Machine Learning for Predicting Motor Improvement After Acute Subcortical Infarction Using Baseline Whole Brain Volumes

Abstract

Background. Neuroimaging biomarkers are valuable predictors of motor improvement after stroke, but there is a gap between published evidence and clinical usage. Objective. In this work, we aimed to investigate whether machine learning techniques, when applied to a combination of baseline whole brain volumes and clinical data, can accurately predict individual motor outcome after stroke. Methods. Upper extremity Fugl-Meyer Assessments (FMA-UE) were conducted 1 week and 12 weeks, and structural MRI was performed 1 week, after onset in 56 patients with subcortical infarction. Proportional recovery model residuals were employed to assign patients to proportional and poor recovery groups (34 vs 22). A sophisticated machine learning scheme, consisting of conditional infomax feature extraction, synthetic minority over-sampling technique for nominal and continuous, and bagging classification, was employed to predict motor outcomes, with the input features being a combination of baseline whole brain volumes and clinical data (FMA-UE scores). Results. The proposed machine learning scheme yielded an overall balanced accuracy of 87.71% in predicting proportional vs poor recovery outcomes, a sensitivity of 93.77% in correctly identifying poor recovery outcomes, and a ROC AUC of 89.74%. Compared with only using clinical data, adding whole brain volumes can significantly improve the classification performance, especially in terms of the overall balanced accuracy (from 80.88% to 87.71%) and the sensitivity (from 92.23% to 93.77%). Conclusions. Experimental results suggest that a combination of baseline whole brain volumes and clinical data, when equipped with appropriate machine learning techniques, may provide valuable information for personalized rehabilitation planning after subcortical infarction.