A watch can record distance, pace, heart rate, sleep and training load. What it still cannot do reliably is tell a runner who feels well this week that they will be injured next week. That distinction matters whenever an app turns incomplete data into a “risk” label.
A 2026 prospective study of 643 runners training for the New York City Marathon combined Strava activity data with weekly surveys over 16 weeks. Its useful finding was not that an algorithm had discovered a fail-safe warning: when the model was applied only to weeks in which a runner had not yet reported an injury requiring a training change, its ability to identify a new injury was low. The strongest signal was largely something far less technological: the runner was already modifying training because of a problem.
What the study examined, and why it matters
The study, published in PM&R, followed 643 adults preparing for the 2022 New York City Marathon. Participants shared Strava activities and completed a baseline survey plus 16 weekly surveys. The outcome was a self-reported injury that required a training modification.
- 307 of the 643 runners, or 48%, reported at least one injury that changed their training during the analysis period.
- The dataset contained 9,002 runner-weeks; no injury was reported in 75% of them.
- The model including all weeks looked strong on a broad discrimination measure, but much of that signal reflected the fact that injury status in the previous week often continued.
- When the researchers looked only at weeks that had previously been injury-free, performance dropped: the precision-recall area under the curve was 0.08. That is the more relevant test for the question, “Can it warn me before a new problem happens?”
That does not make training data worthless. It defines its role. A model may describe patterns in people who are already cutting back or adapting their workload; that does not make it a tool that can diagnose, independently prevent or reliably forecast a new injury for every runner.
Why a good-looking number can mislead
When a problem is relatively uncommon, a global score can look impressive while still being of little practical help in locating the few runners who move from “uninjured” to “injured” next week. That is why the study reports both the ROC area and the precision-recall area. The latter focuses on positive cases and was weak in the previously uninjured subgroup.
“Injury” was not a medical diagnosis or a scan finding. It was a runner reporting that they had to modify their preparation. That is a reasonable real-world outcome, but it does not allow an app to identify tendinopathy, bone stress injury or any other specific condition.
The limitations are not a footnote: they are part of the answer
- The sample consisted of adults in one specific marathon build-up, not every type of runner.
- Activity data do not necessarily contain pain, work stress, surface changes, technique, energy availability, clinical history or the quality of each session.
- Injury status and training modification were self-reported.
- An algorithm that works in one dataset should not be assumed valid for another club, country, watch or race distance without external validation.
- An association between one metric and injury does not prove that changing that metric alone removes risk.
The message fits a 2026 systematic review of knee injuries in endurance runners: prior injury, training load or volume and several biomechanical factors repeatedly appeared, but not as one switch that explains every case. Another 2026 study of road runners likewise describes injury as multifactorial, rather than the product of one watch metric.
What Strava, a watch and a training diary can genuinely add
Data become more valuable when they help you ask better questions, rather than obey an automatic label. A simple record can show whether a niggle began after adding hills, changing shoes, returning too quickly after a race, sleeping worse or raising volume and intensity at the same time.
- Compare the current week with your own history, not with a universal “safe-load” number.
- Record pain, morning stiffness, gait changes, fatigue, illness, sleep and stress alongside kilometres.
- Separate a planned hard session from one that felt unusually difficult.
- Look for trends over several weeks instead of reacting to one poor night of sleep or one unusual heart-rate reading.
- Use the history to have a more specific conversation with a coach, physiotherapist or doctor when needed.
A measured way to respond to an alert
1. Check whether anything has actually changed
Ask what changed in the past 7 to 14 days: volume, intensity, elevation, shoes, surface, strength work, racing, sleep, nutrition or stress. A watch alert can be the beginning of that review.
2. Put symptoms and function first
An app cannot judge whether pain worsens while you run, changes your gait or persists afterwards. Symptoms, their trajectory and your ability to walk, climb stairs or complete normal activities matter more than a readiness score.
3. Adjust the plan, not only the number
When there is a meaningful niggle or a clear build-up of fatigue, the sensible decision may be to reduce, replace, move or cancel a session. There is no single percentage reduction that fits everyone. The aim is to stop an early signal becoming a string of forced workouts.
4. Seek assessment before looking for more data
Focal bone pain, inability to bear weight or run normally, major swelling, night pain, fever, numbness, weakness, chest pain or unexpected shortness of breath need medical attention. No metric dashboard can replace that assessment.
What not to do
- Do not use a risk prediction to ignore pain because the watch shows green.
- Do not stop running automatically because an app spotted a variation without symptoms or context.
- Do not attribute an injury to one session, one shoe or one metric in hindsight.
- Do not turn another runner’s history into your personal load threshold.
- Do not try to make up every missed kilometre at once after a disrupted week.
So are wearables useless for injury prevention?
They are a very convenient training log, not an oracle. The more complete the context around the data, the more useful it can be for spotting changes and discussing decisions. But injury prevention still depends on a combination of tolerable workload, recovery, history, general health, technique, environment and individual response.
The study of 643 marathon trainees offers a needed correction to enthusiasm around artificial intelligence in sport: predicting that an injury is already changing training is not the same as forecasting a new injury. The better question after an alert is not “What does the algorithm order?” but “What changed, what am I feeling and what information is missing to make a good decision?”
Frequently asked questions
Can a running watch detect an injury?
No. It can record changes in load, pace, heart rate or sleep, but it cannot diagnose an injury. Metrics can provide context; they cannot confirm which tissue is affected or what treatment is appropriate.
Does a mileage increase always cause an injury?
Not always. Load is an important part of the picture, but tolerance also depends on history, intensity, recovery, health, surface, strength and many other factors. The 2026 review found associations, not a universal rule for every runner.
Which metric should I watch most closely?
There is no single winning metric. For most runners, combining training load with symptoms, perceived effort, sleep and routine changes is more useful than chasing one isolated number.
Sources
- Fontana, Egbert and Toresdahl. Predicting self-reported injury status during New York City Marathon training, PM&R (2026).
- Senthil et al. Systematic review of knee-injury risk factors in runners, Orthopaedic Journal of Sports Medicine (2026).
- Garcia et al. Epidemiology and associated factors in road-running injuries, Research in Sports Medicine (2026).