A Closed-Loop Intelligent Management System for Postoperative Orthopedic Rehabilitation: Integrating Deep Reinforcement Learning with Wearable Sensing and Prognostic Prediction

Background: Conventional orthopedic postoperative rehabilitation protocols are typically static and generic, lacking the capacity for dynamic adjustment based on a patient's real-time recovery status. This often results in suboptimal rehabilitation efficiency and an increased risk of re-injury. This study aims to design and provide a proof-of-concept for a novel dynamic closed-loop management system. By deeply integrating Patient-Generated Health Data (PGHD) and Deep Reinforcement Learning (DRL), the system seeks to enable real-time, personalized optimization of rehabilitation regimens and continuous prediction of long-term functional outcomes. Methods: A three-layer system architecture was constructed, comprising an intelligent sensing layer, an AI decision-making & prediction layer, and an interactive feedback layer. Using wearable inertial measurement units (IMUs) and surface electromyography (sEMG) sensors, the system continuously collects multi-dimensional PGHD, including movement quality, exercise intensity, adherence, and pain feedback. This data is encoded into a "patient state space." A DRL agent, trained with the Proximal Policy Optimization (PPO) algorithm, learns an optimal policy to dynamically adjust the subsequent rehabilitation prescription (encompassing exercise type, intensity, frequency, and progression rate) based on the current state. The objective is to maximize a cumulative reward function that balances short-term safety, medium-term adherence, and long-term functional improvement. Furthermore, a Temporal Convolutional Network (TCN) is integrated to provide real-time predictions of future functional recovery curves based on historical state-action sequences. Results: The system was validated in a simulated environment for Anterior Cruciate Ligament Reconstruction (ACLR) postoperative rehabilitation. The trained DRL agent generated differentiated rehabilitation strategies for virtual patients with distinct recovery trajectories. Compared to static protocols, the DRL-based strategy reduced the average time to "safe return to light sports" by 15% and lowered the incidence of simulated "re-injury" events by 40%. The TCN prognostic prediction module achieved Mean Absolute Errors (MAE) of 3.2, 4.8, and 6.5 points (on a 100-point scale) for predicting functional scores at 2, 4, and 8 weeks into the future, respectively, significantly outperforming an ARIMA baseline model. Conclusions: This simulation-based study provides preliminary validation for the feasibility of a dynamic, closed-loop rehabilitation management system for orthopedic postoperative care, founded on PGHD and DRL. The system realizes a real-time intelligent cycle of "monitoring-prediction-intervention," offering a potential tool for highly personalized, adaptive, and prospectively informed management in sports medicine and other orthopedic rehabilitation settings. Subsequent prospective clinical studies are warranted to further validate its safety, efficacy, and clinical utility.