A Machine Learning Approach to Predict Post-stroke Functional Scores According to Individual Connectivity Analyses

Introduction: Stroke induces underlying structural damage which causes brain network dysfunction that contributes to a variety of neurological sequelae. Anatomically specific connectivity analyses combined with advanced statistical approaches may allow us to examine in vivo the prominent neuroanatomic substrates responsible for the clinical pattern of deficits.

Methods: Retrospective analyses of 22 patients a single center with acute ischemic stroke. Functional status scores at hospital presentation were assessed according to the National Institutes of Health Stroke Scale (NIHSS) and Fugl-Meyer Assessment (FMA) scales by two independent neurologists. Individualized connectivity analyses were completed to identify structural and functional connectivity damage or dysfunction within the sensorimotor network. A standard parcellated brain atlas according to the Human Connectome Project (HCP) was utilized. Machine learning analyses were completed utilizing elastic net regularization to assess which structural and functional connectivity features were most essential to be able to best predict FMA and NIHSS total scores as well as individual components of each scale.

Results: Increasing structural damage to the CST, cortical-subcortical projection fibers and deep subcortical-subcortical fibers were each associated with a worse functional score at presentation (p < 0.05). When analyzing functional connectivity, specific parcels were identified within the supplementary motor area and salience networks which had values that were associated with total functional NIHSS and FMA patient scores. These parcels included areas SCEF and 3a contralateral to the lesion site and areas 3a, 6a, and 6mp ipsilateral to the lesion (p < 0.05, each). A number of additional specific parcels could be identified which related to specific aspects of each scale.

Conclusion : Utilizing advanced statistical analyses and a personalized approach, individual neuroanatomic substrates can be identified that may allow us to further predict clinical patterns of deficits following stroke and also anatomically specific targets for subsequent rehabilitation.