Machine Learning-Based Estimation of Upper Extremity Function in Stroke Rehabilitation Using Body-Worn Inertial Sensors.

Neurological injuries such as stroke are a leading cause of disability, significantly impairing upper extremity (UE) function. Standardized clinical assessments are essential to evaluate patient function, monitor progress, and tailor interventions; however, these assessments are time-consuming to administer and require specialized training, limiting their accessibility. We developed a machine learning model to estimate Action Research Arm Test (ARAT) scores with wearable inertial sensors, aiming to reduce patients' and clinicians' workloads in rehabilitation. Twenty-three patients with chronic stroke performed the ARAT with ActiGraph sensors on their wrists and waist. Models used these data to predict the total ARAT score from a minimal set of UE tasks, selecting one item from each ARAT sub-test (grasp, grip, pinch, and gross movement). A nested cross-validation was used to optimize item selection, feature selection, and hyperparameters. The optimized model achieved a median absolute error of 3.81 points, and a coefficient of determination of 0.93 to estimate the total ARAT score. SHapley Additive explanations (SHAP) values identified key contributors to the predictions in each ARAT sub-test. The proposed approach demonstrates the potential of combining wearable sensor data with machine learning to facilitate more frequent and efficient monitoring of rehabilitation progress throughout the care continuum.