Submission ID 131807
| Issue/Objective | As North American football expands globally, particularly into regions with emerging leagues and uneven health infrastructure, injury risk becomes a growing health equity concern. At the same time, the governance of machine learning technologies offers a unique mechanism to address upstream determinants of health equity. Current systems for football injury surveillance outside North America remain reactive and individual-focused, with even greater gaps at the amateur level. This study aims to develop a proactive, data-driven injury surveillance framework to support equitable expansion of this sport by identifying high-risk players and environments before injury occurs. |
|---|---|
| Methodology/Approach | Using web-scraped Canadian Football League play-by-play data, we developed a balanced Random Forest model to predict injury occurrence, applying a stratified 60/20/20 train-validation-test split with F1-based threshold tuning. To incorporate interaction context, we constructed a knowledge graph modeling players as nodes and on-field interactions as edges. Weighted degree centrality was used to quantify exposure, enabling integration of predictive modeling with network-based context. |
| Results | The model demonstrated moderate discriminative performance (ROC AUC = 0.66) in a highly imbalanced setting (injury prevalence ≈2.4%), achieving high sensitivity (0.76). Balanced accuracy (0.64) indicated reasonable class discrimination. Aligning with equitable AI governance, our fairness analysis showed minimal performance disparity across contexts. For intra-division matches, the sensitivity ratio was 1.04, and the specificity ratio was 1.18. Inverse patterns were observed for inter-division matches, supporting consistent application across environments. Feature importance highlighted that injury risk is driven by dynamic exposures, demonstrating how the system forces specific individual roles (receivers) into high-risk interactions. Weighted network centrality functioned as a decision-support layer, identifying high-priority plays where elevated predicted injury risk coincided with high-interaction environments involving highly central players. |
| Discussion/Conclusion | This study reframes injury risk as a function of exposure and interaction context. The Random Forest model represents a low-barrier, scalable method using routinely collected data, requiring only minimal additions such as standardized injury indicators. Combined with network-based insights, this framework enables a shift from reactive reporting to proactive surveillance that identifies both high-risk individuals and environments. In the context of sport and health, such approaches support more equitable safety governance by ensuring that risk monitoring scales alongside participation. |
| Presenters and Affiliations | Paul Mokrzycki McMaster University Paul Mokrzycki McMaster University Bruce Newbold McMaster University Grzegorz Chrobak Wrocław University of Life & Environmental Sciences Przemysław Banat Warsaw University of Technology Jakub Łach Wrocław University of Science and Technology Reeya Kothari McMaster University |
