset and epileptogenic zone labels is required before the framework can be interpreted as a clinical seizure-localization tool.