Behavioral Clusters and Lesion Distributions in Ischemic Stroke, Based on NIHSS Similarity Network.

Stroke, a leading cause of mortality and disability, results in diverse dysfunctions linked to brain lesion locations. The intricate relationship between lesions and symptoms often defies linear analysis methods. Unraveling these connections can yield valuable insights to enhance patient care, optimize rehabilitation strategies, and unveil fundamental principles of healthy brain function. This study introduces a novel unsupervised framework to stratify patients into clinically coherent subgroups based on behavioral symptom profiles and identify their distinct neural correlates. NIHSS assessments are modeled as ordinal feature vectors, integrating symptom prevalence, severity, and covariance patterns into a unified measure of behavioral similarity among stroke survivors. The resulting similarity network is partitioned using Repeated Spectral Clustering, which accumulates partition evidence for stable subgroup discovery. Voxel-wise lesion analysis subsequently highlights each subgroup's collective neuroanatomical signatures. Despite being identified in a completely unsupervised manner based solely on NIHSS scores, the emergent clusters correspond to well-documented syndromes, validating the purely data-driven symptom groupings alongside established neurological knowledge. Clusters exhibit critical voxels in group-specific anatomical locations, even when average lesion maps spatially overlap, suggesting that our method disentangles functionally distinct substrates within shared vascular territories. Our workflow represents a significant methodological advancement, providing robust, clinically relevant insights into symptom phenotyping and lesion patterns. The framework's mathematical transparency and validation against canonical knowledge underscore its potential for generalization to multimodal biomarkers and broader biomedical research. To foster reproducibility, we provide open-source code.