{"resourceId":"monmouth-part-tennis-injury-classification-2026","versions":[{"version":"external-4a02cc18deaeecc903a1fc65d809750b7126004270b58f69bc44c22dd31dfd5a","resource":{"id":"monmouth-part-tennis-injury-classification-2026","title":"Collegiate tennis model shows why overall accuracy can conceal weak injury detection","organization":"Monmouth University","sector":"Higher education athletics","geography":"United States","publishedAt":"arXiv deposited August 25, 2026; manuscript identifies a 2025 IEEE conference publication","publicationDate":"2026-08-25","eventDate":null,"sourceName":"arXiv / accepted author manuscript","sourceLabel":"Collegiate tennis modeling study","sourceUrl":"https://arxiv.org/html/2608.25126v1","evidenceClass":"academic-research","outcomeClass":"cautionary","topics":["knowledge-work","data-security","governance-procurement","operating-model"],"finding":"PART combines wearables, questionnaires, jump testing and video. Its injury classification results warrant caution.","sledRelevance":"Direct collegiate evidence, with limited transfer beyond the studied tennis cohort.","evidence":"Nine players were followed for 16 weeks. Table II reports XGBoost accuracy 0.852, F1 0.100 and AUC 0.695; logistic regression accuracy is 0.849 with F1 0.000.","architectureImplications":"Interpretation: separate ingestion, label generation and model evaluation; test time and athlete separation before integrating forecasts.","governanceImplications":"Interpretation: reserve participation decisions for accountable staff; require a documented challenge route.","securityPrivacyImplications":"Interpretation: restrict identifiable wellness data, separate research access from coaching views, and audit exports.","caveats":"Injury-label construction and classifier split details are insufficiently specified. No external validation or injury-reduction trial is established.","streamIds":["college-athletics"],"roles":{"sales":"Interpretation: Frame the customer problem as evaluating whether readiness analytics can support review without overwhelming staff. Engage the athletic trainer, tennis coach, athlete representatives and athletics IT. Ask how actual injuries are labeled, what triggers a review and how many alerts staff can handle. Offer a bounded retrospective data-and-evaluation assessment with an explicit stop/go decision. The value hypothesis is better evidence for purchasing and workflow design. Do not promise fewer injuries, competitive gains or savings from the reported accuracy. Applicability is strongest for programs able to support careful tennis-specific evaluation.","engineering":"Interpretation: Build a restricted research workspace with versioned inputs and independently adjudicated outcome labels before any live integration. Establish permitted exports, identifier mapping and timestamp alignment. Fit preprocessing on training data only; use future-period and held-out-athlete tests, with a simple baseline and confidence intervals. Examine confusion matrices and threshold-specific alert volume rather than selecting on accuracy alone. Treat insufficient injury events as a reason to defer conclusions. Cloud, on-premises or hybrid placement should follow institutional data controls; the paper does not establish a preferred hosting model. Do not expose identifiable records through a general-purpose copilot.","delivery":"Interpretation: Assign a sports-medicine owner and a data engineer, with athlete representation at design review. First document consent, access, retention and escalation; then pilot in shadow mode. Train reviewers to record disagreement and correction reasons. Proposed acceptance criteria are complete provenance for every evaluated prediction, zero unauthorized exports, a reproducible held-out evaluation and an agreed maximum alert workload. Set performance thresholds with the clinical owner before testing. These are proposed gates, not observed results. Key dependencies are reliable outcome labels and staff review time; key risks are false reassurance, unnecessary restrictions and coercive participation."},"retrievedAt":"2026-09-14T03:00:55Z","enrichedAt":"2026-09-14T03:03:04Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Interpretation: offer accessible questionnaires and a non-digital reporting route; train staff to explain uncertainty.","procurementImplications":"Interpretation: request reproducible event-level testing and data-exit rights before procuring a readiness service.","operatingModelImplications":"Interpretation: sports medicine should own escalation; analysts should own pipeline quality. Copilot, agent and facilities applicability is limited.","sourceVerification":{"openedUrl":"https://arxiv.org/html/2608.25126v1","referenceExcerpt":"In contrast, Logistic Regression had an F1 score of 0","promptVersion":"sled-research-v3.2","model":null,"basis":"agent-reported inspection"}}}]}