Explainable Machine-Learning based Detection of Knee Injuries in Runners

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.

💡 Research Summary

This paper investigates the use of optical motion‑capture data combined with explainable machine‑learning techniques to detect knee‑injury‑related gait patterns in recreational runners. The authors exploit a publicly available treadmill dataset (Ferber et al., 2024) that contains 1 798 participants, from which 839 stance‑phase recordings are selected (576 healthy, 137 with patellofemoral pain syndrome (PFPS), 126 with iliotibial‑band syndrome (ITBS)). For each recording, three‑dimensional marker trajectories of the feet, shanks, thighs and pelvis are processed to compute Cardan joint angles for ankles, knees, hips and the pelvis. A principal‑component‑analysis‑based algorithm identifies touchdown and toe‑off events, allowing the stance phase to be isolated and resampled to 101 equally spaced points via cubic interpolation.

Two families of features are constructed: (1) full‑length time‑series of the twelve joint/segment angles (mean, upper‑ and lower‑envelope) and (2) a set of discrete spatio‑temporal and kinetic descriptors (step width, stride rate, peak knee flexion, hip adduction, ankle eversion, power in low‑ (0‑1 Hz), medium‑ (1‑3 Hz) and high‑ (3‑99 Hz) frequency bands, etc.). The authors evaluate three binary classification tasks—healthy vs. any injury, healthy vs. PFPS, and healthy vs. ITBS—using both classical algorithms (K‑Nearest Neighbors, Gaussian Process Classifier, Decision Tree) and deep learning models (1‑D Convolutional Neural Network and Long Short‑Term Memory network). Hyper‑parameter optimisation is performed via grid search (classical models) and Bayesian optimisation (deep models). Performance is assessed with accuracy, precision, recall and F1‑score, employing five‑fold cross‑validation.

Results show that hybrid inputs (time‑series + point descriptors) consistently outperform either representation alone, yielding a 5‑7 percentage‑point gain in accuracy across all models. Among the classifiers, the CNN achieves the highest scores: 77.9 % accuracy for PFPS, 73.8 % for ITBS, and 71.43 % for the combined injury class. The LSTM, while capturing temporal dependencies, lags behind the CNN by 2‑3 % due to over‑fitting and longer training times. Classical methods reach at most ~65 % accuracy, confirming the advantage of deep architectures for this high‑dimensional, temporally rich data.

Explainability is addressed through four techniques: Shapley values, Partial Dependence Plots (PDP), saliency maps and Grad‑CAM. Shapley analysis identifies knee‑flexion peak, hip adduction/abduction angles, and ankle eversion as the most influential features. PDP visualises monotonic relationships, e.g., higher knee‑flexion peaks increase the probability of PFPS. Saliency and Grad‑CAM applied to the CNN highlight the early‑stance and mid‑stance windows, especially the knee and hip angle trajectories, as critical for the model’s decision. These findings align with biomechanical literature that links PFPS to excessive frontal‑plane knee loading and ITBS to altered hip mechanics.

The study’s limitations include reliance on treadmill data (limiting ecological validity), class imbalance (particularly fewer ITBS samples), exclusion of the swing phase, and lack of external validation on an independent cohort. Future work is suggested to integrate wearable inertial sensors and video‑based pose estimation for multimodal input, to develop real‑time feedback systems for injury prevention, and to test the models on larger, multi‑site datasets to confirm generalisability.

Overall, the paper demonstrates that high‑resolution motion‑capture combined with explainable deep learning can reliably discriminate knee‑injury gait patterns, and that interpretability tools can uncover clinically meaningful biomechanical markers, paving the way for data‑driven injury‑prevention strategies in sports medicine.