Author(s): LECKEY, C., VAN DYK, N., BERNDSEN, J., DOHERTY, C., LAWLOR, A., DELAHUNT, E., Institution: UNIVERSITY COLLEGE DUBLIN, Country: IRELAND, Abstract-ID: 929

INTRODUCTION:
The application of data analytics in professional sport has been facilitated by the proliferation of electronic performance and tracking systems, leading sporting organisations and researchers to seek data-driven solutions to a variety of problems. Despite the growth in ML applications for injury risk prediction in sport, a recent scoping review reported no relevant research in relation to rugby union. Since the publication of this review, two studies have sought to apply ML for injury prediction in rugby union, although both papers cite their analysis of a single team as a significant limitation of their research. Moreover, the aforementioned scoping review discussed the need for greater clinical relevance within sports injury risk prediction research, from definition and scope of ‘injury’, to model evaluation and dissemination. Accordingly, this research intends to bridge these gaps by synthesising diverse datasets from multiple professional rugby union teams. The aim of this research is to critically assess the practicality and clinical utility of predictive ML risk models for strictly defined non-contact time-loss lower limb injuries in elite-level male rugby union players.
METHODS:
Data for this research was provided by the Irish Rugby Football Union, and encompasses the four professional men’s provincial rugby teams in Ireland (Connacht Rugby, Leinster Rugby, Munster Rugby and Ulster Rugby), and the Ireland senior men’s national team, for three consecutive seasons (2021/22, 2022/23 and 2023/24). The data domains included: (i) GNSS units, worn during training sessions and matches, (ii) players self-perceived wellness measures reported on training and match days, and (iii) in-season musculoskeletal (MSK) tests. Machine learning next-day injury risk models were created for logistic regression, SVM, Random Forest, XGBoost, and CatBoost.
RESULTS:
CatBoost provided the highest statistical performance when applied to the unseen testing data, achieving values of ROC AUC (0.66), F1-Score (0.024), and AUPRC (0.008). XGBoost was the only other model to surpass a test ROC AUC of 0.6, achieving 0.61, 0.021, and 0.007 for ROC AUC, F1-Score and AUPRC respectively. Logistic regression was found to match the performance of Random Forest when applied to the test set, yet failed to outperform chance during cross-validation. Kernel-based SVM achieved the lowest ROC AUC during testing, with AURPC failing to outperform prevalence.
CONCLUSION:
This study addressed key methodological limitations in sports injury risk prediction research, including imprecise injury definitions and reliance on single-team cohorts. Despite these methodological refinements, predictive performance across the models was modest, with low precision scores highlighting each model’s inability to effectively identify impending injuries. The resulting high false-positive rates underscore the restricted clinical applicability of current machine learning approaches in real-world elite sport settings.