{"resourceId":"nature-liu-acl-core-stability-ml-2026-09-29","versions":[{"version":"external-e3d6b5af66ccb72c9ce8372f019ccc4e02b68e82f453f7de4f2faf33bf3ec2b4","resource":{"id":"nature-liu-acl-core-stability-ml-2026-09-29","title":"Scientific Reports (same-day): PE-student ACL risk ML classifies surrogate phenotype, not observed injury","organization":"Guangzhou Institute of Applied Science and Technology; Gdansk University of Physical Education and Sport; Hong Kong Polytechnic University","sector":"University physical education / sports medicine ML research","geography":"China and Poland (PE students; transfer limits to U.S. NCAA)","publishedAt":"29 September 2026","publicationDate":"2026-09-29","eventDate":null,"sourceName":"Scientific Reports (Nature Portfolio)","sourceLabel":"Peer-reviewed academic research","sourceUrl":"https://www.nature.com/articles/s41598-026-72807-1","evidenceClass":"academic-research","outcomeClass":"cautionary","topics":["data-security","governance-procurement","operating-model","accessibility-workforce"],"finding":"Liu et al. (published 2026-09-29) recruit 135 physical education students and train seven ML classifiers on a 10-feature core-stability battery to identify a composite biomechanical high-risk phenotype from landing kinematics and strength ratios. XGBoost reached AUC-ROC 0.847 (sensitivity 78.6%, specificity 82.8%). Authors repeatedly state no participant was followed for incident ACL injury—so models classify a surrogate phenotype—and call the framework exploratory pending multi-center prospective validation against observed ACL injury.","sledRelevance":"Same-day university-athlete ACL-risk ML that pairs usable AUC with unusually clear non-deployment language—directly useful when collegiate sports medicine is pitched injury-screening AI.","evidence":"Cross-sectional n=135; 10×10 repeated stratified CV; SHAP feature ranking (Y-Balance, side bridge, FMS); explicit surrogate-outcome and external-validation limitations in abstract/conclusion.","architectureImplications":"Offline tabular/explainable screening research stack—not a real-time wearable injury alarm; outputs must stay in clinician review queues without write-back to practice plans.","governanceImplications":"Requires IRB/ethics framing that high-risk labels are phenotype hypotheses, plus rules against scholarship/playing-time automation until prospective injury endpoints exist.","securityPrivacyImplications":"Biomechanical and strength assessments on student-athletes are sensitive; live NCAA replication would need FERPA-aware access controls and careful athlete communication of labels.","caveats":"Surrogate phenotype only; single-center PE students (not NCAA varsity); no external validation; China/Poland context.","streamIds":["college-athletics"],"roles":{"sales":"Sports medicine buyers see AUC 0.85 ACL slides and may assume injury prediction. Discovery must ask whether the endpoint is observed ACL tears or a biomechanical surrogate, whether any athlete was followed prospectively, and how labels are shared. Bound value is interpretable core-stability screening hypotheses (Y-Balance, side bridge, FMS)—not clinical automation. Unsupported claims include predicting ACL tears, NCAA roster readiness, or that outperforming logistic regression equals deployable screening.","engineering":"Fit is research-grade, explainable classification of a landing/strength phenotype. Architecture should batch-score assessments, attach SHAP drivers, and route to AT/physician review without auto practice changes. Prerequisites: local labeled data, temporal validation, and ethics review of label communication. Constraints: surrogate endpoint and single-center PE sample. Security: restrict phenotype scores and raw assessments to sports medicine roles. Proof of value: prospective season holdout against observed ACL/MSK events at real prevalence—not CV AUC alone.","delivery":"Delivery is a staged research path: replicate features locally, validate prospectively, then pilot decision-support with counseling resources. Dependencies: sports medicine ownership and IRB capacity. Ownership: team physician/ATC for clinical use; data science for model maintenance. Skills adoption teaches staff to treat scores as hypotheses. Governance checkpoints: ban on playing-time automation; athlete-facing explanation protocol; annual external-validation review. Acceptance criteria: documented surrogate-vs-injury distinction in all vendor materials; prospective protocol registered. Risks include false reassurance, stigma from high-risk labels, and overfitting to PE cohorts."},"roleDetails":{"sales":{"customerProblem":"Collegiate sports medicine wants ACL-risk tools, but many models quietly predict surrogates rather than tears.","stakeholders":["Head athletic trainer / team physician","Sport coaches","Student-athlete wellness","Compliance / privacy"],"discoveryQuestions":["Is the labeled endpoint observed ACL injury or a biomechanical surrogate?","Was anyone followed prospectively for tears?","Who sees high-risk labels and how are they explained to athletes?"],"valueHypothesis":"Explainable core-stability screening can prioritize Y-Balance/side-bridge/FMS follow-up if treated as hypothesis generation after local prospective validation.","potentialEngagement":"Protocol workshop defining surrogate vs injury endpoints and a season-long prospective validation plan before any roster-facing use.","unsupportedClaims":["Model predicts ACL tears in NCAA athletes","AUC 0.847 equals clinical screening readiness","Safe to automate playing time from phenotype scores"]},"engineering":{"fit":"Explainable tabular phenotype classifier for PE/university athletes—not an autonomous injury device.","architectureIntegration":"Assessment capture → batch score + SHAP → clinician queue; no write-back to practice plans or scholarships.","prerequisites":["Local longitudinal injury labels","Ethics-approved label communication","Role-based access to assessment data"],"constraints":["Surrogate outcome only","Single-center PE sample","No external validation"],"security":"Treat biomechanical assessments and risk scores as sensitive student-athlete data; FERPA-aware logging and minimization.","proofOfValue":["Prospective validation against observed ACL/MSK events","Calibration at local injury prevalence","Athlete communication pilot with documented overrides"]},"delivery":{"work":"Stand up assessment pipeline, replicate features, run prospective validation, then limited decision-support pilot.","dependencies":["Sports medicine staffing","IRB/ethics approval","Counseling capacity for high-risk conversations"],"ownership":"Team physician/ATC owns clinical interpretation; data science owns model lifecycle.","skillsAdoption":"Train staff that scores are phenotype hypotheses; document overrides.","governanceCheckpoints":["Pre-pilot surrogate-vs-injury disclosure","Quarterly label-use audit","Ban on automated roster decisions"],"acceptanceCriteria":["Written non-deployment until prospective validation","Athlete-accessible explanation path","No scholarship automation"],"risks":["Stigma from high-risk labels","False reassurance","China/Poland-to-NCAA transfer failure"]}},"retrievedAt":"2026-10-01T03:20:10Z","enrichedAt":"2026-10-01T03:26:10Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Low-cost field tests (Y-Balance, side bridge, FMS) could widen screening access, but misapplied labels may harm scholarships and mental health without counseling capacity.","procurementImplications":"Require written endpoint definition (injury vs surrogate), prospective validation plan, and non-automation clauses before buying ACL-risk AI.","operatingModelImplications":"Keeps sports medicine as decision owner; AI may rank athletes for further assessment only after local validation.","sourceVerification":{"openedUrl":"https://www.nature.com/articles/s41598-026-72807-1","referenceExcerpt":"No participant was followed for incident ACL injury, so all models classify this surrogate phenotype rather than predict injury.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}