IMU-Based Estimation of the Knee Contact Force using Artificial Neural Networks

Abstract

Knee joint contact force (KCF) plays a significant role in the occurrence and progression of knee osteoarthritis (KOA) disease. KCF can be used in monitoring rehabilitation progress after knee arthroplasty surgery and the design of prostheses. Currently, measuring KCF is dependent on the data extracted from gait laboratories. The combination of artificial neural networks (ANNs) and wearable technology can overcome the limitations imposed by lab-based analysis in measuring KCF. Therefore, the present study aimed to investigate the potential of a fully-connected neural network (FCNN) in predicting the KCF via three inertial measurement unit (IMU) sensors attached to the pelvis, thigh, and shank segments. Ten healthy male volunteers participated in this study. The 3D marker trajectories and ground reaction forces (GRFs) were captured at 200 Hz and 1000 Hz sampling frequencies during level-ground walking. Using a generic OpenSim model, the KCF was estimated through static optimization. The resultant KCF estimated by the musculoskeletal model was then used as the target of the neural network, while linear acceleration and 3D angular velocity data captured by three IMUs were considered as the network inputs. The network performance was investigated at intra- and inter-subject levels. Based on our findings, the proposed network of this study enables the prediction of KCF with 89% and 79% accuracy (based on the Pearson correlation coefficient) at the intra- and inter-subject levels, respectively. The results of this study promise the possibility of using IMU sensors in predicting KCF outside the lab and during daily activities.