Explainable Machine-Learning based Detection of Knee Injuries in Runners

TL;DR

This paper tackles knee-injury detection in runners (PFPS and ITBS) by leveraging optical motion capture to extract full stance-phase time-series of joint angles and kinetic descriptors, combined with a broad set of classical ML and DL classifiers. CNNs and LSTMs outperform traditional methods, achieving up to about 77.9% accuracy for PFPS and 73.8% for ITBS using time-series inputs, with explainability tools (Shapley values, Saliency maps, Grad-CAM) revealing biomechanical patterns across stance phases. The work demonstrates that temporal dynamics carry richer injury signals than point-based features, and shows how explainability can render DL decisions clinically interpretable, promoting potential for personalized injury prevention and decision support in athletic populations.

Abstract

Running is a widely practiced activity but shows a high incidence of knee injuries, especially Patellofemoral Pain Syndrome (PFPS) and Iliotibial Band Syndrome (ITBS). Identifying gait patterns linked to these injuries can improve clinical decision-making, which requires precise systems capable of capturing and analyzing temporal kinematic data. This study uses optical motion capture systems to enhance detection of injury-related running patterns. We analyze a public dataset of 839 treadmill recordings from healthy and injured runners to evaluate how effectively these systems capture dynamic parameters relevant to injury classification. The focus is on the stance phase, using joint and segment angle time series and discrete point values. Three classification tasks are addressed: healthy vs. injured, healthy vs. PFPS, and healthy vs. ITBS. We examine different feature spaces, from traditional point-based metrics to full stance-phase time series and hybrid representations. Multiple models are tested, including classical algorithms (K-Nearest Neighbors, Gaussian Processes, Decision Trees) and deep learning architectures (CNNs, LSTMs). Performance is evaluated with accuracy, precision, recall, and F1-score. Explainability tools such as Shapley values, saliency maps, and Grad-CAM are used to interpret model behavior. Results show that combining time series with point values substantially improves detection. Deep learning models outperform classical ones, with CNNs achieving the highest accuracy: 77.9% for PFPS, 73.8% for ITBS, and 71.43% for the combined injury class. These findings highlight the potential of motion capture systems coupled with advanced machine learning to identify knee injury-related running patterns.

Figures (7)

• Figure 1: Methodology flowchart from the joint and segment angles in the database, including the time series in the stance phase extraction and running spatio-temporal parameters calculation, until the injury pattern identification with the ML and DL models, along with their explainability analysis.
• Figure 2: Cumulated normalized SHAP for the most relevant features in the identification of PFPS+ITBS with SVML.
• Figure 3: Cumulated normalized SHAP for the most relevant features in the identification of PFPS with SVML.
• Figure 4: Cumulated normalized SHAP for detection of ITBS with SVML.
• Figure 5: CNN explainability for PFPS+ITBS identification with temporal features from three methods: (a) depicts the saliency map, (b) corresponds to Grad-CAM and (c) is the shapley analysis result.