Publication

Abstract : Epilepsy affects over 50 million people worldwide, yet timely and reliable diagnosis is still not widely available, especially in resource-constrained environments. Improving the accessibility, interpretability, and reliability of non-invasive diagnostic tools is essential for advancing equitable neurological healthcare and improving clinical outcomes worldwide. Although electroencephalography is the clinical standard, its low sensitivity in interictal periods encourages the development of accessible imaging-based alternatives. In this work, we introduce an interpretable structural MRI-based framework for epilepsy diagnosis and phenotype stratification based on structural covariance networks (SCNs). To capture local morphological relationships between brain regions, a novel subject-specific k-nearest-neighbor (kNN)-based SCN formulation is introduced. Both conventional Pearson correlation and the proposed kNN-based approach were used to construct subject-specific SCNs for GM, WM and CSF, and graph-theoretical features were extracted and classified using classical machine learning models. The models trained on kNN-derived SCNs demonstrated F1-scores of > 0.90 with perfect recall with fused multi-tissue features, which were superior to both single-tissue and correlation-based networks. Explainability using SHAP demonstrated that white matter and cerebrospinal fluid network features played a significant role in the diagnostic decisions, thereby improving biological interpretability. In addition to diagnosis, phenotype-stratified analyses revealed clinical significance: the burden of focal-to-bilateral tonic–clonic seizure (FBTCS) was correlated with extensive white matter network disruption, and elevated thalamic nodal efficiency was correlated with seizure generalization. Compared with a deep learning benchmark, the proposed framework was significantly more interpretable and computationally efficienct, with a 30-fold reduction in network construction time. Collectively, these findings establish an explainable multi-tissue SCN framework for structural MRI-based epilepsy diagnosis and seizure burden stratification that supports earlier diagnosis, personalized clinical decision-making, and improved neurological healthcare, while providing a scalable approach that may be extended to other network-level neurological disorders.