Machine Learning-Based Prediction of Transition to Functional Upper Limb Recovery After Intensive Inpatient Rehabilitation in Early Subacute Stroke

Background/Objectives: Recovery of upper limb function after stroke is highly heterogeneous, and accurate prediction of clinically meaningful functional transition remains a major challenge in rehabilitation medicine. We developed and temporally validated machine learning (ML)-based prognostic models for predicting transition from non-functional movement to functionally usable upper limb capacity in patients undergoing intensive inpatient rehabilitation during the early subacute phase of stroke. Methods: This retrospective cohort study included 960 patients with ischemic or hemorrhagic stroke admitted to a tertiary rehabilitation center between 2010 and 2025. Three functional recovery outcomes were defined: motor impairment recovery, defined as Fugl-Meyer Assessment for Upper Extremity score ≥ 32; gross manual dexterity recovery, defined as Box and Block Test score ≥ 2 blocks/min; and functional pinch strength recovery, defined as pinch strength ≥ 1.1 kgf. Multidimensional predictors spanning demographic, clinical, neurophysiological, neuroimaging, and rehabilitation-related domains were integrated. Four ML algorithms were evaluated using stratified 5-fold cross-validation and temporal validation in a chronologically independent cohort (2024–2025). Models were developed under two tracks: Track A, incorporating only baseline variables available at admission (primary prognostic model), and Track B, additionally incorporating cumulative rehabilitation-related variables (exploratory). Results: Random Forest demonstrated the best overall performance. During temporal validation, models achieved AUROC of 0.800 for motor impairment recovery, 0.958 for gross manual dexterity recovery, and 0.888 for functional strength recovery. Baseline motor severity and corticospinal tract integrity were the dominant biological determinants of recovery. Earlier rehabilitation initiation and greater upper-limb robot-assisted therapy exposure were also associated with improved outcomes; however, these findings should be interpreted as observational associations subject to treatment-selection bias rather than evidence of causal effects. Conclusions: Probabilistic ML prediction integrating neural reserve and rehabilitation-related exposure variables can support individualized precision rehabilitation planning and improve functional outcome stratification in early subacute stroke.