Estimating Upper-extremity Function with Raw Kinematic Trajectory Data after Stroke using End-to-end Machine Learning Approach.

Although there are some studies on the automatic evaluation of impairment levels after stroke using machine learning (ML) models, few have delved into the predictive capabilities of raw motion data. In this study, we captured kinematic trajectories of the trunk and affected upper limb from 21 patients with chronic stroke when performing three reaching tasks. Employing ML models, we integrated the recorded trajectories to predict scores of the Fugl-Meyer Assessment of the Upper Extremity (FMA-UE) of stroke patients. A transformer-based model achieved better metrics than Residual Neural Network (ResNet) and support vector regression (SVR). The trajectory successfully predicted FMA-UE scores, with the forward task (R2=0.905±0.028) outperforming the vertical task (R2=0.875±0.019) and horizontal task (R2=0.868±0.031). This pilot study demonstrated the capability of original trajectory data in tracking personal motor function after stroke and extended possibility of application in telerehabilitation.