Evaluation of machine learning methods to predict knee loading from the movement of body segments

Abstract

Abnormal joint moments during gait are validated predictors of knee pain in osteoarthritis. Calculation of moments necessitates measurement of forces and moment arms about joints during walking. Dynamically changing moment arms can be calculated from motion trackers either optically or with wireless inertia sensing units, but the measurement of forces is more problematic. Either the patient has to walk over a force platform or a force sensing device has to be built into the sole of the shoes. One possible means of registering abnormal joint moments without the restrictions due to force measurements is to predict moments from the movement of body segments using advanced machine learning techniques. To test the viability of this approach, we aimed to predict the frontal plane internal knee abduction moment form 3D Euler angles of the ankle, knee, hip and pelvis during a single gait cycle of 31 patients with alkaptonuria. Four machine-learning algorithms were used in our experiment to predict moments namely: Decision Tree, Random Forest, Linear Regression and Multilayer Perceptron neural network. Based on performance measures of prediction (R2, root mean squared error and area under the recall curve), the random forest algorithm performed best but this was also the slowest by a factor of 10. Considering both performance and speed, the Multilayer Perceptron neural network method was superior with R2, root mean square of error, area under the recall curve and required training time of 0.8616, 0.0743, 0.874 and 730 ms, respectively.