{"resourceId":"frontiers-collegiate-load-injury-wearable-xgboost-2026-09-30","versions":[{"version":"external-96027add0ac3dae4b5290b1da6825fa492abed61b5f6d74198fa3ee98806d556","resource":{"id":"frontiers-collegiate-load-injury-wearable-xgboost-2026-09-30","title":"Prospective collegiate cohort: wearable plus movement-screen machine learning predicts load-related injury with modest accuracy (AUC 0.70) and no external validation","organization":"Xu, Zhang and Liao (Aviation University of Air Force, Changchun) in Frontiers in Physiology","sector":"Collegiate sports medicine and athlete performance analytics","geography":"China (international; transferability to U.S. NCAA programs limited)","publishedAt":"September 30, 2026","publicationDate":"2026-09-30","eventDate":null,"sourceName":"Frontiers in Physiology","sourceLabel":"Peer-reviewed journal article","sourceUrl":"https://www.frontiersin.org/journals/physiology/articles/10.3389/fphys.2026.1920109/full","evidenceClass":"academic-research","outcomeClass":"mixed","topics":["data-security","infrastructure","operating-model"],"finding":"In 161 collegiate soccer, basketball and track athletes followed for one season, 37 load-related time-loss injuries occurred (23.0%). An XGBoost model fusing wearable tibial and sacral sensor features with Functional Movement Screen task scores reached an out-of-fold AUC of 0.70 (95% CI 0.61 to 0.79), compared with 0.62 for screening alone and 0.64 for wearables alone. Biomechanical features added 0.07 AUC beyond previous-injury history. The Brier score was 0.171 against 0.177 for a no-skill baseline; sensitivity was 0.57 and specificity 0.76.","sledRelevance":"Why now: published Sep 30 and not yet in the archive, it gives a carefully reported benchmark as athletics departments keep adding predictive injury and performance AI. It shows that even a well-designed preseason model separates risk only modestly.","evidence":"Single-season prospective cohort reported against STROBE and TRIPOD+AI; nested five-fold cross-validation with embedded LASSO feature selection. 168 recruited, 161 analyzed. Fold-level AUC ranged 0.58 to 0.79. Events per variable were below 10 in most folds; a chronological split had too few events to report. No external validation cohort. No in-season workload data.","architectureImplications":"Requires three IMUs per athlete (tibial sensors at 1000 Hz, sacral at 200 Hz), standardized screening by certified raters, and a modeling pipeline that keeps feature selection inside cross-validation. Interpretation: departments evaluating vendor injury models should ask whether the vendor guards against this kind of leakage.","governanceImplications":"With sensitivity of 0.57 and positive predictive value of 0.46, such scores should inform screening conversations, not availability or roster decisions; sports medicine should own their use.","securityPrivacyImplications":"Biomechanical and injury data are health-related; collection needs informed consent, access limits and retention rules.","caveats":"Small event count; composite injury outcome; one season; no external or prospective validation; authors state findings are hypothesis-generating; non-U.S. setting.","streamIds":["college-athletics"],"roles":{"sales":"Sales framing: this study is a reality check for injury-prediction pitches. Even a careful design reached AUC 0.70, with sensitivity of 0.57 and no external validation. A bounded offer is an independent evaluation of a department's existing or proposed injury-risk tool against its own injury records. Do not claim injury reduction, accurate individual prediction, or transfer of these numbers to U.S. programs.","engineering":"Engineering framing: the methods are the useful part. Feature selection is nested inside cross-validation, calibration is checked on the original class balance, and the authors report fold-level spread and per-fold event counts. Engineers can use this as a checklist when reviewing vendor models. Proof means local retrospective validation, calibration plots and a held-out season before any model influences training decisions. Interpretation: the authors call results hypothesis-generating.","delivery":"Delivery framing: any injury-risk score needs a clinical owner and a defined use. Delivery should place model outputs inside sports medicine workflows, define what actions a high-risk flag may trigger, collect consent for sensor data, and review performance each season against observed injuries. Scores should never become sole grounds for playing-time or roster decisions. Interpretation: start with a retrospective check on past seasons before collecting new sensor data."},"roleDetails":{"sales":{"customerProblem":"Departments are offered injury-prediction tools but cannot tell how well they work for their own athletes.","stakeholders":["Head team physician","Director of sports medicine","Director of sport science","Strength and conditioning","Compliance and privacy office"],"discoveryQuestions":["What injury-risk tools do you use or plan to buy?","Have they been validated on your athletes?","Who decides what a high-risk flag means?","How is consent for sensor data collected?"],"valueHypothesis":"Independent local validation may prevent departments from overrelying on modest injury-risk scores and focus spending on tools that hold up.","potentialEngagement":"A retrospective validation and governance review of one injury-risk tool against a department's past seasons.","unsupportedClaims":["Injury reduction","Reliable individual-athlete prediction","Transfer of these results to U.S. NCAA programs","Superiority of XGBoost over simpler models"]},"engineering":{"fit":"Programs already running wearable sensors and movement screening with recorded time-loss injuries.","architectureIntegration":"Sensor data and screening scores joined with injury records in a sports science data store; model scores surfaced to sports medicine.","prerequisites":["Consistent injury definitions","Linked athlete identifiers across sources","Enough events to validate"],"constraints":["Small event counts per team","No external validation in this study","Non-U.S. cohort"],"security":"Treat sensor and injury data as health information with role-based access and retention limits.","proofOfValue":["Local AUC and calibration on past seasons","Sensitivity and positive predictive value at the chosen threshold","Stability across seasons"]},"delivery":{"work":"Place any injury-risk model under sports medicine control with defined actions and seasonal review.","dependencies":["Sports medicine sign-off","Athlete consent process","Data engineering support"],"ownership":"Head team physician or director of sports medicine.","skillsAdoption":"Train staff to read probabilities and false-positive rates rather than treating flags as diagnoses.","governanceCheckpoints":["Defined permitted uses before rollout","Seasonal performance review","Ban on use for roster or playing-time decisions"],"acceptanceCriteria":["Proposed: local validation report completed before operational use","Proposed: documented action protocol for flagged athletes"],"risks":["Overconfidence in modest scores","Stigmatizing athletes with false positives","Health-data privacy exposure"]}},"retrievedAt":"2026-10-06T22:17:02-05:00","enrichedAt":"2026-10-06T22:30:00-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Requires certified screening raters and sports science staff able to interpret model outputs.","procurementImplications":"Ask vendors for external validation, calibration and per-season event counts before purchase.","operatingModelImplications":"Predictive scores become one input to sports medicine review, not an automated decision.","sourceVerification":{"responseId":"curl-frontiers-1920109-20261006T221702","openedUrl":"https://www.frontiersin.org/journals/physiology/articles/10.3389/fphys.2026.1920109/full","referenceExcerpt":"No external validation cohort was available, a limitation acknowledged in the Discussion.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}