Cross-sport transportability of workload-based machine learning for injury risk: a two-cohort study

Machine-learning models for sports injury risk are commonly evaluated within a single cohort, leaving uncertainty about transportability across sports and monitoring systems. We evaluated whether an athlete-relative workload model developed in competitive runners retained performance in elite women’s football. This secondary two-cohort study used public longitudinal data from competitive runners for model development and SoccerMon elite women’s football for locked external evaluation. Predictors were prior 7-day athlete-relative load mean, load variability, session frequency, perceived recovery, perceived exertion, and previous recorded injury history. All rolling features ended at t − 1. The primary model was prespecified L2-regularized logistic regression without class rebalancing. Associations used athlete-clustered standard errors; discrimination, precision–recall, Brier score, calibration, and decision curves were evaluated with athlete-cluster bootstrap confidence intervals. The runner cohort contributed 41,286 observations, 551 injury events, and 73 athletes; football contributed 18,614 observed risk days, 78 incident player-days, and 50 athletes. Load variability was associated with recorded injury in runners (odds ratio 1.21, 95% confidence interval 1.06–1.39) but not football (0.96, 0.74–1.24). Previous injury history was associated in both cohorts. In locked football evaluation, ROC-AUC was 0.768 (athlete-cluster 95% confidence interval 0.616–0.853) and precision–recall AUC was 0.0144 versus prevalence 0.00419. The Brier score was 0.00422, calibration slope 1.35, and observed-to-expected ratio 0.37. A history-only model had ROC-AUC 0.815, whereas a workload-only model had ROC-AUC 0.608. The locked model retained partial cross-sport ranking ability, but workload-instability effects did not replicate and raw probabilities were miscalibrated. The transported signal was driven mainly by previous injury history; therefore, moderate discrimination should be interpreted as relative ranking rather than accurate individual risk estimation. The model is not ready for individual clinical or training decisions; prospective multi-team evaluation with harmonized, clinician-adjudicated outcomes and local recalibration is required.