EXACT: A Meta-Learning Framework for Precise Exercise Segmentation in Physical Therapy

Wearable sensor technology has significantly enhanced healthcare quality, including physical therapy. However, due to the design of current deep learning models, existing works often ignore the unique variations of rest intervals between repetitions and variations in individual user progress, potentially hindering effective therapy outcomes. To address these limitations, we introduce EXACT, a novel framework designed to improve exercise segmentation and differentiate between exercise variations and rest intervals in PT applications. EXACT leverages a unique combination of a U-Net architecture integrated with Model-Agnostic Meta-Learning (MAML), enhanced with residual connections and attention mechanisms to capture subtle variations in exercise patterns and rest intervals. This approach addresses key challenges in segmenting dense, multivariate IMU data, providing a robust solution that adapts to new tasks with minimal retraining. EXACT achieves up to 20% improvement in segmentation Dice score over state-of-the-art U-Net models, demonstrating superior performance in distinguishing queried exercises from other exercises and rest intervals and handling variability in patient movements. Through rigorous evaluation and ablation studies, we demonstrate that attention and residual connections are essential for propagating relevant feature information and maintaining generalizability across varied exercise contexts. EXACT's adaptability and precision make it a valuable tool for real-time monitoring in PT, offering enhanced insights into patient progress and exercise quality in rehabilitation tracking. The code for our project is available at the EXACT Codebase1.