Profiling Risk Through Structure: Dimensionality Reduction of Drop Vertical Jump Biomechanics and Clustering for Targeted Neuromuscular Intervention

Authors:

Abstract
Machine learning has the potential to reveal latent biomechanical profiles that link athlete physiology, injury predisposition, and responses to interventions to reducing injury risk. This study developed a manifold-learning-based clustering approach to stratify athletes into biomechanically distinct subgroups. We evaluated combinations of four dimensionality-reduction techniques (t-SNE, KPCA, Diffusion Maps, Autoencoders) with three clustering algorithms (GMM, Agglomerative, DBSCAN) on biomechanical data from 312 female athletes (soccer, volleyball, or basketball) during a drop vertical jump (DVJ), selecting the optimal pipeline. All participants received a neuromuscular training (NMT) intervention aimed at reducing anterior cruciate ligament injury risk and were additionally randomized to receive either augmented NMT (aNMT) or sham NMT (sNMT), respectively involving accurate or uninformative visual feedback about training exercise performance. The optimal pipeline (t-SNE with GMM) identified three biomechanically distinct clusters that aligned with sport type, injury incidence, and intervention allocation. Profiles included a moderate-risk group, an injury-resilient group, and a high-risk, injury-prone group with elevated sprains, concussions, and non-preventable injuries, demonstrating that athletes can be grouped based on latent biomechanical structure using this approach. Cluster validity was confirmed by silhouette scores (0.49) and statistical testing at 95% confidence for separation of sport (0.00002), injury (0.036), and intervention (0.017). These results demonstrate that learned manifolds capture physiologically meaningful structure, offering a generalizable framework for stratifying athletes into biomechanically distinct subgroups associated with injury patterns, sport participation, and the distribution of interventions.