<p>Stroke leads to gait impairment and reduced physical function, making accurate functional assessment essential for effective rehabilitation planning. However, conventional clinical scales may be limited by discrete scoring and examiner dependency, whereas laboratory-based motion analysis systems are costly and difficult to implement routinely. This study aimed to classify physical function in stroke patients using gait features derived from a single shank-mounted inertial measurement unit (IMU) and to compare the classification performance with that of a motion capture (MoCap) system. Thirty-two stroke patients completed the Short Physical Performance Battery (SPPB), and participants were divided into low- and high-physical-function groups using a cutoff score of 10. MoCap-based features included spatiotemporal gait parameters and joint kinematics, whereas IMU-based features included descriptive statistical, gait-related, frequency, and variability features extracted from raw sensor signals. A filtering-based feature selection procedure was applied, and the top 40 features were selected using mutual information. Six machine learning models were evaluated using leave-one-out cross-validation. Among the evaluated models, the IMU-based linear discriminant analysis model achieved an accuracy of 0.91, which was comparable to that of the MoCap-based XGBoost model (0.94). These findings indicate that a single shank-mounted IMU may provide sufficient gait information for objective classification of physical function in stroke patients, supporting its use as a practical and low-burden assessment tool in stroke rehabilitation.</p>