A Machine Learning and Wearable Sensor Based Approach to Estimate External Knee Flexion and Adduction Moments During Various Locomotion Tasks

Bernd J Stetter; Frieder C Krafft; Steffen Ringhof; Thorsten Stein; Stefan Sell

doi:10.3389/fbioe.2020.00009

A Machine Learning and Wearable Sensor Based Approach to Estimate External Knee Flexion and Adduction Moments During Various Locomotion Tasks

Front Bioeng Biotechnol. 2020 Jan 24:8:9. doi: 10.3389/fbioe.2020.00009. eCollection 2020.

Authors

Bernd J Stetter¹, Frieder C Krafft¹, Steffen Ringhof^{1

2}, Thorsten Stein¹, Stefan Sell^{1

3}

Affiliations

¹ Institute of Sports and Sports Science, Karlsruhe Institute of Technology, Karlsruhe, Germany.
² Department of Sport and Sport Science, University of Freiburg, Freiburg, Germany.
³ Joint Center Black Forest, Hospital Neuenbuerg, Neuenbuerg, Germany.

Abstract

Joint moment measurements represent an objective biomechanical parameter of knee joint load in knee osteoarthritis (KOA). Wearable sensors in combination with machine learning techniques may provide solutions to develop assistive devices in KOA patients to improve disease treatment and to minimize risk of non-functional overreaching (e.g., pain). The purpose of this study was to develop an artificial neural network (ANN) that estimates external knee flexion moments (KFM) and external knee adduction moments (KAM) during various locomotion tasks, based on data obtained by two wearable sensors. Thirteen participants were instrumented with two inertial measurement units (IMUs) located on the right thigh and shank. Participants performed six different locomotion tasks consisting of linear motions and motions with a change of direction, while IMU signals as well as full body kinematics and ground reaction forces were synchronously recorded. KFM and KAM were determined using a full body biomechanical model. An ANN was trained to estimate the KFM and KAM time series using the IMU signals as input. Evaluation of the ANN was done using a leave-one-subject-out cross-validation. Concordance of the ANN-estimated KFM and reference data was categorized for five tasks (walking straight, 90° walking turn, moderate running, 90° running turn and 45° cutting maneuver) as strong (r ≥ 0.69, rRMSE ≤ 23.1) and as moderate for fast running (r = 0.65 ± 0.43, rRMSE = 25.5 ± 7.0%). For all locomotion tasks, KAM yielded a lower concordance in comparison to the KFM, ranging from weak (r ≤ 0.21, rRMSE ≥ 33.8%) in cutting and fast running to strong (r = 0.71 ± 0.26, rRMSE = 22.3 ± 8.3%) for walking straight. Smallest mean difference of classical discrete load metrics was seen for KFM impulse, 10.6 ± 47.0%. The results demonstrate the feasibility of using only two IMUs to estimate KFM and KAM to a limited extent. This methodological step facilitates further work that should aim to improve the estimation accuracy to provide valuable biofeedback systems for KOA patients. Greater accuracy of effective implementation could be achieved by a participant- or task-specific ANN modeling.

Keywords: accelerometers and gyroscopes; artificial neural networks; biofeedback; biomechanics; knee joint loading; knee osteoarthritis; reduced sensor set.