Two machine-learning injury studies. The one that predicted forward got 77 false alarms.

Two studies published within days of each other apply the same family of methods to athlete injury and reach opposite verdicts. The difference is not in the algorithms. It is in the direction of time.

Looking backwards works

The first analysed inertial measurement unit data from 14 NCAA Division I football players enrolled in the Hamstring Injury (HAMIR) study, comparing on-field sprinting at pre-injury baseline against sprinting at return to sport after a hamstring strain [s1].

Statistical parametric mapping — a technique that tests for differences across the whole stride rather than at selected time points — found the changes were not confined to the injured leg. In exploratory analyses, the healthy limb showed the most pronounced differences, in horizontal thigh and shank acceleration during initial swing at return to sport [s1]. In the injured limb, horizontal shank acceleration was significantly reduced for 55% of initial swing, and the pelvis showed the greatest vertical difference, at 20% across the stride [s1].

Eight supervised machine learning models were then trained to classify whether a stride came from the baseline or return-to-sport condition. After data augmentation, XGBoost reached a mean AUC of 0.75 and Random Forest 0.79 [s1]. Interpretability analysis identified thigh and pelvis segment features as the most influential variables [s1].

That is a real and useful result, and it is worth being precise about what it is. The models are classifying a condition that has already occurred, in data that has already been collected, from athletes who were already cleared to return. The finding is that a cleared athlete does not sprint the way they used to — including on the uninjured side — and that the difference is large enough for a model to detect [s1].

Looking forwards does not

The second study asked the question clinicians actually want answered: who is about to get hurt.

Investigators assembled three seasons of data (2021/22 to 2023/24) on 129 professional male rugby union players from the four Irish provincial teams — Connacht, Leinster, Munster and Ulster — and the Ireland men's national team [s2]. Inputs were external workload from GNSS units, self-reported wellness scores, and musculoskeletal screening measures. The outcome was a non-contact soft-tissue lower limb injury, logged by team medical staff in a centralised athlete management system [s2].

They fitted logistic regression, a support vector machine, Random Forest, XGBoost and CatBoost. The best performer, CatBoost, achieved a ROC AUC of 0.66, an area under the precision-recall curve of 0.008 against a prevalence of 0.0039, and a precision of 0.013 [s2].

The authors translate that precision figure themselves, and it is the most quotable number in either paper: one correct prediction for every 77 false predictions [s2]. Their conclusion is that predictive performance was poor across all models, and that the low precision scores highlight each model's inability to identify impending injuries within the modelling framework analysed [s2].

Why the gap is structural, not fixable by a better model

The rugby study's own numbers explain the problem. Injury prevalence in the dataset was 0.0039 — roughly four events per thousand player-observations [s2]. When the thing you are predicting is that rare, even a model with respectable discrimination produces overwhelmingly more false positives than true ones, because the false positives are drawn from a vastly larger pool. A ROC AUC of 0.66 sounds mediocre; the precision-recall figures are what show the practical consequence.

This is the same arithmetic that makes screening tests for rare conditions difficult, and no choice of gradient-boosting library changes it. What changes it is either a much stronger signal in the inputs, or a decision framework that accepts a 1-in-78 hit rate because acting on a false positive is cheap.

What each is good for

The hamstring paper is not a prediction tool and does not claim to be. Its authors position the work as supporting return-to-sport decisions — using interpretable IMU-based models to characterise whether an athlete's mechanics have returned to their own baseline [s1]. That is a comparison against a personal reference point, in a population where the event of interest has already happened, and it is a much easier problem than forecasting.

The limits are real on both sides. Fourteen athletes is a very small sample, the hamstring analysis was a secondary analysis of existing data, and the authors describe their statistical parametric mapping results as exploratory [s1]. The rugby study is larger and multi-team, but it is one sport, one country, one definition of injury, and one modelling framework — which the authors are careful to specify when stating that the approaches failed within the framework they analysed [s2].

What to watch is whether anyone publishes a forward-looking model with the precision-recall numbers stated as plainly as this rugby group did. The field has produced many injury prediction papers reporting AUC alone. On a base rate of four per thousand, AUC alone is not enough information to judge whether a tool would be usable, and this pair of papers makes the reason unusually concrete.

Sources

• [s1] Sprint Mechanics Differ Between Pre-injury and Return-to-Sport After Hamstring Strain Injury: A Comparative Statistical Analysis with Machine Learning. Medicine & Science in Sports & Exercise, 28 August 2026. https://doi.org/10.1249/MSS.0000000000004121
• [s2] Machine Learning Model Development and Evaluation for Non-contact Lower Limb Injury Risk Prediction in Elite Male Rugby Union: A Multi-team, Multi-season Analysis. Sports Medicine - Open, 26 August 2026. https://doi.org/10.1186/s40798-026-01060-7