Combining causal inference with deep learning to identify modifiable injury risk factors in elite team-sport athletes

Abstract Athletic injuries continue to disrupt competitive sport, eroding athlete welfare, on-field performance and career longevity. Conventional injury-prediction pipelines have leaned heavily on correlational signal and reactive triage, leaving the mechanistic basis of risk largely opaque. We build an analytical pipeline that pairs causal-inference reasoning with deep-representation learning so that two distinct objectives can be pursued under a single architecture. Our framework treats prediction and causal estimation as separable tasks rather than blurring them: a temporal-convolutional encoder with attention pooling drives discrimination, while a treatment-agnostic representation network with confounder balancing supports counterfactual estimation under the explicit assumptions of conditional ignorability, positivity and no interference. Inter-rater reliability of the expert-derived reference standard is reported (Fleiss’ kappa = 0.71), and learning curves, sensitivity to unmeasured confounding, calibration and cluster-level allocation are examined alongside the headline metrics. Model development follows TRIPOD + AI; because the prospective component was neither randomised nor registered, it is reported against the TREND statement for non-randomised evaluations. Drawing on longitudinal monitoring data from 342 elite soccer and basketball athletes, the framework recovered injury-relevant causal factors with 0.92 precision and 0.92 recall against an expert-consensus reference under five-fold athlete-stratified cross-validation. That pairing flatters the method, since the same panel drafted the candidate list. Re-benchmarked against an independent 18-factor literature-derived set and a semi-synthetic simulation with known ground truth, precision held at 0.85 while recall fell to 0.61 and 0.79 respectively; we treat this lower band, not the 0.92 pair, as the defensible operating range. Mean absolute error in causal-effect estimation was 56% lower than that of standard deep learning, discrimination remained robust (AUC-ROC > 0.82) and the calibration slope sat close to unity (0.97). The accompanying non-randomised, propensity-score-matched prospective comparison—89 intervention athletes against 87 controls, with 11 and 18 time-loss injuries over six months—is an exploratory pilot rather than a confirmatory effect estimate. Allocation was effectively at the institution level, so leave-one-cluster-out, cluster-level permutation and informative-prior Bayesian multilevel analyses accompany the Rosenbaum and E-value bounds. The point estimate is a 40% relative risk reduction, and 29 events cannot pin it down: the individual-level interval runs from a 19% increase to a 70% reduction ( p = 0.14), every cluster-aware interval is wider still, and the cluster-level permutation p of 0.05 is simply the smallest value six clusters can yield. What this paper offers is a methodological template for joining causal reasoning with predictive modelling in sports medicine, with the explicit caveat that any clinical claim will need a properly powered, multi-centre cluster-randomised trial.