Accelerometry and the Capacity-Performance Gap: Case Series Report in Upper-Extremity Motor Impairment Assessment Post-Stroke.

This case series investigates whether traditional machine learning (ML) and convolutional neural network (CNN) models trained on wrist-worn accelerometry data collected in a laboratory setting can accurately predict real-world functional hand use in individuals with chronic stroke. Participants (N = 4) with neuroimaging-confirmed chronic stroke completed matched activity scripts-comprising instrumental and basic activities of daily living-in-lab and at-home. Participants wore ActiGraph CenterPoint Insight watches on the impaired and unimpaired wrists; concurrent video recordings were collected in both environments. Frame-by-frame annotations of the video, guided by the FAABOS scale (functional, non-functional, unknown), served as the ground truth. The results revealed a consistent capacity-performance gap: participants used their impaired hand more in-lab than at-home, with the largest discrepancies in patients with moderate to severe impairment. Random forest ML models trained on in-lab accelerometry accurately classified at-home hand use, with the highest performance in mildly and severely impaired limbs (accuracy = 0.80-0.90) and relatively lower performance (accuracy = 0.62) in moderately impaired limbs. CNN models showed comparable accuracy to random forest classifiers. These pilot findings demonstrate the feasibility of using lab-trained ML models to monitor real-world hand use and identify emerging patterns of learned non-use-enabling timely, targeted interventions to promote recovery in outpatient stroke rehabilitation.