Related Experiment Video
Updated: May 12, 2025
Movement Retraining using Real-time Feedback of Performance
Published on: January 17, 2013
Optimizing the dynamic treatment regime of outpatient rehabilitation in patients with knee osteoarthritis using
Sijia Liu1, Jiawei Luo2, Chengqi He3
1Institute of Orthopedics, West China Hospital, Sichuan University, Chengdu, 610041, Sichuan, China.
Background:
Knee osteoarthritis (KOA) is a prevalent chronic disease worldwide, and traditional treatment methods lack personalized adjustment for individual patient differences and cannot meet the needs of personalized treatment.
Methods:
In this study, a dedicated knee osteoarthritis bank (KOADB) was constructed by collecting extensive clinical data from patients. Random forest was used to select the features that had the greatest impact on treatment decisions from 122 questionnaire items. The questionnaire design was optimized to reduce the burden on patients and ensure the validity of data collection. Then, based on the key features screened out, a dynamic treatment recommendation system was constructed by using deep reinforcement learning algorithms, including Deep Deterministic Policy Gradien(DDPG), Deep Q-Network(DQN) and Batch-Constrained Q-learning(BCQ). A large number of simulation experiments have verified the effectiveness of these algorithms in optimizing the treatment strategy of KOA. Finally, the applicability and accuracy of the model were evaluated by comparing the treatment behaviors with actual patients.
Results:
In the application of deep reinforcement learning algorithms to treatment optimization, the BCQ algorithm achieves the highest success rate (79.1%), outperforming both DQN (68.1%) and DDPG (76.2%). These algorithms significantly outperform the treatment strategies that patients actually receive, demonstrating their advantages in dealing with dynamic and complex decisions.
Conclusions:
In this study, a deep learning-based KOA treatment optimization model was developed, which was able to adjust the treatment plan in real time and respond to changes in patient status. By integrating feature selection and reinforcement learning techniques, this study proposes an innovative method for treatment optimization, which offers new possibilities for chronic disease management and demonstrates certain feasibility in the development of personalized medicine and precision treatment strategies.
