{"resourceId":"arxiv-soccermon-landmark-session-injury-labels-2026-09-03","versions":[{"version":"external-cd85a9acb8aa1bf528ecca6eb7292cdee722cded567923da8ca459b63b65c821","resource":{"id":"arxiv-soccermon-landmark-session-injury-labels-2026-09-03","title":"arXiv: minute-resolution monitoring cannot invent minute injury labels—five positive athletes limit precision","organization":"National Technical University of Athens","sector":"Academic sports analytics / athlete monitoring research","geography":"International (elite Norwegian women's football SoccerMon 2020 data; Greek authorship)","publishedAt":"September 3, 2026","publicationDate":"2026-09-03","eventDate":null,"sourceName":"arXiv","sourceLabel":"Preprint academic research","sourceUrl":"https://arxiv.org/html/2609.03790","evidenceClass":"academic-research","outcomeClass":"cautionary","topics":["data-security","infrastructure","governance-procurement","operating-model"],"finding":"Chatzidimitriou and Tserpes (submitted 2026-09-03) show that when injuries are labeled only at athlete-session level, analysts must not replicate positives across every minute row. On SoccerMon elite women's football data (3,743 sessions; 22 positive sessions from only five athletes), landmark session discrimination is non-monotonic (logistic ROC-AUC 0.499 to 0.607 to 0.367 across 10-60 min); TabPFN peaks at 0.692 at 30 min but does not consistently beat Random Forest; SMOTE/CTGAN help only conditionally and CTGAN drifts. Authors reject minute-alert deployment claims. Why now: collegiate wearable/injury-AI procurement continues while onset-timestamp quality remains the binding constraint.","sledRelevance":"Methodological control for college athletics sports-medicine/analytics teams evaluating wearable injury ML—protects student-athletes from overclaimed alerts and distinguishes elite/pro evidence from collegiate readiness.","evidence":"Athlete-disjoint validation; landmark AUC tables; augmentation grid with cluster bootstrap; explicit limitations on positive-athlete N and self-reported injuries.","architectureImplications":"One representation per athlete-session at fixed landmarks; keep targets at true label resolution; prefer athlete-disjoint evaluation over minute-row inflation.","governanceImplications":"Require onset-timestamp policy and positive-athlete N disclosure before any alerting language in vendor or internal AI claims.","securityPrivacyImplications":"Minute-resolution physiology is sensitive biometric/performance data even when used for research-style discrimination scores.","caveats":"Not NCAA/collegiate sample; elite Norwegian women's football; self-reported injuries; Team B had zero positives; preprint not peer-reviewed.","streamIds":["college-athletics"],"roles":{"sales":"Vendors and sports-science units often pitch minute-resolution wearable AI for injury risk, but this preprint shows session-level labels cannot support minute-level supervision. Discovery should ask what temporal resolution injury labels actually have, how many independent injury-positive athletes exist, and whether claims are discrimination scores or operational alerts. Bound value is method governance—label-unit alignment—not a deployable warning system. Unsupported claims include minute-specific onset prediction, clinical intervention timing at 30 minutes, or that synthetic data creates new injured athletes.","engineering":"Fit is a landmark, one-representation-per-athlete-session design for high-frequency monitoring with coarse labels—transferable caution for collegiate GPS/wearable pipelines. Architecture should keep targets at session/day resolution unless onset is medically timestamped; use athlete-disjoint validation. Prerequisites: clear injury ontology, onset fields, and positive-athlete counts. Constraints: elite Norwegian women's football SoccerMon 2020; only five injury-positive athletes; self-reported injuries; no collegiate sample. Security/privacy: athlete monitoring data remain sensitive even in research settings. Proof of value: refuse minute-alert pilots until onset-labeled cohorts exist; report PR-AUC vs prevalence, not ROC alone.","delivery":"Delivery implication is study design and go/no-go for injury ML, not a product install. Work includes auditing label timestamps, positive-athlete independence, and forbidding label replication across minutes. Dependencies: sports medicine documentation quality and institutional research agreements. Ownership: sports performance/medicine for labels; analytics for models. Skills adoption: train staff that dense sensors do not equal dense supervision. Governance checkpoints: label-resolution review before any alerting claim; external validation plan. Acceptance criteria: written non-deployment statement for minute alerts; documented positive-athlete N; athlete-disjoint metrics. Risks: false confidence from large minute-row counts and synthetic augmentation theater."},"roleDetails":{"sales":{"customerProblem":"Athletic departments are offered high-frequency wearable injury AI while injury documentation often lacks within-session onset timestamps.","stakeholders":["Sports medicine / ATC leadership","Sports performance analytics","Compliance / research integrity","Head coaches receiving risk flags"],"discoveryQuestions":["At what temporal resolution are injuries recorded (minute, session, day), and who adjudicates them?","How many independent injury-positive athletes exist in the local dataset?","Is the vendor claiming minute alerts, session discrimination, or research-only scores?"],"valueHypothesis":"Aligning model targets to true label resolution prevents false operational promises and focuses spend on better medical documentation before alerting ML.","potentialEngagement":"Label-resolution and positive-athlete audit of current wearable/injury data with a written go/no-go for any alerting use case.","unsupportedClaims":["Minute-specific injury onset prediction from session labels","30-minute landmark as a clinical intervention optimum","Synthetic oversampling creates new independent injured athletes"]},"engineering":{"fit":"Methodological pattern for landmark session representations under weak injury supervision; caution for collegiate GPS/HR pipelines with session/day labels.","architectureIntegration":"One feature vector per athlete-session at fixed elapsed landmarks; athlete-disjoint folds; no replication of session positives onto every minute row.","prerequisites":["Injury records with explicit timestamp policy","Athlete identifiers suitable for disjoint validation","Clear non-alert interpretation of discrimination scores"],"constraints":["Elite Norwegian women's football 2020 SoccerMon cohort—not NCAA","Only five injury-positive athletes; wide uncertainty","Self-reported injuries; Team B had zero positives"],"security":"Treat minute-resolution physiology as sensitive biometric/performance data; restrict research exports and re-identification risk.","proofOfValue":["Documented label-resolution statement before any pilot","PR-AUC contextualized to ~1% prevalence","Refusal criteria when positive-athlete N is too small"]},"delivery":{"work":"Audit injury label fields, freeze any minute-alert roadmap, and redesign evaluation to athlete-disjoint session units.","dependencies":["Sports medicine documentation practices","Wearable vendor export schemas","IRB/data-use agreements if research reuse"],"ownership":"Sports medicine owns injury truth; performance analytics owns model claims.","skillsAdoption":"Train analysts that sensor density does not create onset labels and that CTGAN/SMOTE is not biological replication.","governanceCheckpoints":["Pre-model label audit","External validation plan across teams/sexes/levels","Publication/marketing review for non-deployment language"],"acceptanceCriteria":["No production minute-level injury alerts without onset-labeled data","Reported metrics include positive-athlete N and athlete-disjoint validation","Synthetic augmentation, if used, audited for fidelity and not sold as new cases"],"risks":["Operationalizing research discrimination scores as alerts","Overfitting to a handful of positive athletes","Transferring pro/elite findings to collegiate populations without new evidence"]}},"retrievedAt":"2026-09-27T03:19:50Z","enrichedAt":"2026-09-27T03:22:00Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Pushes labor toward better medical documentation and label governance rather than more complex models on weak labels.","procurementImplications":"Reject minute-alert injury products that cannot show onset-labeled validation and independent positive-athlete counts.","operatingModelImplications":"Separates research discrimination scoring from operational alerting; sports medicine owns truth data before analytics ships flags to coaches.","sourceVerification":{"openedUrl":"https://arxiv.org/html/2609.03790","referenceExcerpt":"inferential precision remains limited by only five injury-positive athletes","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}