With the TCS New York City Marathon four days away, a study published in PM&R examines whether runners' own data can predict who will get hurt in training. Mark Fontana, Jamie Egbert and Brett Toresdahl followed 643 adults training for the 2022 New York City Marathon, about 53 per cent of them women with a mean age of around 43, using a baseline survey, 16 weekly surveys and shared Strava running logs.
Nearly half of the group, 307 of 643 runners or 48 per cent, reported at least one injury that forced them to modify or stop training. Across 9,002 runner-weeks of observations, 75 per cent were injury-free. Injuries tended to persist from week to week, recovery was more common than worsening, and most were gradual-onset overuse problems.
The researchers trained two machine learning models, using an Explainable Boosting Machine, to predict whether a runner would report a training-limiting injury in the following week. The first, applied to all runner-weeks, performed well, with an AUROC of 87.3 per cent. The second, restricted to runners not already modifying their training, performed poorly, with an AUROC of 67.3 per cent and an AUPRC of just 7.7 per cent from only 50 injury events in the test set.
The gap exposes the central finding: the strongest predictors were prior-week pain and prior-week injury status, so the first model was largely recognising injuries that were already under way. For uninjured runners, the useful signals included the acute-to-chronic workload ratio two weeks earlier, monthly moving time and the longest daily distance. Risk appeared to rise around week seven of training, and a workload ratio near 1.3 was associated with higher risk.
The authors concluded that the approach is feasible but not yet accurate enough to guide individual recommendations. They also flagged limitations, including self-reported injuries, no clinical diagnoses, a small number of injury events among initially uninjured runners and a sample that skewed female and technologically engaged. For marathon trainees, the practical message is that persistent pain is the clearest warning sign in the data, and no algorithm yet replaces paying attention to it.