{"resourceId":"nature-ems-stroke-ml-triage-20260203","versions":[{"version":"external-d9c78240a1625a147c9bdef64393832f21de0cc20cad8e82e6d2122efd17a0ec","resource":{"id":"nature-ems-stroke-ml-triage-20260203","title":"Scientific Reports: EMS-data ML reaches ROC-AUC 0.843 for prehospital stroke triage (retrospective)","organization":"Saban et al. (multi-institution U.S. academic/clinical EMS research team)","sector":"Academic EMS / prehospital stroke triage research","geography":"United States (single EMS–ED linked retrospective cohort; transferability limits apply)","publishedAt":"2026-02-03","publicationDate":"2026-02-03","eventDate":null,"sourceName":"Scientific Reports (Nature Portfolio)","sourceLabel":"Peer-reviewed retrospective ML study on EMS stroke triage","sourceUrl":"https://www.nature.com/articles/s41598-026-37069-x","evidenceClass":"academic-research","outcomeClass":"mixed","topics":["data-security","operating-model","knowledge-work","accessibility-workforce"],"finding":"In a retrospective study of 4,333 patients / 8,221 ambulance encounters (2015–2020; stroke rate 2.0%), XGBoost best detected stroke (ROC-AUC 0.843 [0.77–0.89], PR-AUC 0.293 [0.16–0.45]) and Random Forest best detected severe stroke (ROC-AUC 0.826 [0.75–0.90], PR-AUC 0.186). Prehospital vitals were available in >88% of encounters for key measures but glucose/temperature were sparse; EMS CPSS screening missed many strokes (EMS identified only 32.3% of strokes across all encounters). Authors position ML as supplemental to clinical judgment and call for prospective multi-setting validation and better EMS documentation standardization.","sledRelevance":"Direct EMS clinical-AI evidence for stroke destination/triage support—distinct from non-emergency call bots and already-archived OHCA/stigma triage papers. Why now: as U.S. PSAPs and EMS medical directors expand from admin AI to clinical assist, this quantifies discrimination vs low PR-AUC and CPSS gaps that buyers must pressure-test locally.","evidence":"Opened publisher HTML and PDF via curl; confirmed online date 2026/02/03, cohort sizes, ROC/PR-AUC figures, CPSS/EMS identification rates, and prospective-validation call from Conclusions.","architectureImplications":"If piloted, score at ePCR/CAD time using structured vitals/dispatch codes with calibrated probabilities and SHAP-style explanations—not free-text LLM triage alone.","governanceImplications":"Require prospective silent-mode evaluation against destination and LVO outcomes before any routing change; disclose model use to medical direction and QA.","securityPrivacyImplications":"EMS–ED linked PHI training/inference needs BAAs, minimum necessary features, and clear prohibition on vendor secondary use/training without approval.","caveats":"Retrospective single-system cohort; low prevalence yields modest PR-AUC; not a live RCT of patient outcomes; feature availability (glucose/temp) and documentation quality limit transportability.","streamIds":["emergency-services"],"roles":{"sales":"Frame the Scientific Reports EMS stroke ML results as promising discrimination (ROC-AUC ~0.84) with low positive predictive value and explicit supplemental-tool framing—not as proven routing automation. Discovery should cover current stroke scales, destination protocols, ePCR completeness, and medical-direction appetite for silent-mode scoring. Bounded engagement: retrospective local replay plus prospective silent alerts measuring sensitivity at a fixed false-alert budget. Unsupported claims include reduced disability, guaranteed LVO capture, or replacement of CPSS/paramedic judgment.","engineering":"Fit is calibrated risk scoring on structured prehospital vitals and dispatch codes inside ePCR/CAD. Architecture needs reliable device-to-ePCR vitals, timestamp integrity, and explanation features for medics. Prerequisites include labeled stroke outcomes linkage and prevalence-aware thresholds. Constraints: retrospective training site, sparse glucose/temp, modest PR-AUC. Proof of value is prospective sensitivity/PPV at agreed alert rates versus current scale performance.","delivery":"Delivery includes silent-mode pilot design, medic training on score meaning, and QA sampling of disagreements. Dependencies: EMS medical director, ePCR vendor, receiving stroke centers, privacy officer. Ownership of destination remains with clinicians under protocol. Acceptance criteria: no automatic reroute without human confirm; published local calibration. Risks: alert fatigue, documentation-driven feature missingness, and over-trust in AUROC."},"roleDetails":{"sales":{"customerProblem":"Prehospital stroke scales miss atypical presentations while destination decisions are time-critical.","stakeholders":["EMS medical directors","Paramedic leadership","Stroke center coordinators","ePCR/IT","Privacy officers"],"discoveryQuestions":["What stroke scale is required today?","How often is CPSS documented?","Who can change destination?","What false-alert rate is tolerable?"],"valueHypothesis":"Calibrated ML on routine vitals may flag strokes missed by scales if used as a silent or advisory signal under medical direction.","potentialEngagement":"Local retrospective validation plus 90-day prospective silent-mode study.","unsupportedClaims":["Proven outcome improvement","Replaces CPSS","High PPV in low-prevalence field use"]},"engineering":{"fit":"Advisory stroke/severe-stroke scoring from structured EMS data.","architectureIntegration":"Monitor/ePCR vitals + dispatch codes → calibrated model → medic-visible advisory → destination protocol unchanged unless human confirms.","prerequisites":["Outcome-labeled cohort","Vitals completeness","Threshold policy","Audit logging"],"constraints":["Retrospective external evidence","Low prevalence/PR-AUC","Missing glucose/temp"],"security":"BAA-covered PHI pipeline; no vendor training on identifiable EMS data without approval.","proofOfValue":["Sensitivity at fixed alert rate","PPV","Agreement with stroke-center diagnosis"]},"delivery":{"work":"Stand up silent-mode scoring, medic education, and disagreement QA.","dependencies":["Medical direction","ePCR vendor","Receiving hospitals","Privacy/compliance"],"ownership":"EMS medical director owns clinical use policy.","skillsAdoption":"Train crews to treat scores as prompts to reassess, not orders.","governanceCheckpoints":["Monthly calibration review","Override audit","Destination protocol check"],"acceptanceCriteria":["No auto-reroute","Logged overrides","Prospective KPI report"],"risks":["Alert fatigue","Missing vitals degrade score","AUROC over-trust"]}},"retrievedAt":"2026-09-30T22:13:18-05:00","enrichedAt":"2026-09-29T22:02:40-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Paramedics/EMTs need workflows that surface scores without increasing scene time; medical directors must define override and documentation of disagreement.","procurementImplications":"Buy clinical stroke ML only with local prospective KPIs (sensitivity at fixed alert rate, destination appropriateness)—not ROC-AUC from another system.","operatingModelImplications":"Keep destination decisions under EMS medical direction; use ML to flag missed atypical strokes for clinician review.","sourceVerification":{"responseId":"curl-nature-stroke-ems-html-2026-09-30T221318-0500","openedUrl":"https://www.nature.com/articles/s41598-026-37069-x","referenceExcerpt":"The XGBoost model performed best for stroke detection (ROC-AUC 0.843","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}