Lesion-Derived Estimated Functional and Structural Connectomes Predict Post-Stroke Recovery Outcomes

Abstract

Accurate prognosis after stroke remains challenging, limiting the ability to predict individual recovery. While advanced connectome MRI can provide further prognostically-useful information regarding affected brain networks, its impractical nature has given rise instead to lesion-based estimates of network disruptions. Here, we present a novel pipeline that leverages our recently developed connectome analysis tools to estimate structural disconnection (eSDC) and functional connectivity (eFC) directly from N=96 patient lesion masks for use in predicting motor, language, attention and memory outcomes at multiple timepoints after stroke. Our approach jointly leverages our Network Modification (NeMo) Tool, a database of white matter tractography results that estimates structural disconnectivity (eSDC), and the Krakencoder, a joint autoencoder that estimates functional connectomes (eFC) from the NeMo Tool's output. Models based on eSDC and eFC predicted almost all outcomes, particularly motor outcomes, significantly better than models that used patients' fMRI-based observed FC (oFC) and models based on only lesion information and covariates. Estimated connectome measures also i) retained meaningful inter-individual variability when compared to observed functional connectomes, ii) reflected population-level, stroke-related changes in functional connectomes, and iii) resulted in outcome prediction models with feature importances that aligned with expected networks of relevance to the given outcome. Feature importance analyses also identified distinct predictive signatures across both modality and outcome domain, indicating that eSDC and eFC capture non-redundant, biologically meaningful information about post-stroke network disruption. Together, these findings establish our lesion-based estimates of post-stroke structural and functional connectomes as a practical, information-rich alternative to advanced imaging in patients. This approach allows for probing an individual's affected neural circuitry underlying stroke impairment and recovery, paving the way toward more efficient, accessible, and personalized approaches to post-stroke prognosis and intervention.