Novel Deep Learning Model to Estimate Knee Flexion and Adduction Moments with Wearable IMUs during Treadmill and Overground Walking

A major issue after total knee replacement (TKR) surgery is asymmetric gait kinetics, which increases knee loads on the non-operated knee. This imbalance accelerates osteoarthritis (OA) progression, often leading to a second contralateral TKR. There is a clear need for an advanced wearable system with multiple sensors to accurately estimate gait kinetics in natural environments. This study aims to develop a machine learning framework that exclusively uses wearable inertial measurement units (IMUs) during overground and treadmill walking to estimate knee flexion moment (KFM) and knee adduction moment (KAM), significant biomechanical factors linked to OA. We introduce a novel deep learning model that combines a Long Short-Term Memory (LSTM)-based Autoencoder and Variational Gaussian Process (VGP) to estimate the mean and uncertainty region of the KAM and KFM. Seventeen healthy participants performed treadmill walking trials, while a separate group of seventeen healthy participants performed overground walking trials for model training and validation. Results demonstrated Root Mean Square Errors (RMSE) of 0.49%BW·BH (body weight × body height) and 0.73%BW·BH for KAM and KFM, respectively, during treadmill walking and 0.74%BW·BH and 0.49%BW·BH for KAM and KFM respectively during overground walking, which is more accurate than existing approaches. The proposed model with wearable IMUs could enable knee health monitoring and rehabilitation for these key biomechanical factors linked to the progression of knee joint pathologies outside of traditional biomechanical laboratories with large, tethered equipment and into clinics, hospitals, and the community.