Biomechanical phenotypes of 90° change of direction in football players: Unsupervised machine learning in anterior cruciate ligament injury prevention
Francesco Della Villa, Alessandro Ghibellini, Matthew Buckthorpe, Luciano Bononi, Maurizio Gabbrielli, Alberto Grassi, Stefano Zaffagnini, Stefano Di PaoloAbstract
Purpose
Neuromuscular and biomechanical deficits contribute to anterior cruciate ligament (ACL) injury risk in football. Artificial intelligence (AI) could be adopted to analyse large datasets of complex multidirectional movements, aiding to unravel hidden biomechanical profiles. This study aimed to identify clusters of features describing the 90° change of direction (COD) task in football players—that is, biomechanical phenotypes —through unsupervised machine learning algorithms.
Methods
One thousand and two healthy football players (mean age: 16.3 years) from the ‘CutTheACL’ project performed a series of 90° COD tasks. Two‐dimensional (2D) kinematics and ground reaction forces (GRF) were extracted from three high‐speed cameras and one embedded force platform (VICON, AMTI). Unsupervised agglomerative clustering was performed to inspect the full dataset including kinematics, kinetics and 2D scoring system variables for each of the 6008 trials. The Silhouette coefficient was used to inspect the goodness of cluster differentiation. The Kruskal–Wallis and χ 2 tests were performed to identify statistical differences among the clusters ( p < 0.05).
Results
Four clusters with distinct biomechanical traits (phenotype) emerged (silhouette = 0.43). The clusters' interpretation was simplified according to GRFs and ACL‐injury risk related 2D kinematics: Cluster 0 ‘low forces, poor movement control’; Cluster 1 ‘low forces, acceptable movement control’; Cluster 2 ‘high forces, poor movement control’; Cluster 3 ‘high forces, acceptable movement control’. Statistically significant differences emerged among all clusters ( p < 0.001).
Conclusion
Four biomechanical COD phenotypes with distinct neuromuscular control and force profiles were identified through unsupervised clustering. Such an approach might offer a foundation to inspect young football players' motion and target primary ACL injury prevention.
Level of Evidence
Level IV, cohort study.