KneeGuard: A Calibration-free Wearable Monitoring System for Knee Osteoarthritis Gait Re-training via Effortless Wearing

Abstract

Listen

Translate article

Gait re-training is an effective approach to slow disease progression and alleviate pain in knee osteoarthritis (KOA) patients. Personalized gait re-training strategies, based on knee loading and muscle forces, have shown promise in improving rehabilitation outcomes. Laboratory systems to monitor these metrics are unsuitable for daily use owing to the complicated setup and high-cost. Recently proposed wearable solutions try to fill this gap, but their in-lab calibration requirement still hinders practical usage. This paper introduces KneeGuard, a calibration-free gait re-training monitoring system that can estimate knee loading and muscle forces via effortless wearing. We identify the main issue of current calibration-needed systems is insufficient biomechanical information retrieval and modeling. To address this, we propose a user-friendly wearable prototype incorporating inertial measurement unit (IMU) and surface electromyography (sEMG) to obtain comprehensive biomechanical information including body geometrical changes and muscle contractions. For modeling, we design a biomechanic-inspired fusion framework based on multi-task learning and cross-modality attention to capture inter-modality biomechanical correlations. Additionally, since precise sensor placement required by current sEMG-based solutions is difficult to locate, we develop a circular sEMG array and propose a spatial-aware feature extraction module, achieving effective biomechanical feature extraction under effortless wearing. We collaborate with a medical center and collect a dataset from 21 KOA patients and 17 healthy subjects at different speeds. Notably, our dataset includes six gait types for KOA gait re-training, making it the first gait dataset with comprehensive re-training strategies. Evaluation demonstrates that KneeGuard achieves an average normalized root-mean-square error (NRMSE) of 9.95% in knee loading estimation and an average NRMSE of 8.75% in the estimation of muscle forces, comparable to the with-calibration results in existing works. We have open-sourced the code and a sample dataset in https://github.com/KneeGuard/KneeGuard.