From the College Athletics edition of September 22, 2026
UGA Zenodo study: DI football GPS ML reaches ~0.878 ROC-AUC for non-contact injury risk amid severe imbalance
University of Georgia / collaborating NCAA Division I football program · College athletics / football sports science and injury-risk analytics · United States (NCAA Division I football; program unnamed in preprint)
- Publisher
- Zenodo
- Original publication
- May 25, 2026 (Zenodo v1)
- Source retrieved
- 2026-09-23
What happened
Patel and Rasheed present an end-to-end ML pipeline on de-identified daily GPS monitoring from a prominent NCAA Division I football program to identify elevated non-contact injury risk. With roughly 60 injury cases versus about 15,000 healthy sessions, an LSTM reached ROC-AUC approximately 0.878; ensembles improved sensitivity/specificity balance. Interpretable drivers include sudden spikes in acceleration, deceleration, and session duration, plus elevated risk after rest. Force-plate readiness was largely unavailable. Authors position the work as decision-support for workload management, not perfect prediction.
Why it matters
Newly relevant NCAA DI football primary-data signal not in the archive: complements archived Monmouth tennis and USF concussion papers with varsity football GPS methods, and contrasts vendor injury-AI claims by foregrounding imbalance and incomplete readiness data.
Evidence and measured results
Zenodo PDF (10.5281/zenodo.20385737): abstract AUC; imbalance counts; feature patterns; Limitations section on imbalance, GPS-only scope, and CMJ gaps; player-independent evaluation language.
Limitations and uncertainty
Zenodo preprint; single unnamed program; extreme imbalance; GPS-external features dominate; not an operational outcome study.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-23; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Football performance staff collect dense GPS workload data but struggle to turn rare non-contact injury events into trustworthy daily risk triage.
- Stakeholders
- Head football strength/performance
- Sports medicine / ATC
- Football coaching staff
- Athletics IT/analytics
- Compliance/privacy
- Discovery questions
- Which GPS vendor and export grain (athlete-day vs minute) are in use?
- How complete are force-plate/CMJ readiness feeds?
- Who may change practice load when a model flags risk?
- Have any prior models been tested player-independent under class imbalance?
- Value hypothesis
- If GPS-based temporal models flag elevated-risk days with interpretable spike/post-rest drivers, staff can prioritize human review—without claiming injury elimination.
- Potential engagement
- Four-to-six-week methods pilot: label non-contact events, evaluate LSTM/ensemble baselines with player-independent folds, and brief sports medicine on alert thresholds.
- Unsupported claims
- Guaranteed reduction in non-contact injuries
- Multi-school external validation already completed
- That force-plate readiness is fully modeled in this study
Pre-sales engineering
Role takeaway
- Fit
- Decision-support ML on collegiate football GPS athlete-day sequences for elevated non-contact injury risk.
- Architecture and integration
- Batch ingest daily GPS features; build athlete-day labels from sports-medicine coding; train temporal/tabular models offline; serve risk scores and feature explanations to staff tools only.
- Prerequisites
- De-identified GPS history with injury labels
- Player-independent evaluation protocol
- Clear human owner for load changes
- Constraints
- Extreme class imbalance (~60 vs ~15,000)
- Incomplete CMJ/force-plate coverage
- Single DI program; preprint status
- Security
- Restrict injury and GPS identifiers; no athlete-facing raw scores without clinical gate; retain access logs.
- Proof of value
- Held-out ROC-AUC with sensitivity/specificity/G-mean
- Documented false-alarm burden at chosen threshold
- Sports-medicine review of flagged high-risk days
Delivery
Role takeaway
- Work
- Pilot labeling pipeline, model card, coach/ATC briefing, and threshold policy for advisory alerts.
- Dependencies
- GPS data license/export
- Sports-medicine coding time
- Performance staff scheduling
- Ownership
- Sports performance owns operational use; sports medicine owns medical authority; analytics owns model maintenance.
- Skills and adoption
- Staff train on imbalance metrics and interpreting spike/post-rest drivers rather than accuracy headlines.
- Governance checkpoints
- Written advisory-only policy
- Clinician sign-off before any remove-from-play automation
- Seasonal re-validation
- Acceptance criteria
- Player-independent metrics documented
- Alert review workflow live for one unit
- CMJ gap explicitly listed as out-of-scope
- Risks
- Alert fatigue from low specificity
- Overfitting to one program
- Privacy incidents from broad score sharing
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?
Build athlete-day stores from GPS first; evaluate temporal models under player-independent folds; treat CMJ joins as optional until coverage is complete; ship scores only into staff tools.
Governance
Who approves, reviews and stays accountable for outcomes?
Keep alerts advisory; sports medicine retains unchallengeable medical authority; publish model cards stating imbalance and single-program limits.
Security and privacy
What data, permissions and controls need testing?
De-identify exports; restrict who sees individual risk ranks; avoid secondary AI-training licenses on injury labels without consent review.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Less-resourced DI programs need clear imbalance education so AUC headlines do not outrun staffing for alert review.
Procurement
What should contracts, pricing and exit terms secure?
Require player-independent metrics, sensitivity at usable specificity, and readiness-data coverage disclosures in injury-AI RFPs.
Operating model
Which teams own the service once it runs?
Separate model ops from medical decision rights; performance staff adjust load only within clinician-approved policies.
Publication history
- 2026-09-22College Athletics · Issue 164 resources
Stable resource ID: uga-zenodo-di-football-gps-injury-ml-2026