{"resourceId":"researchsquare-sweden-ohca-ml-vs-dispatchers-discordant-90-calls-20261018","versions":[{"version":"external-1e30440959ea1f4fb416e049a0f18405759bbe5ea35a47a618d0890ff3f5aedb","resource":{"id":"researchsquare-sweden-ohca-ml-vs-dispatchers-discordant-90-calls-20261018","title":"Research Square preprint: in 90 Swedish cardiac arrest calls missed by one side, ML caught 52 dispatchers missed and dispatchers caught 38 ML missed","organization":"Swedish OHCA dispatch research team (Research Square preprint)","sector":"Emergency medical dispatch","geography":"Sweden","publishedAt":"2026-10-18","publicationDate":"2026-10-18","eventDate":null,"sourceName":"Research Square","sourceLabel":"Research Square preprint: Characteristics of out-of-hospital cardiac arrest calls unrecognised by machine learning versus medical dispatchers","sourceUrl":"https://www.researchsquare.com/article/rs-11153391/v1","evidenceClass":"academic-research","outcomeClass":"mixed","topics":["knowledge-work","operating-model","governance-procurement"],"finding":"Using a nationwide 2018 Swedish OHCA cohort in which the ML model recognized 729 of 851 cardiac arrest calls and dispatchers 715, the authors audited the 90 discordant 112 calls with CARES criteria. ML recognized arrest that dispatchers missed in 58% (52/90); dispatchers recognized arrest the ML missed in 42% (38/90). ML did better when the caller knew the patient, was a layperson, was alone, the arrest was at home, or the call was in the evening; dispatchers did better when CPR was already under way or the caller was not alone. The authors conclude the two show different failure patterns and that ML support may complement dispatchers.","sledRelevance":"Directly relevant to 911 and EMS dispatch centers considering AI cardiac arrest recognition: it shows where machines and call-takers each fail, which matters for alert design.","evidence":"Observational secondary analysis of 2018 audio; preprint, not peer reviewed; small discordant sample; odds ratios have wide confidence intervals.","architectureImplications":"Real-time ML listening on emergency call audio with alerts to the call-taker; the study does not test live alerting.","governanceImplications":"Because each side misses different calls, policy should treat ML as a second listener with protocols for when its alert conflicts with the dispatcher's judgement.","securityPrivacyImplications":"Requires processing live caller audio; retention and access rules for recordings apply.","caveats":"Data from 2018; Swedish language and protocols; only 90 calls; association analysis, not a trial of live alerting.","streamIds":["emergency-services"],"roles":{"sales":"For sales, the point is that AI and dispatchers miss different cardiac arrest calls, so a second listener has a credible case. Dispatch directors care about the roughly one in four arrests missed on calls, and this study shows where ML catches some of them, especially lone lay callers at home. It does not show lives saved, faster CPR or that dispatchers will act on alerts. Any engagement should be a bounded retrospective audit of the agency's own calls before any live alerting.","engineering":"For pre-sales engineering, the system is ML speech analysis on live emergency call audio raising an alert to the call-taker. The study's value is its failure-mode breakdown: ML struggles when CPR is already under way or several people are talking. That tells engineers which call types to test and how to present alerts. Integration with call-handling and CAD systems, language coverage and latency are prerequisites. Data is from 2018 Swedish calls, so local language, accents and protocols must be validated on the agency's own recordings.","delivery":"For delivery, start with a retrospective audit of local cardiac arrest calls comparing ML output with dispatcher recognition, then design alert handling rules. Ownership sits with the dispatch center director and EMS medical director. Call-takers need training on what to do when an alert conflicts with their judgement. Acceptance should be measured on recognition rate and time to recognition by call type. Risks include alert fatigue, dispatchers ignoring alerts, and weaker model performance in other languages."},"roleDetails":{"sales":{"customerProblem":"Dispatchers miss some out-of-hospital cardiac arrests on calls, delaying CPR instructions.","stakeholders":["911/PSAP director","EMS medical director","Quality assurance lead","CAD vendor"],"discoveryQuestions":["What share of your confirmed arrests were recognized on the call?","Do you audit missed-arrest calls by caller type?","How would call-takers be told to handle an AI alert they disagree with?"],"valueHypothesis":"An ML second listener may catch some arrests dispatchers miss, mainly with lone lay callers (interpretation).","potentialEngagement":"A bounded retrospective audit of local cardiac arrest call recordings against an ML model.","unsupportedClaims":["Lives saved or faster CPR","That dispatchers will act on alerts","Performance in English-language U.S. centers"]},"engineering":{"fit":"Fits PSAPs with recorded call audio and EMD protocols for cardiac arrest.","architectureIntegration":"Real-time speech model on call audio producing alerts in the call-taker's screen or CAD.","prerequisites":["Access to call audio streams","Local labeled cardiac arrest calls","CAD or call-handling integration"],"constraints":["2018 Swedish data","Small discordant sample","Preprint status","Language and protocol differences"],"security":"Live caller audio needs secure processing and retention limits matching recording policy.","proofOfValue":["Retrospective recognition comparison on local calls","Error breakdown by CPR-in-progress and multi-caller calls","Time-to-recognition measurement"]},"delivery":{"work":"Audit local calls, set alert handling protocols, then trial silent-mode alerts before live use.","dependencies":["Call recording access","Medical director protocol","QA reviewers"],"ownership":"PSAP director with the EMS medical director.","skillsAdoption":"Call-takers learn how to treat and document AI alerts.","governanceCheckpoints":["Medical protocol approval","Silent-mode review before live alerts","Quarterly missed-arrest audit"],"acceptanceCriteria":["Proposed: higher combined recognition than dispatcher baseline in silent mode","Proposed: no added call time"],"risks":["Alert fatigue","Dispatchers ignoring alerts","Weaker accuracy in other languages"]}},"retrievedAt":"2026-10-23T22:05:03-05:00","enrichedAt":"2026-10-23T22:08:44-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Dispatchers outperform ML in multi-caller and CPR-in-progress calls, so the human role remains central.","procurementImplications":"Buyers should ask vendors for failure-mode analysis by call type, not only overall sensitivity.","operatingModelImplications":"Supports a complementary model with ML as a second listener rather than a replacement.","sourceVerification":{"responseId":"curl-rs-2026-10-23","openedUrl":"https://www.researchsquare.com/article/rs-11153391/v1","referenceExcerpt":"OHCA was recognised by the ML model but missed by medical dispatchers in 58% (52/90) of calls","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}