MIND-based cortical similarity networks reveal distinct cortical alterations in motor subtypes of Parkinson's disease

Parkinson's disease (PD) is a biologically heterogeneous disorder, yet conventional motor classifications may not fully reflect underlying cortical organization. We combined Morphometric INverse Divergence (MIND) networks, data-driven clustering, and explainable machine learning to investigate cortical morphometric heterogeneity in 219 patients with PD and 146 healthy controls. HYDRA clustering of regional MIND strength identified two stable cortical subgroups characterized by globally lower (Cluster 1) versus higher (Cluster 2) morphometric similarity, which only partially overlapped with tremor-dominant (TD) and postural instability/gait difficulty (PIGD) phenotypes. Combining cluster membership with clinical classification yielded four stratified subgroups. Supervised machine learning within a nested cross-validation framework then tested whether MIND-derived patterns discriminated these subgroups better than clinical labels. Discrimination between TD and PIGD was limited (AUC ≈ 0.53-0.55), but markedly higher within clinical phenotypes (AUC up to 0.96 for TD Cluster 1 vs. Cluster 2 and 0.85 for PIGD Cluster 1 vs. Cluster 2) and across subgroup contrasts (AUC 0.90-0.93). The two-cluster organization was replicated in the independent multi-site PPMI cohort, where the discovery-trained classifier also transferred without retraining (AUC = 0.79), supporting generalizability. SHAP analyses showed that the discriminative features were distributed across frontotemporal, frontoparietal, cingulate, and temporo-occipital similarity pairs, involving regions anatomically assigned to the Default Mode, Frontoparietal Control, Salience/Ventral Attention, Visual, and Somatomotor networks. In PIGD Cluster 1, reduced morphometric similarity between regions assigned to the Somatomotor and Frontoparietal Control networks was associated with greater motor severity. These findings indicate that individualized morphometric similarity networks capture a dimension of PD heterogeneity beyond conventional motor phenotypes and may provide a biologically informed framework for patient stratification.