Machine learning predicts concussion recovery for return-to-play

The hard part of return-to-play after concussion is not reciting the graduated progression. It is recognizing, early enough, the athlete who will not move through that progression cleanly. In clinic and on the sideline, AI in concussion diagnosis and return to play protocols becomes useful only if it answers that practical question: who is likely to have prolonged recovery before the usual week-by-week picture makes it obvious?

A 2025 BMJ Open Sport & Exercise Medicine study is attention-grabbing because it stays close to that problem. Investigators used a random forest model to predict whether athletes with mild traumatic brain injury would miss more than five games, combining demographics, injury history, MRI white matter findings, and SCAT5 data. In a single UK center cohort of 375 athletes, the model reported 94.6% accuracy, 100% sensitivity, 93.8% specificity, and an AUC of 0.963 for that outcome.[1]

The study’s most memorable detail is not the accuracy line. It is a case in which the treating physician expected a good prognosis, while the model predicted the athlete would miss more than 20 games. The athlete did miss more than 20 games.[1] That case is both impressive and uncomfortable. It shows why clinicians are interested in prediction models, especially when early symptoms and a hopeful examination do not tell the whole story. It does not show that an algorithm can clear, hold, or manage an athlete by itself.

What the model actually predicted

Return-to-play is often discussed as if the core decision is a single yes-or-no clearance moment. The evidence here is narrower. The BMJ model predicted games missed after mild traumatic brain injury, using more information than a sideline assessment or symptom checklist alone. Its strongest practical value is not replacing the graduated return-to-sport process, but flagging athletes whose recovery may be longer than early clinical impressions suggest.

That distinction matters. A model that predicts more than five games missed is not the same as a model that proves an athlete is biologically recovered on a specific day. Missing games reflects recovery, medical decision-making, sport schedules, roster context, and institutional practice. It is still a meaningful clinical outcome, especially for teams trying to plan care, but it should not be mistaken for a direct brain-recovery biomarker.

The input set is also clinically revealing. SCAT5 data were included, but SCAT5 alone was weaker than the fuller model, with SCAT5-only performance reported at AUC 0.697. Injury history and MRI white matter findings carried stronger predictive signal, with injury history reported at AUC 0.963 as a single predictor in the BMJ study context.[1] That pattern fits what many concussion clinicians already suspect: symptoms and early exam findings are necessary, but they rarely capture the whole recovery trajectory.

Why random forests fit this kind of concussion question

Concussion recovery is rarely explained by one clean variable. Two athletes can report similar symptom scores and still diverge over the next month. Prior injury, sex and age distributions, acute presentation, imaging findings, and the way symptoms evolve can interact in ways that are difficult to reduce to a linear rule at the bedside.

Random forest models are suited to mixed clinical datasets because they can handle nonlinear relationships and interactions among variables without forcing a single assumed pathway. That does not make them inherently clinically valid. It explains why they may perform well when the question is prognostic and the inputs include both structured clinical measures and imaging-derived findings.

For a care team, the attraction is obvious. An athletic trainer may see the athlete every day and notice the messy recovery details. A physician may be asked to make a clearance decision under time pressure. A model that elevates risk based on patterns across prior injury, SCAT5 results, and MRI findings could make the conversation less dependent on optimism, roster pressure, or the athlete’s desire to return.

The larger CARE dataset supports the signal, but keeps the claim modest

The second 2025 study widens the lens. Using data from 971 college athletes in the FITBIR/CARE Consortium context, researchers trained a random forest model to classify typical recovery, defined as 28 days or less, versus prolonged recovery, defined as more than 28 days. The model reported 89% accuracy and an AUC of 0.85.[2]

The important detail is that MRI white matter hyperintensities were the strongest single predictor, with AUC 0.9.[2] That is not a small point. It means the higher-performing models in this evidence base are not simply applying machine learning to symptom forms. They are drawing strength from multidimensional data, including imaging features that many athletic settings do not routinely obtain after every concussion.

The same study also makes the ceiling of the evidence visible. The authors described the work as proof-of-concept, and the reported R² range was 0.17–0.21.[2] In plain clinical language, a useful classifier may still leave much of the recovery variability unexplained. That is not a failure; it is a warning against turning an encouraging research model into a clearance authority.

High accuracy is not the same as clinical replacement

The graduated return-to-play protocol remains the familiar clinical scaffold because it does something prediction cannot do by itself: it observes how the athlete responds to increasing cognitive and physical load. A risk estimate made near the front end of recovery cannot substitute for symptom monitoring, exertional progression, neurologic reassessment, and the clinician’s responsibility for clearance.

Accuracy also depends on the environment that produced it. A model developed in one center can learn patterns tied to that center’s scanner, referral practices, athlete mix, documentation habits, and return-to-play culture. A model trained in college athletes may not behave the same way in younger athletes, professional players, combat sports, or programs with different access to imaging. The question is not whether the published numbers are interesting. They are. The question is whether they survive contact with another clinic.

This is where external multi-center validation becomes more than a methodological nicety. A sideline decision has consequences for the athlete who returns too soon, for the athletic trainer who sees subtle deterioration during practice, and for the physician whose signature appears on the clearance form. Before an ML estimate enters that chain of responsibility, clinicians need to know how often it misses prolonged recovery, in whom, and under what data conditions.

There is also a practical burden hidden inside the best-performing inputs. MRI white matter findings are clinically attractive when they improve prognostic signal, but routine access to MRI varies widely. Programs without imaging access should not assume they can reproduce the same performance with symptom data alone. The evidence points toward richer data improving prediction; it does not show that every setting can collect that data without cost, delay, or selection bias.

Secondary injury risk is related, not proof of RTP readiness

A separate 2025 University of Delaware model is useful for the adjacent question of what happens after concussion, but it should not be folded into the recovery-prediction evidence as if it answered the same question. That study used a random forest model in 194 athletes and reported 95% accuracy for predicting musculoskeletal injury after concussion, with the caveat of a single-university setting and about 35% missing data.[3]

That finding matters because return-to-play is not only about symptom resolution. Athletes who have recovered enough to progress may still carry altered neuromotor control, reaction-time changes, conditioning gaps, or risk patterns that affect subsequent injury. But a model predicting post-concussion musculoskeletal injury is not direct evidence that ML can clear an athlete from concussion. It supports a broader point: machine learning may help identify risk around the return process, not replace the process itself.

Where this leaves clinical use in Q3 2026

The best current reading is neither dismissal nor adoption. The 2025 random forest studies show unusually strong prognostic performance for prolonged concussion recovery, especially when models include MRI white matter findings and injury history. They also show why a single SCAT5-centered view of recovery is too thin for individualized prediction.

They do not show that machine learning has been integrated into formal return-to-play guidelines, and they do not remove the need for a graduated progression. The evidence supports risk stratification under research conditions: identifying athletes who may need closer follow-up, slower progression, or a more guarded prognosis. It does not support handing clearance decisions to a model.

For now, the most defensible category is clinical application in development. Machine learning can sharpen the early recovery conversation, particularly for prolonged recovery risk, but the models still need external multi-center validation across sports, scanners, and athlete populations. In Q3 2026, they are complementary prognostic tools waiting to prove they can travel—not replacements for graduated return-to-play protocols.

References

1. BMJ Open Sport & Exercise Medicine 2025 machine learning study, BMJ Open Sport & Exercise Medicine, 2025.
2. FITBIR/CARE Consortium study, PubMed, 2025.
3. University of Delaware post-concussion musculoskeletal injury risk model, Sports Medicine, 2025.