AI-DRIVEN PHYSIOTHERAPY REHABILITATION FRAMEWORK USING WEARABLE MOTION SENSORS, REAL-TIME MOVEMENT ANALYSIS, AND PERSONALIZED EXERCISE PRESCRIPTION FOR FUNCTIONAL RECOVERY AND PATIENT OUTCOME OPTIMIZATION

Abstract

Physiotherapy rehabilitation requires appropriate movement and an exercise dose matched to functional capacity. However, home sessions are rarely observed, compensatory movements may go undetected, and fixed protocols may not reflect individual recovery. This study presents a closed-loop, artificial intelligence (AI)-driven framework integrating wearable inertial measurement units (IMUs), movement recognition, exercise-quality assessment, clinician decision support, and personalized weekly exercise prescription. Because no dataset combining sensor signals, prescribed dose, and longitudinal outcomes was available, the evaluation used a literature-informed synthetic design, not human recordings. Sixty virtual patients completed 5,040 repetitions of six lower-limb exercises, recorded by IMUs on the thigh, shank, and trunk. Four model families were assessed using patient-grouped validation and bootstrap intervals. Gradient boosting achieved 0.994 accuracy for exercise recognition, a macro-F1 of 0.857 for movement-form classification, and an area under the receiver operating characteristic curve (AUC) of 0.9999 for compensation detection. Exercise-quality scoring achieved R² = 0.846 and a mean absolute error of 4.65 points.

The three-sensor configuration outperformed individual sensors and sensor pairs, while sampling rates as low as 12.5 Hz maintained similar classification performance. A 30° sensor rotation reduced macro-F1 from 0.848 to 0.720; rotation augmentation improved it to 0.800. Feature extraction and inference required approximately 6 ms per window on one server core. In a 12-week simulation with 150 virtual patients per arm, the combined framework improved the Timed Up and Go test by 3.11 seconds, compared with 1.99 seconds for a fixed protocol and 2.76 seconds for simulated clinician adjustment. Eighty percent met the assumed composite responder definition, compared with 51% under the fixed protocol. Feedback and dose personalization both contributed to improvement, while measurement noise reduced the benefit of personalization. These results describe framework behavior under stated assumptions, not clinical efficacy: sensor signals and recovery trajectories were simulated, so model performance represents an upper-bound estimate. Prospective evaluation using real patient data is necessary to establish generalizability, safety, clinical utility, and effects on functional recovery.