Machine learning models can flag marathon runners' injury risk week by week, new study finds

A newly published study in the journal PM&R has used machine learning to predict week-by-week injury risk among runners training for the New York City Marathon, adding to a growing body of research aimed at moving injury prevention from generic advice toward individualised, data-driven guidance. The prospective observational study asked participants to complete a baseline survey followed by sixteen weekly interval surveys across their marathon training block, while also sharing GPS watch and smartphone-based running logs pulled largely from Strava.

The scale of the injury problem the researchers were addressing is striking. Across the training block, 38.4% of runners reported an injury, with a further 14.1% reporting one during the marathon itself, for an overall injury prevalence of 42.6%. Those figures sit consistent with prior research showing that somewhere between a third and a half of marathon trainees experience some form of injury before or during race day.

The location of injuries shifted depending on training phase. Foot, knee and hip injuries were the most common complaints during the build-up, while knee, thigh and foot injuries dominated during the marathon itself, suggesting that the demands placed on the body change meaningfully in the final, fatigued miles compared with steady-state training runs. Hamstring injuries had the single highest prevalence of any specific diagnosis, at 6.7% of the study population.

What distinguishes this research from earlier injury-prevalence surveys is its attempt to build predictive models rather than simply describe outcomes after the fact. By combining the weekly self-reported survey data with objective GPS and activity-log data, the researchers trained models intended to flag elevated injury risk in the coming week, potentially giving coaches and runners a tool to adjust training load before a niggle becomes a diagnosed injury rather than after.

The work follows a broader trend the site has tracked this year, from UC Davis's wearable accelerometer research on collegiate runners to meta-analyses on how carbon-plated shoes alter loading and metabolic cost. Together, these studies point toward a future in which marathon training plans increasingly draw on individual biomechanical and physiological data rather than one-size-fits-all mileage templates, even if translating research-grade prediction models into consumer running apps remains a work in progress.