Hybrid CNN-LSTM and PPO Architecture for Adaptive Home-Based Physical Therapy

Abstract

This study outlines an intelligent feedback system for personalized, home-based physical rehabilitation. It integrates a convolutional neural network long short-term memory (CNNLSTM) model to detect joint movement deviations and a proximal policy optimization (PPO) agent for real-time adaptive feedback. By combining these technologies with multimodal sensory inputs, such as skeletal tracking, inertial measurement units (IMUs), and electromyography (EMG) signals, the system provides tailored guidance, corrects posture, regulates rest, and enhances engagement. Key features include a closed-loop framework that merges deep learning and reinforcement learning (RL) for dynamic, real-time feedback, as well as multimodal sensing for a comprehensive view of user activity. Scenario-based simulations tested performance, showing reduced missed corrections (38.4%), increased productive exercise time (21.7%), improved fatigue management (35%), and robust PPO agent stability after 3,000 training episodes, even during sensor failures. Results highlight the system’s potential for personalized, adaptive intervention in both home and clinical settings, enhancing rehabilitation effectiveness and accessibility despite common challenges such as sensor errors.