From the College Athletics edition of September 20, 2026
USF study: longitudinal ML predicts sport-related concussion clearance with XGBoost accuracy 0.84
University of South Florida / Digital Health (SAGE) · College athletics / sports medicine return-to-play risk modeling · United States (USF Concussion Center clinic cohort)
- Publisher
- Digital Health (PMC full text)
- Original publication
- May 20, 2026
- Source retrieved
- 2026-09-21
What happened
Using 217 sports-related concussion patients (mean age 26.94) from the USF Concussion Center (2021–2025), Tran et al. train six ML classifiers to label normal (<30 days) versus prolonged (≥30 days) medical clearance. Adding Visit-2 longitudinal features improved accuracy in 66% of models; XGBoost achieved the highest accuracy (0.84, +5% over Visit 1). Specificity remained low (0.00–0.34) with 81.1% prolonged cases. VOR Vertical Headache and its change score were consistent prolonged-recovery predictors; treatment presence predicted normal recovery. Authors require external validation before clinical deployment.
Why it matters
Newly relevant athlete-support / sports-medicine evidence for collegiate athletic trainers considering ML on VOMS and related assessments already used in RTP—complements archived Claros post-concussion MSK risk modeling with a clearance-time outcome. Transferability is limited because the cohort is a specialty clinic sports-related sample, not an NCAA CARE multi-site labeled set.
Evidence and measured results
PMC full text of Digital Health article (doi:10.1177/20552076261450858): abstract metrics, methods (LOOCV; Visit 1 vs Visit 2 feature sets; class imbalance), Table performance figures, feature-frequency discussion, limitations, and explicit external-validation requirement; IRB STUDY003514; public code link.
Limitations and uncertainty
Clinic cohort not exclusively NCAA-labeled; mean age ~27; low specificity; no external validation; binary 30-day threshold; self-report bias; gender imbalance.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-21; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Sports-medicine staff struggle to anticipate which concussed athletes will need prolonged clearance timelines using single-visit impressions alone.
- Stakeholders
- Head athletic trainer
- Team physician
- Sports performance director
- Sport coaches
- Compliance/privacy officer
- Discovery questions
- Which VOMS/BESS/ImPACT fields are already mandated and stored longitudinally?
- Who may view cognitive or symptom scores today?
- What false-positive rate is tolerable given AT workload?
- Would recovery-risk tiers ever influence scholarship or roster decisions?
- Value hypothesis
- If local revalidation improves balanced accuracy beyond chart review alone, ATs could target vestibular/rehab pathways earlier using assessments they already collect—without claiming faster RTP.
- Potential engagement
- Six-to-eight-week retrospective validation on the school’s last two seasons of SRC cases, plus a draft clinical-use policy.
- Unsupported claims
- Guaranteed shorter clearance times
- That 0.84 accuracy generalizes without local validation
- NCAA multi-site representativeness of the USF clinic cohort
Pre-sales engineering
Role takeaway
- Fit
- Clinical decision-support scoring on existing concussion assessment pipelines for collegiate sports medicine.
- Architecture and integration
- Export visit-level VOMS subscores and related fields; compute difference features; score in a controlled environment; return risk tiers to AT workflows with audit logs.
- Prerequisites
- Longitudinal two-visit concussion records with clearance labels
- Standardized assessment timing
- Clinical governance approval
- Constraints
- Source study used LOOCV only
- Severe class imbalance and low specificity
- Clinic demographics may differ from campus athlete population
- Security
- Restrict cognitive/mental-health features; no secondary model-training rights for vendors without athlete-facing consent review.
- Proof of value
- Local balanced accuracy and specificity on held-out seasons
- Calibration within agreed error bands
- Clinician override rate reviewed monthly
Delivery
Role takeaway
- Work
- Historical validation, threshold setting with ATs, coach-access policy, and season-one monitoring dashboard.
- Dependencies
- Athletic training documentation quality
- IT access to assessment and clearance systems
- Team physician sign-off
- Ownership
- Sports medicine / head athletic trainer owns clinical use; AD sponsors policy.
- Skills and adoption
- ATs learn score interpretation and specificity limits; coaches learn they do not receive raw cognitive detail without policy.
- Governance checkpoints
- IRB/clinical review before prospective use
- Written prohibition on scholarship use
- Quarterly drift review
- Acceptance criteria
- Documented local validation report
- Signed access matrix
- No unauthorized roster uses in pilot
- Risks
- Over-trust in 0.84 accuracy headline despite low specificity
- Leak of symptom/cognitive scores to coaching staff
- Distribution shift versus clinic training data
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Score VOMS/BESS/ImPACT-like exports in a clinician-controlled service with longitudinal difference features and treatment flags; do not expose raw cognitive scores to coaches without policy.
Governance
Who approves, reviews and stays accountable for outcomes?
Keep model output under sports-medicine ownership for clearance support only; prohibit scholarship/roster use; require local revalidation before any prospective alert.
Security and privacy
What data, permissions and controls need testing?
Concussion and mental-health (GAD-7/PHQ-9) features are highly sensitive; minimize secondary AI-training rights and apply role-based access even inside athletics.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
ATs need plain-language cards that emphasize specificity limits so junior staff are not flooded by prolonged-class alerts.
Procurement
What should contracts, pricing and exit terms secure?
Buy validation and integration around existing assessment vendors; do not procure a black-box clearance AI that ignores VOMS field definitions.
Operating model
Which teams own the service once it runs?
Keep medical clearance with sports medicine; models triage attention, they do not clear athletes.
Publication history
- 2026-09-20College Athletics · Issue 144 resources
Stable resource ID: usf-src-clearance-ml-2026