Can artificial intelligence predict a running injury?

It would be helpful to receive a warning before an injury. The challenge is to distinguish a tool that identifies patterns in one database from another that actually helps people avoid problems in the real world.

A research published in npj Digital Medicine investigated this possibility by gathering information from different areas of sports life. The contribution is promising as a research method, but it is still far from an infallible individual prediction.

What the researchers have done

We analyzed 142 runners followed by 12 months, with 6181 weekly observations. The data included training, history, strength, biomechanics, nutrition, body composition and genetic factors.

The random forest model reached AUC close to 0.784. There was no validation in an independent external cohort. The authors recognize limitations of sample size, self-reported injuries and characteristics of the collection. The study did not test whether providing alerts to runners would reduce the number of injuries.

An AUC of 0.784 does not mean correctly predicting 78.4% of injuries

The AUC assesses discrimination: in simplified terms, the ability to assign a higher score to a positive case than to a negative one. It does not directly inform the proportion of correct diagnoses or the probability of a specific runner being injured.

To turn scores into alerts, it would be necessary to define a threshold. A more sensitive threshold can produce more false warnings; another more restrictive one can pass on cases. It is also important to check if the estimated probabilities correspond to what happens in the population.

These properties need to be demonstrated in the intended use. A global number of performance does not replace the evaluation of how a tool works for different people.

Predicting and preventing are different tasks

Even a reasonable forecast does not automatically determine the best action. If the system identifies high risk, what changes? Reduce volume? Change session? Forward for evaluation? Which one helps and which one can unnecessarily interrupt a preparation?

These questions require implementation studies and, ideally, comparison of results between people who received an intervention and people who did not receive it.

Another care is causal. A variable that contributes to a prediction does not turn into a cause of injury or target that should be modified. The algorithm can use a useful association without revealing the responsible mechanism.

Editorial example: how to evaluate the promise of an application

If a tool promises to anticipate injuries, it is worth asking in whom it was tested, how it defines an injury and whether it was evaluated outside the base on which it was built. It is also necessary to know how many alerts are false and what happened when users followed the recommendations.

In the study, thousands of records do not represent thousands of independent people. Repeated observations enrich temporal analysis, but remain linked to a group of 142 corridors. This detail matters when interpreting the possibility of generalization.

What can the runner learn now?

The research reinforces the value of organizing information of different dimensions, without reducing the discussion to a single load number. Its main product is a base and structure to investigate prediction.

The step between recognizing patterns and guiding decisions needs to be validated. For the Enduran-se reader, this is a useful rule for evaluating promises of technology: asking not only if the algorithm predicts, but whether its use improves results and under what conditions.

Scientific reference

Multidisciplinary prediction of running-related injuries using machine learning. npj Digital Medicine, 2026. DOI: 10.1038/s41746-026-02413-y.

Editorial text in our own words. Examples identified as editorials are illustrative and do not represent additional participants or interventions of the research.

Cada um no seu ritmo. Juntos no caminho.