{"resourceId":"monmouth-part-tennis-readiness-ensemble-2026","versions":[{"version":"external-8609ab6071f040682ddb0b3d83687f5721ae6f21ba31762c45a8dfd6f8ea5c54","resource":{"id":"monmouth-part-tennis-readiness-ensemble-2026","title":"Monmouth PART ensemble shows middling injury AUC and weak injury-class F1 in nine-player tennis pilot","organization":"Monmouth University (arXiv / IEEE Systems, Man, and Cybernetics Magazine Early Access)","sector":"College athletics / collegiate tennis sports science","geography":"United States — Monmouth University collegiate tennis cohort","publishedAt":"August 21, 2026 (arXiv v1); IEEE Early Access 2026","publicationDate":"2026-08-21","eventDate":null,"sourceName":"arXiv","sourceLabel":"Academic preprint — PART multimodal tennis readiness","sourceUrl":"https://arxiv.org/html/2608.21530v1","evidenceClass":"academic-research","outcomeClass":"mixed","topics":["data-security","developers-agents","operating-model","accessibility-workforce"],"finding":"Qu et al. present PART, a multimodal weighted ensemble producing an Athlete Readiness Score from wearables, questionnaires, jump tests, and match video for nine collegiate tennis players (5 male, 4 female). Injury-risk classifiers reach AUC-ROC about 0.61–0.65 with overall accuracy near 0.74 for tree/boosting models but injury-class F1 only about 0.40–0.43 amid ~24% injury observations; authors flag pilot scope, heuristic playing-style weights, and need for larger validation. Distinct from the previously archived Monmouth tennis injury-classification preprint (2608.25126).","sledRelevance":"Why now: Complements 2026-09-15 fatigue-monitoring and prior Monmouth injury-class work with an unarchived readiness-ensemble paper that again shows headline accuracy can obscure weak injury detection—directly informing acceptance tests for collegiate monitoring AI.","evidence":"Table II reports Logistic Regression AUC 0.61±0.01 accuracy 0.48±0.09 F1(injury) 0.40±0.01; XGBoost/Decision Tree/Random Forest AUC ~0.64–0.65, accuracy 0.74, F1(injury) ~0.43. Top injury features: HRV 0.334, RHR 0.184, Recovery Score 0.112. IRB FA2584; written consent. Authors state limited cohort restricts statistical power.","architectureImplications":"Separating-stage models plus supervised ensemble weights; evaluate injury-class metrics and component contributions before exposing ARS to coaches.","governanceImplications":"Keep experimental ARS out of scholarship/discipline decisions; require sports-medicine ownership of load changes.","securityPrivacyImplications":"Continuous wearable and video biometrics need purpose-limited storage and athlete consent aligned to NCAA performance-tech expectations.","caveats":"n=9 one-semester pilot; not an injury-reduction trial; playing-style weights heuristic; IEEE Early Access version may differ slightly from arXiv HTML; recreational extension is speculative.","streamIds":["college-athletics"],"roles":{"sales":"Sales framing: Collegiate tennis and Olympic-sport programs seeking multimodal readiness scores will see PART-style ensembles that fuse wearables, wellness surveys, jump tests, and match video—but the Monmouth pilot’s injury classifiers show middling AUC (~0.64–0.65) with low injury-class F1 despite higher overall accuracy. Sell shadowed readiness dashboards and load-management decision support, not autonomous injury prediction. Stakeholders are tennis coaches, sports medicine, and compliance/IRB. Unsupported claims include season-wide injury reduction or transfer of nine-player results to multi-sport NCAA departments.","engineering":"Engineering framing: Architecture is a separating-stage set of wellness regression, injury classification, physical-capability, and playing-style models integrated via supervised weighted ensemble into an Athlete Readiness Score. Fit for small roster sports with daily wearable and questionnaire compliance. Prerequisites: IRB/consent (Monmouth FA2584 cited), stratified splits, class-weighting for ~24% injury observations, and video pipelines. Constraints: n=9 one-semester pilot, heuristic playing-style weights, imbalance hurting injury F1. Security: protect HRV/RHR/recovery and video biometrics with role-based coach views. Proof: injury-class F1/precision at operating threshold, not headline accuracy alone.","delivery":"Delivery framing: Reproduce features on local tennis roster, validate offline with player-wise splits, then run case-study style readiness reviews without automatic practice bans. Dependencies: wearable vendor exports, strength-testing protocols, and sports-medicine ownership of interventions. Skills adoption teaches staff to read ARS components and override when video or clinical context disagrees. Governance checkpoints: consent refresh, retention limits, and prohibition on scholarship uses of experimental scores. Acceptance criteria: documented improvement in injury-class detection versus baseline plus trainer-rated usefulness of two load-management recommendations per month."},"roleDetails":{"sales":{"customerProblem":"Tennis/Olympic-sport staff want multimodal readiness insight but risk acting on high overall accuracy that conceals weak injury-class detection.","stakeholders":["Tennis coach","Sports medicine / athletic training","Strength staff","Compliance/IRB"],"discoveryQuestions":["What is the local injury-class base rate?","Which wearable and wellness feeds are reliable daily?","Who may change practice from an ARS flag?","Is video motion analysis already in workflow?"],"valueHypothesis":"A weighted multimodal ARS can support load management if validated on injury-class F1/precision locally—without promising PART’s pilot metrics.","potentialEngagement":"Offline validation on one semester of tennis data plus shadowed readiness reviews.","unsupportedClaims":["Injury reduction from deploying PART","Generalization beyond nine collegiate players","Learned playing-style weights (authors call them heuristic)"]},"engineering":{"fit":"Multimodal ensemble: wellness regression, injury classification, physical capability, playing-style, fused to ARS.","architectureIntegration":"Daily wearable/questionnaire/jump/video features into secure sports-med store; ensemble weights from supervised reliability; coach UI shows component scores.","prerequisites":["Consent/IRB","Wear compliance","Outcome labeling","Stratified validation"],"constraints":["n=9 pilot","~24% injury observations","Low injury F1 vs higher accuracy","Heuristic style weights"],"security":"Protect HRV/RHR/recovery and biomechanics video; minimize secondary use.","proofOfValue":["Injury-class F1 at chosen threshold","Upper vs lower-body AUC checks","Trainer agreement on case-study flags"]},"delivery":{"work":"Feature parity check, offline validation, two case-study style reviews, then decide on limited production shadowing.","dependencies":["Wearable exports","Jump-test protocol","Video capture ops"],"ownership":"Sports medicine owns interventions; coach owns practice plan changes.","skillsAdoption":"Train staff to distrust headline accuracy when injury F1 is weak.","governanceCheckpoints":["Consent refresh","No scholarship use of experimental ARS","Retention schedule"],"acceptanceCriteria":["Local injury-class metric meets pre-set floor","Documented human overrides","No automatic penalties from scores"],"risks":["Acting on false negatives","Overfitting small n","Athlete survey burden","Scope creep to LLM coach advice before validation"]}},"retrievedAt":"2026-09-17T03:04:50Z","enrichedAt":"2026-09-17T03:04:50Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Small-roster Olympic sports can pilot multimodal stacks, but labeling and daily survey burden may exceed staffing at thinner programs.","procurementImplications":"Budget for wearables, jump testing, video, and validation labor—not only model licenses; demand injury-class performance in RFPs.","operatingModelImplications":"Supports human-in-the-loop readiness review rather than automatic practice bans when minority-class detection is weak.","sourceVerification":{"openedUrl":"https://arxiv.org/html/2608.21530v1","referenceExcerpt":"The limited cohort size (nine players) restricts statistical power and the ability to capture the full diversity of injury patterns.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}