{"resourceId":"bmc-raju-collegiate-injury-ml-kaggle-2026-01-05","versions":[{"version":"external-c6d5a6638f4aeb383d8d932ef7a05dff4c06517c2e1df099e5641f696510a7dc","resource":{"id":"bmc-raju-collegiate-injury-ml-kaggle-2026-01-05","title":"BMC study: Random Forest hits 0.98 accuracy on Kaggle collegiate injury table—authors still call for external validation","organization":"BMC Sports Science, Medicine and Rehabilitation (Raju et al.)","sector":"Academic sports science / collegiate athlete analytics","geography":"International authors; public collegiate dataset (Kaggle); open-access journal","publishedAt":"January 5, 2026","publicationDate":"2026-01-05","eventDate":"2026-01-05","sourceName":"BMC Sports Science, Medicine and Rehabilitation","sourceLabel":"Peer-reviewed open-access research article","sourceUrl":"https://link.springer.com/article/10.1186/s13102-025-01502-x","evidenceClass":"academic-research","outcomeClass":"cautionary","topics":["data-security","developers-agents","governance-procurement","operating-model"],"finding":"Raju et al. (published 2026-01-05) present an AutoML+SHAP injury-risk framework on a public Kaggle multi-sport college-athlete dataset (n=200). Random Forest achieved the best reported performance (accuracy = 0.98; ROC-AUC = 0.97) using workload, recovery, performance, and demographic features; SHAP highlighted ACL risk, load balance, fatigue, and training hours. The paper states the work is secondary analysis of Kaggle data with no new athlete collection and calls for real-time monitoring and external validation. Why now: newly relevant caution as departments buy injury/performance AI while retrospective metrics on public tables can look deployment-ready without prospective NCAA validation.","sledRelevance":"Collegiate injury ML evidence with explicit secondary-data and external-validation limits—useful brake on vendor and research hype for athletic-department buyers.","evidence":"Peer-reviewed open-access article with abstract metrics, methods describing Kaggle secondary analysis, SHAP feature discussion, and limitations calling for external/real-time validation.","architectureImplications":"Offline tabular ensemble ML with explainability—not a streaming wearable architecture; transfer requires local labels and leakage controls.","governanceImplications":"High retrospective metrics must not authorize coach-facing risk scores without prospective validation, clinical ownership, and athlete notice.","securityPrivacyImplications":"Study used public anonymized data; campus analogs involve student-athlete health information needing access control and GenAI paste bans.","caveats":"Optimistic metrics on secondary public data; class imbalance and lack of daily wearable time-series limit deployability; not U.S. NCAA field validation.","streamIds":["college-athletics"],"roles":{"sales":"Sports-performance and athletic-training buyers see headline injury-prediction accuracy near 98% on collegiate data and may over-read deployability. Discovery must ask whether models were prospectively validated on the school's athletes, what the injury definition was, and whether data were wearable streams or secondary public tables. Bound value is an interpretable AutoML+SHAP research framework on a Kaggle collegiate dataset—not a production athletics system. Unsupported claims include real-world injury prevention, wearable replacement, or generalizability to U.S. NCAA programs from this study alone.","engineering":"Fit is a retrospective Random Forest/XGBoost/ANN comparison with SHAP on n=200 multi-sport college athletes from a public Kaggle set (accuracy 0.98; ROC-AUC 0.97 for RF). Architecture is offline tabular ML—not real-time GPS/HRV pipelines; authors note absence of daily time-series and need for external validation. Prerequisites for any transfer: local labeled injuries, leakage checks, and calibration on the department's sports. Constraints: secondary data, possible optimistic metrics, no field deployment. Security/privacy: public anonymized data here; campus use would implicate student-athlete health data controls. Proof of value: prospective holdout on institutional data beating simple baselines.","delivery":"Treat as research input to an evaluation checklist, not a turnkey product. Work includes reproducing metrics on local data, defining injury labels with sports medicine, and setting acceptance thresholds before any coach-facing risk scores. Dependencies: IRBs/data agreements, sports medicine ownership, IT security. Ownership: sports performance/medicine for clinical interpretation; data science for model risk. Skills adoption: coaches learn SHAP drivers are associations not mandates. Governance: no roster decisions from unvalidated scores. Acceptance: prospective AUC and calibration documented; athlete notice/consent path defined. Risks: over-trust in 0.98 accuracy and equity harms from biased labels."},"roleDetails":{"sales":{"customerProblem":"Athletic departments want injury-risk analytics but struggle to separate research metrics from production-ready systems.","stakeholders":["Sports medicine / athletic training","Strength & conditioning","Athletics data science","Compliance/privacy"],"discoveryQuestions":["Was validation prospective on your athletes or only retrospective on public data?","How is injury defined and labeled locally?","Who may see athlete-level risk scores?"],"valueHypothesis":"Explainable tabular ML can support research triage of workload-recovery factors if prospectively validated—not if buyers treat 0.98 accuracy as deployment-ready.","potentialEngagement":"Model-risk assessment workshop: reproduce pipeline on de-identified local data and compare to a naive baseline before any pilot.","unsupportedClaims":["Real-world injury reduction","Generalization from Kaggle collegiate data to NCAA DI programs","Replacement of wearable monitoring programs"]},"engineering":{"fit":"Offline AutoML+SHAP injury classification on aggregated collegiate tabular features (n=200), RF best at accuracy 0.98 / ROC-AUC 0.97.","architectureIntegration":"Batch feature table from training/recovery logs to train/eval with stratified CV and SHAP explanations; not a streaming wearable architecture.","prerequisites":["Local labeled injury outcomes","Leakage and imbalance review","Calibration and threshold policy with clinicians"],"constraints":["Secondary Kaggle analysis; no new athlete data collection","No daily wearable time-series in the used dataset","Authors call for external/real-time validation"],"security":"Campus deployment would require student-athlete health-data controls, access logs, and limits on coach-facing raw scores.","proofOfValue":["Prospective season holdout beats baseline","Calibration plots accepted by sports medicine","Documented false-negative cost analysis"]},"delivery":{"work":"Literature-informed evaluation checklist, local data feasibility study, and go/no-go before coach dashboards.","dependencies":["Sports medicine labeling protocol","IRB/data-use agreements","IT security review"],"ownership":"Sports medicine owns clinical use decisions; analytics owns model risk documentation.","skillsAdoption":"Train staff that SHAP drivers are associative and scores are not automatic play/no-play rules.","governanceCheckpoints":["Pre-pilot model card and limitations review","Athlete notice/consent check","Mid-season performance monitoring"],"acceptanceCriteria":["Prospective metrics pre-specified and met","No roster/medical decisions from unvalidated scores","Written escalation when model and clinician disagree"],"risks":["Over-trust in headline 0.98 accuracy","Bias against athlete subgroups","Privacy exposure if health features enter GenAI tools"]}},"retrievedAt":"2026-09-29T03:24:20Z","enrichedAt":"2026-09-29T03:25:30Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"SHAP explanations can help non-ML staff interpret drivers, but risk over-trust if presented without limitations literacy.","procurementImplications":"Demand prospective validation, calibration, and model cards before buying injury-prediction products that cite similar headline accuracies.","operatingModelImplications":"Keep injury models in research/evaluation mode under sports medicine until prospective acceptance criteria are met.","sourceVerification":{"openedUrl":"https://link.springer.com/article/10.1186/s13102-025-01502-x","referenceExcerpt":"Random Forest has shown the best performance (accuracy = 0.98; ROC-AUC = 0.97)","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}