From the College Athletics edition of October 19, 2026
Self-archived paper claims XGBoost injury model (AUC 0.89) and 34% fewer injuries in an unnamed NCAA Division I cohort, but provenance and design limit trust
Lakshmibai National Institute of Physical Education (LNIPE), Guwahati, India (authors); deposited on Zenodo · Sports medicine and athlete support · Study claims a single unnamed NCAA Division I institution (United States); authors based in India
- Publisher
- Zenodo (Bhattacharjee, Bhattacharjee and Das)
- Original publication
- Zenodo record published October 17, 2026 (PDF states published October 15, 2026)
- Source retrieved
- 2026-10-20
What happened
The paper reports a two-season prospective cohort of 412 NCAA Division I athletes in six sports (388 analyzed, 8,946 athlete-weeks, 204 injuries) combining GPS/IMU load, neuromuscular tests, wellness self-reports and injury history. Using athlete-grouped cross-validation, XGBoost reached AUC 0.89 (sensitivity 0.85, specificity 0.88) versus logistic regression at 0.75; lowering the threshold to 0.35 raised sensitivity to 0.93 but cut specificity to 0.76. In season two, athletes above a 0.65 risk score were referred to sports medicine, who used clinical judgement on load changes. The authors report four-week injury incidence of 8.9% versus 13.5% in season-one high-risk athletes who were not referred, a 34% decrease, and state this is observational, not randomized. The paper does not name the institution or the approving board beyond 'the Institutional Review Board', and lists the authors as two master's students and an assistant professor at an Indian physical-education institute. Its limitations section cites a single institution, few rare-injury events and the observational design.
Why it matters
Why now: deposited October 17, 2026 and not archived. Collegiate injury-prediction claims like this are what vendors and staff may cite when buying athlete-monitoring AI. Interpretation: the claimed AUC is far above the 0.70 reported by the archived Frontiers collegiate cohort (Sep 30), and two recent reviews opened this run say external validation is rare and internal validation inflates estimates. That makes this a cautionary example of an attractive figure without verifiable provenance.
Evidence and measured results
Self-archived paper on Zenodo; the PDF shows received/accepted/published dates but no journal name, and peer review could not be confirmed. No data or code availability statement was found. Internal validation only; the intervention comparison uses a historical control.
Limitations and uncertainty
Unverifiable site and ethics approval; no journal identified; internal validation only; observational historical comparison; internal inconsistencies (wellness compliance described as 'low' at 91.4%). Treat all figures as author claims.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-19; 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 teams are offered AI injury-risk tools backed by impressive but unverified accuracy and reduction claims.
- Stakeholders
- Head athletic trainer and team physicians
- Director of sports performance
- Strength and conditioning coaches
- University compliance and privacy office
- Athletics IT
- Discovery questions
- Has the model been validated on a different institution or season?
- What false-alarm rate can staff absorb each week?
- Who reviews flags, and what actions follow?
- Which medical and wellness data would the tool hold, and under what consent?
- Can the vendor share calibration results?
- Value hypothesis
- A clinician-reviewed risk flag could help focus load-management attention, if validated locally and calibrated.
- Potential engagement
- An evidence and validation review of an injury-prediction tool using the department's own historical data before purchase.
- Unsupported claims
- A 34% injury reduction
- AUC 0.89 transferring to another program
- Causal injury prevention
- Peer-reviewed status
Pre-sales engineering
Role takeaway
- Fit
- Departments with consistent GPS/IMU use, routine jump testing, daily wellness reporting and reliable injury records.
- Architecture and integration
- Weekly join of wearable load, neuromuscular testing, wellness and injury-history data into rolling features feeding a gradient-boosted model and SHAP dashboard for sports medicine.
- Prerequisites
- Clean multi-season injury records
- High wearable and questionnaire compliance
- Data-sharing agreement between performance and sports medicine
- Defined referral threshold
- Constraints
- Single-site evidence only
- Rare injuries poorly predicted
- HIPAA and FERPA boundaries on medical data
- Observational evidence for benefit
- Security
- Medical history and biometric data require role-based access, audit logs, consent records and retention limits, and must stay within sports medicine control.
- Proof of value
- Temporal hold-out test on the department's own prior season
- Report calibration and weekly false-alarm counts
- Compare outcomes with a concurrent, not historical, group
- Clinician review of a sample of flags
Delivery
Role takeaway
- Work
- Build the data pipeline, validate locally, set the referral workflow and monitor flags and outcomes.
- Dependencies
- Sports medicine sign-off
- Wearable vendor data access
- Injury surveillance records
- Privacy office review
- Ownership
- Head athletic trainer or team physician, with sports performance as data steward.
- Skills and adoption
- Clinicians need to read SHAP explanations and probabilities; coaches need to accept that flags prompt review, not automatic load cuts.
- Governance checkpoints
- Privacy and consent review before data integration
- Local validation report before use
- Monthly false-alarm and outcome review
- Annual model revalidation
- Acceptance criteria
- Proposed: local temporal-validation AUC and calibration reported before go-live
- Proposed: every flag reviewed by a clinician within 48 hours
- Proposed: weekly false alarms below an agreed ceiling
- Risks
- Overreliance on inflated accuracy figures
- Alert fatigue from false positives
- Medical data exposure
- Misattributing injury changes to the model
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?
Describes a weekly pipeline joining wearable load, jump testing, wellness questionnaires and injury records into rolling 7- and 28-day features, with SHAP explanations on a staff risk dashboard. Interpretation: this data integration, not the model choice, is the main build effort for any department.
Governance
Who approves, reviews and stays accountable for outcomes?
Model flags were referred to clinicians rather than acted on automatically, a human-in-the-loop design worth copying. Interpretation: provenance gaps (unnamed site and board, no data statement) mean buyers should not treat the reported effect as evidence.
Security and privacy
What data, permissions and controls need testing?
Combines biometric, wellness and medical-history data on student-athletes, which brings HIPAA (where applicable) and FERPA boundaries, consent and access controls into scope. The paper says consent was obtained but gives no data-handling detail.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Adds interpretation work for athletic trainers; staff need training to use probabilistic flags alongside clinical judgement.
Procurement
What should contracts, pricing and exit terms secure?
Vendor claims citing studies like this should be checked for external validation, named sites and data availability before purchase.
Operating model
Which teams own the service once it runs?
Places an AI referral step between performance data and sports medicine decisions, requiring clear ownership of the threshold and response.
Publication history
- 2026-10-19College Athletics · Issue 413 resources
Stable resource ID: lnipe-preprint-ml-injury-risk-ncaa-di-cohort-unverified-2026-10-17