Personalized Machine Learning Approach to Estimating Knee Kinematics Using Only Shank-Mounted IMU

Abstract

Knee kinematics is a valuable measure of knee joint function. However, collecting that data outside the clinic is difficult, especially with a limited number of wearable sensors and when you only use an ankle-mounted inertial measurement unit to estimate knee kinematics. Due to the cyclic nature of gait, it is possible to use machine learning to extract joint angles from only ankle-mounted sensors. This study aimed to use time-series feature extraction and a random forest regressor to generate a person-specific surrogate model for estimating knee joint flexion angles from a single-mounted inertial measurement unit above the ankle. Optical motion capture and inertial data from 10 healthy participants walking on a treadmill were collected to create 10 personalised surrogate models for estimating right knee flexion angles during gait. An additional 10 models were created for a leave-one-out analysis to test the generalisability of the models. Temporal cross-validation of the personalised models and a leave-one-out analysis was performed on the selected feature set. The personalised models achieved an average root mean square error of 2.45 ± 0.65 degrees (R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.98) compared to a gold-standard optical motion capture. The generalised models achieved an average root mean square error of 6.77 ± 3.38 degrees (R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.83) in the leave-one-out analysis. Time series feature-based personalised surrogate models could be used to accurately estimate knee kinematics by using a single ankle-mounted sensor. However, more data is required to train a generalised model using the presented method.