Advanced Machine Learning Techniques for Explainable Detection of Knee Injuries in Runners

Abstract

Running is a widely practiced recreational activity, but it is associated with high incidence of knee injuries, the most frequent being the Patellofemoral Pain Syndrome (PFPS) and Iliotibial Band Syndrome (ITBS). This work aims to enhance the detection of the running pattern associated with these injuries. We analyze a publicly available database of 1798 healthy and injured runners, with treadmill running motion records. Our study focuses on the stance phase of the running gait, analyzing time series of joint and segment angles estimated from motion records, as well as point values. Three classification problems are evaluated: identifying knee-injured runners vs. healthy ones, detecting PFPS, and identifying ITBS. The performance of eight classifiers were evaluated including Support Vector Machines, Random Forests, and Convolutional Neural Networks (CNN), among others. We evaluate models' performance by reporting accuracy, precision, recall, F1-score metrics and saliency maps for explainability. CNNs achieve the highest accuracy for detecting the PFPS pattern (81.59 ± 2.15%), when using time series data, and the ITBS pattern (76.19 ± 3.17%) with the combination with point values. The results emphasize analyzing full time series over point values. This research proves the effectiveness of advanced Machine Learning techniques in detecting common running-related knee injury patterns.