Predicting medial knee contact force from kinematic data using a neural network

Gait modifications, such as altering foot progression angle (FPA), have potential as a non-invasive treatment option for patients with knee osteoarthritis (OA) by reducing peak knee joint loading, thus reducing pain and improving knee function. However, estimating the effectiveness of these interventions requires a gait laboratory and expertise in musculoskeletal simulation to compute joint contact forces. We propose a machine learning approach to leverage portable motion capture systems by predicting peak medial knee contact force from kinematic input data. We trained a feed-forward neural network using data from 68 individuals with knee OA walking with five different FPAs. We predicted the early stance peak of medial contact force with 11.4% ± 8.1% mean absolute error and the late stance peak with 9.0% ± 6.9% error. When given kinematic data from five different FPA conditions, our model was able to assign the FPA modification to maximally reduce contact force in 6 out of 7 test subjects.