Abstract
Stroke is a leading cause of long-term disability worldwide, necessitating efficient motor function assessment to guide personalized rehabilitation. The Fugl-Meyer Assessment for the Upper Extremity (FMA-UE) is a clinical gold-standard tool, but it is time consuming and requires trained clinicians, which limits its frequency of use and accessibility. While wearable sensors and deep learning offer promising avenues for remote assessment, accurately estimating detailed sub-scores of specific motor functions remains a significant challenge. This work introduces a deep learning framework for automated estimation of FMA-UE total and subdivision scores. Data was collected from 15 participants using four inertial measurement units (IMUs) positioned on the arm and trunk. Each participant performed seven specialized functional motions designed for comprehensive joint synergy involvement within ten minutes. A therapist-rated FMA-UE provided true scores. The proposed model leverages the integration of an LSTM-based autoencoder and mixup augmentation to enhance generalization and robustness. Evaluated through a leave-one-subject-out cross-validation (LOSOCV), the estimator demonstrated strong performance, achieving R(2) values exceeding 0.82. Pearson's correlation coefficient r was more than 0.90, and the normalized root-mean-square errors (NRMSE) were below 0.14 for all subparts (A-D). Crucially, the total FMA-UE score was estimated with an NRMSE of 0.0678. These results show that a concise, sensor-based assessment can reliably predict detailed motor function scores.