Investigating the structure-function relationship of the corticomotor system early after stroke using machine learning

BACKGROUND: Motor outcomes after stroke can be predicted using structural and functional biomarkers of the descending corticomotor pathway, typically measured using magnetic resonance imaging and transcranial magnetic stimulation, respectively. However, the precise structural determinants of intact corticomotor function are unknown. Identifying structure–function links in the corticomotor pathway could provide valuable insight into the mechanisms of post-stroke motor impairment. This study used supervised machine learning to classify upper limb motor evoked potential status using MRI metrics obtained early after stroke. METHODS: Retrospective data from 91 patients (49 women, age 35–97 years) with moderate to severe upper limb weakness within a week after stroke were included in this study. Support vector machine classifiers were trained using metrics from T1- and diffusion-weighted MRI to classify motor evoked potential status, empirically measured using transcranial magnetic stimulation. RESULTS: Support vector machine classification of motor evoked potential status was 81% accurate, with false positives more common than false negatives. Important structural MRI metrics included diffusion anisotropy asymmetry in the supplementary and pre-supplementary motor tracts, maximum cross-sectional lesion overlap in the sensorimotor tract and ventral premotor tract, and mean diffusivity asymmetry in the posterior limbs of the internal capsule. INTERPRETATIONS: MRI measures of corticomotor structure are good but imperfect predictors of corticomotor function. Residual corticomotor function after stroke depends on both the extent of cross-sectional macrostructural tract damage and preservation of white-matter microstructural integrity. Analysing the corticomotor pathway using a multivariable MRI approach across multiple tracts may yield more information than univariate biomarker analyses.