Development and application of a knee loading score for change-of-direction-specific movements

Sammanfattning

The anterior cruciate ligament (ACL) is the most frequently injured ligament in the knee, mainly due to its role as a primary joint stabiliser. In sports science, non-contact ACL injuries are commonly linked to change-of-direction (COD) manoeuvres, especially in multidirectional team sports. Given the long rehabilitation times and risk of long-term consequences, ACL injury prevention remains a major research focus. Effective prevention requires evaluation of athletes in natural sporting environments and tools to identify injury-prone movement techniques. This master’s thesis explores machine learning to predict knee loading from movement data captured by inertial measurement units (IMUs). A support vector machine was trained to classify high and low knee loading and externally tested on custom trials designed to elicit varying loads. The model classified peak knee abduction moment (pKAM), a proxy for knee loading, based on IMU-derived joint angles, angular velocities, and centre of mass velocity at initial contact. Ground truth labels were obtained using marker-based motion capture and inverse dynamics. To evaluate generalisability, an independent dataset of 45° and 135° COD trials was collected, featuring three movement conditions provoking varied knee loads through differences in speed, cut angle, and technique focusing on upper body posture. The model was calibrated using Platt scaling to produce probabilistic outputs for each condition. The model showed strong performance on internal data, with an AUC of 0.80, indicating accurate classification of high and low pKAM. In the external test, the custom dataset elicited statistically significant varying knee loading levels, particularly in the technique condition, aligning with expectations. The model did recognize differences in knee loading among unseen participants and tasks, suggesting potential to identify injury-prone movement patterns. Still, predictions were biased toward low loading, highlighting the need for more training data with higher variability in tasks. Predicted probabilistic outputs were skewed toward lower values (0-0.5) and had a limited spread across the 0–1 range, proving non-sufficient for clinical use. Nevertheless, these findings support the central aim of this thesis, that a biomechanically informed dataset could support the development of a model capable of distinguishing loading patterns, marking a promising step toward a practical, externally validated knee loading score.