Multivariate Searchlight Decoding of Structural Disconnection and Network Features for Predicting Post-Stroke Deficits

Abstract Objective. Lesion-symptom mapping and structural disconnection analyses have advanced our understanding of post-stroke deficits but remain predominantly based on forward inference, limiting identification of regions specifically predictive of behavioral impairment. We developed a multivariate voxelwise searchlight decoding framework for inverse inference integrating probabilistic structural disconnection with network-topological features. Approach. The analysis included 326 stroke patients from the GRECogVASC cohort assessed six months after stroke. For each patient, an age-specific probabilistic structural disconnection map was reconstructed using normative diffusion MRI tractography from 403 healthy controls. Within each voxel-centered searchlight, four features were extracted: disconnection probability, degree, clustering coefficient, and closeness centrality. Support vector machine classifiers were trained within deficit-specific decoding spaces using balanced bootstrap aggregation and majority voting to predict motor, executive, and processing-action-speed impairments. Statistical significance was assessed using 1,000 permutations, retaining voxels that consistently performed above chance across bootstrap iterations. Decoding-derived regions were subsequently assessed for within-cohort individual-level classification. Main results. Initial validation focused on left and right motor impairments, identifying predictive regions primarily within the contralateral precentral gyrus, internal capsule, and corticospinal tract, with mean accuracy ranges of 68.7–77.9%, 66.4–74.6%, and 64.8–71.8%, respectively. At the individual level, decoding-derived masks preserved overall classification performance, with accuracy improvements of up to 2.6% for individual deficits while reducing voxel dimensionality by up to 95.4%. Executive dysfunction and processing-action-speed deficits showed statistically above-chance but lower accuracies (54.5–66.9%). Executive dysfunction involved predominantly left-lateralized frontotemporal regions, whereas processing-action-speed deficits involved distributed fronto-insular, temporal, parietal, visual-association, and commissural pathways. Significance. The multivariate searchlight decoding framework integrating structural disconnection and network topology provides an inverse-inference approach complementary to lesion- and disconnection-symptom mapping for identifying spatially specific predictive patterns after stroke. The strong motor findings provide proof-of-concept support for the approach, whereas the cognitive results support its extension to distributed behavioral domains while underscoring the need for independent replication before neuroanatomical or clinical interpretation (Clinical trial registration: NCT01339195).