Public Safety & Emergency ServicesIssue 446 records
Emergency Services
Tonight's edition is about the gap between promising AI results and the evidence needed to trust them. Jeju, South Korea says AI reading of ambulance heart tests, relayed to doctors, nearly doubled field ECG use and cut transfers between hospitals to 0.6%, though those figures come from the program and its vendor. A Barcelona park's AI smoke cameras cut useless alerts by 86% after local tuning, with people checking each one. Portugal is funding drone swarms that aim to reach fires within 10 minutes, with tests a year away. Korea's local-government IT institute now recommends AI that analyzes 119 calls in real time. But two recent studies urge caution: dispatch AI models almost never get tested outside their home data, and general chatbots misread ambulance ECGs.
- Evidence records
- 6
- Cross-source patterns
- 2
- Evidence classes
- 2 vendor claim2 academic research1 standards or public-body guidance1 independent reporting
- Outcomes
- 3 emerging2 cautionary1 effective
- Source freshness
- 4 new this fortnight2 recent
- Research completed
- 2026-10-21
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Reported AI gains came from locally tuned systems with a person checking each output
Jeju's ECG results run through an online physician who decides destination, and the program changed steps and training at the same time; Collserola's alert burden fell only after weeks of local tuning, and specialists validate every plume before alerting. The dispatch scoping review finds models rarely tested outside their development data and calls for local validation and recalibration. Together they suggest gains are system- and site-specific, not properties of the algorithm alone.
Operating questionWhat local tuning period, human check and before-after baseline will we require before crediting AI with an improvement?
Supporting evidenceJeju Special Self-Governing Province Fire Safety Headquarters (with Jeju National University Hospital and ARPI)Mobile World Capital Barcelona, Collserola Natural Park Consortium, Barcelona Metropolitan AreaYonsei University College of Medicine
Guidance promoting AI for 119 and EMS decisions is moving faster than external validation
Korea's KLID guide names real-time 119 call-audio analysis as a key service without stated validation requirements, while the scoping review finds no externally validated dispatch model among 11 development studies and inconsistent trial benefits, and the Heart & Lung study shows general-purpose LLMs give unreliable prehospital ECG calls. Buyers need task-specific, local evidence before deployment.
Operating questionWhat external validation and task-specific evidence must a 119 or EMS AI tool show before it is allowed to influence a live decision?
Supporting evidenceKorea Local Information Research and Development Institute (KLID)Yonsei University College of MedicineLSU Health Sciences Center-Shreveport
Evidence ledger
Full record · every source keeps its link and limitations
Jeju 119 says AI ECG reads relayed to physicians nearly doubled ambulance 12-lead ECG use and cut cardiac re-transfers to 0.6%
Seoul Shinmun (published 2026-10-20 10:00 KST) reports that Jeju Special Self-Governing Province's Fire Safety Headquarters won the top (presidential) prize at the 38th 119 Fire Policy Conference on Sept 17 for an AI-based severe cardiovascular emergency system built on ARPI's ECG Buddy AI ECG analysis. Built from 2024 with Jeju National University Hospital, the province and ARPI, it was piloted by 11 ambulance teams from January 2025 and has run across all 32 Jeju ambulance teams since March 2026. Paramedics record a 12-lead ECG and send the AI result and patient details to an online medical-direction physician, who decides the destination; the process went from seven steps to three. Per Jeju Fire and partners, 12-lead ECG use rose from 47.2% (2024) to 89.7%, hospital transmission reached 98.5%, direct transport to PCI-capable hospitals was 96% (March-May 2026), and re-transfer of suspected severe cardiac patients fell from 4.9% to 0.6% (one patient); 13 of 14 re-transfers in 2024 did not use ECG Buddy. A hospital team's 2025 pilot analysis found teams using it had 28.3% higher ECG use and 38.2% more medical-direction requests, 90% PCI within 90 minutes of arrival, and AUROC 0.94. A data platform linking EMS and hospital records is planned.
Why it matters, evidence and limitations
- Why it matters
- A rare EMS AI deployment reporting process and transport outcomes across an entire provincial fleet. U.S. EMS agencies and regional STEMI systems can use it to frame what to measure, while recognizing the numbers are operator-reported before-after figures.
- Evidence and measured results
- Newspaper report of operator, hospital and vendor figures plus a hospital pilot analysis summarized without methods; no independent evaluation.
- Limitations and uncertainty
- Figures are operator- and vendor-supplied (photos credited to ARPI; ARPI CEO quoted) and repeated in near-identical Digital Times coverage; before-after comparison against 2024 with no control and concurrent protocol changes (seven to three steps, medical direction, training); re-transfer rate rests on very small counts; AUROC target and denominators not given; no adverse events, false positives or survival outcomes reported; Korean-language source; single island province with one regional cardiac center.
General-purpose LLMs misread prehospital ECGs for cath-lab activation; monitor algorithm outperformed all three
A retrospective single-center study in Heart & Lung (published online Sept 15, 2026; LSU Health Sciences Center-Shreveport) tested three general-purpose multimodal LLMs (ChatGPT-5, Claude Sonnet 4, Gemini 2.5 Pro) on 615 prehospital 12-lead ECG images from 270 Shreveport EMS encounters (2019-2024) with suspected acute MI, against cardiology STEMI-pathway activation as the reference (101 activations, 232 positive ECGs). Gemini had 95.3% sensitivity but 9.4% specificity; ChatGPT 68.1%/42.3% and Claude 67.2%/46.5%. The monitor's built-in Glasgow algorithm was more balanced at 67.7% sensitivity and 64.2% specificity, with the highest accuracy (65.5% versus 41.8-54.3%); LLM PPV ranged 38.9-43.2%. Authors conclude general-purpose LLMs are not appropriate for cath-lab activation decisions and should be limited to education and documentation, and stress the results do not generalize to dedicated AI-ECG models. Why now: tonight's Jeju result credits a purpose-built ECG AI; this study shows why buyers must distinguish dedicated ECG tools from general chatbots that paramedics can reach on a phone.
Why it matters, evidence and limitations
- Why it matters
- EMS agencies need policy on clinicians using consumer chatbots for ECG reads; this U.S. study gives medical directors evidence to prohibit it while evaluating purpose-built tools separately.
- Evidence and measured results
- Peer-reviewed retrospective diagnostic-accuracy study with open dataset on Figshare.
- Limitations and uncertainty
- Single-center, retrospective, zero-shot prompt, static cropped images without clinical context; encounter-level labels applied to individual ECGs; reference standard is activation decisions, not angiographic occlusion for non-activated patients; no repeat runs to measure variability; published five weeks before this edition; opened via WebFetch because box curl returned 403, so publisher excerpt checks may fail.
Scoping review: AI for EMS call-taking and dispatch lacks external validation, and two randomized trials show no consistent benefit
A PRISMA-ScR scoping review in MDPI Healthcare (vol. 14, issue 19, article 3176; Crossref publication date Sept 24, 2026) by Yonsei University authors searched six databases for January 2021 to August 2026 and included 18 studies from 13 countries on AI in EMS call-taking and dispatch; 11 focused on patient recognition and classification, five on dispatch support. Eleven were retrospective model-development studies with internal validation only, and none of these conducted independent external validation or calibration assessment; all 11 were rated high risk of bias under PROBAST+AI. The two randomized trials showed no consistent improvement: in Denmark, ML cardiac-arrest alerts had higher sensitivity (85.0% vs 77.5%) but PPV of 17.8% vs 55.8% and did not improve dispatcher recognition (93.1% vs 90.5%, p=0.15); Sweden's MADLAD showed modestly better low-acuity prioritization (68.3% vs 62.5%) with high risk of bias. The authors call for prospective multicenter studies, local recalibration, drift monitoring, dispatcher override authority and auditable logs. Why now: most recent synthesis of dispatch AI evidence, directly relevant to tonight's Korean guide recommending real-time 119 call-audio AI.
Why it matters, evidence and limitations
- Why it matters
- 911 and EMS dispatch leaders weighing AI call-audio or triage tools can use it as a checklist of what evidence vendors should show: external validation, calibration and operational outcomes.
- Evidence and measured results
- Peer-reviewed PRISMA-ScR scoping review with risk-of-bias assessment.
- Limitations and uncertainty
- No registered protocol; English search terms only; conference papers and preprints excluded, so fast-moving technical work is missed; risk-of-bias assessment by one reviewer and added during peer review; funded by a Korean fire/EMS technology program; MADLAD trial already archived separately; published about four weeks before this edition; opened via WebFetch because box curl returned 403.
Korea's local-government IT institute issues AI service guide naming real-time 119 call-audio analysis and wildfire and flood prediction as key disaster services
Safe Times (published Oct 19, 2026) reports that the Korea Local Information Research and Development Institute (KLID) has published and distributed to all local governments an 'AI service guide for efficient local-government AX', meant to let local governments pick, adapt or combine AI services by local conditions and budget and to overcome fragmented, one-off adoption. In disaster and safety, it presents as key services AI that predicts wildfire and flooding risk in advance and AI that analyzes 119 emergency-call audio in real time. Other domains include 24-hour AI civil-complaint counseling and work assistants, at-risk household detection, intelligent traffic signals and recycling robots. KLID also proposes a standard AI and data ordinance, tailored consulting and standardized distribution of verified AI solutions, and plans visiting consultations for interested local governments.
Why it matters, evidence and limitations
- Why it matters
- Shows a national body packaging disaster AI as standard, reusable services for local governments, a model U.S. states and associations might follow, but with no stated validation requirements.
- Evidence and measured results
- News report of a public-institute guidance document; guide not opened.
- Limitations and uncertainty
- Short news report of a guidance document; the guide itself was not opened and no evidence base, accuracy figures, costs or safeguards for the disaster services are described; 'verified solutions' is not defined; Korean local-government context with national 119 fire agency structure; Korean-language source.
Barcelona's Collserola park AI smoke-detection pilot confirmed 150 incidents and cut irrelevant alerts 86% after local tuning, with human validation before each alert
Mobile World Capital Barcelona announced on Oct 14, 2026 results of a Barcelona Horizon Award pilot with the Collserola Natural Park Consortium and the Barcelona Metropolitan Area using Pantera, an AI and camera system from Brazilian company Umgrauemeio, on the park's watchtowers. Between July 1 and Sept 17 it detected 1,036 potential smoke plumes and confirmed 150 incidents, with smoke detected up to 43.5 km away; 103 of the 150 (68.7%) were outside park boundaries. AI flags plumes within seconds and runs a second check; a specialist team validates visually before an alert goes to the park team, with an average of 3.3 minutes from detection to alert. During an initial adaptation period, non-relevant alerts fell 86%, from 207 to 30 per week, by tuning to local reflections, urban lighting and industrial emissions. Why now: not previously archived and it is one of the few pilots publishing a local false-alert reduction figure, relevant as agencies weigh AI camera claims this season.
Why it matters, evidence and limitations
- Why it matters
- Gives wildland-urban interface agencies a concrete local-tuning figure and a human-validation workflow to compare against U.S. AI camera programs whose false-alert burdens are rarely published.
- Evidence and measured results
- Pilot organizer press release with operational counts; no independent evaluation.
- Limitations and uncertainty
- Pilot organizer's press release on an award-funded pilot; no independent evaluation; 'incidents' not defined and no comparison with how many fires were first reported by people or existing lookouts; residual 30 non-relevant alerts a week still require staff review; no cost or staffing data; urban-edge Mediterranean park; published six days before this edition.
Portugal launches €9.8M Fire Resist pilot for sensor and AI drone swarms aiming to reach fires within 10 minutes, with first tests about a year away
Portugal Resident (Oct 20, 2026, citing Lusa) reports Portugal's government presented Fire Resist, a €9.8 million pilot led by Drivetel that will use forest surveillance towers, smart sensors and AI-equipped drone swarms carrying up to 50 kg of retardant, aiming to reach a potential fire within 10 minutes; the agriculture minister said firefighting services take an average of 30 minutes to reach the scene. AI algorithms will analyze sensor data to distinguish fire outbreaks from fog or mist. Drivetel's chair said first tests will probably occur in about a year; the area is roughly 6,000 hectares in Serra do Marao, implementation began in September and ends in August 2029, with UTAD, CoLAB ForestWISE, Fraunhofer Portugal AWAM and several companies. Asked about cyberattacks, the chair said data processing is secure; equipment will rely on solar or wind where there is no grid.
Why it matters, evidence and limitations
- Why it matters
- An early signal on autonomous initial attack with AI detection; U.S. wildfire agencies can track it, but the project has no results and raises airspace, coordination and cybersecurity questions.
- Evidence and measured results
- News agency report of a government presentation and company statements; no test results.
- Limitations and uncertainty
- Announcement of a funded project with no tests yet; 10-minute goal is aspirational; 30-minute average is a ministerial statement without source data; cybersecurity assurance is an unsupported company claim; autonomous retardant drops raise airspace and coordination questions not addressed; Portuguese national context.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.
- Standards or public-body guidance
- Normative or advisory guidance from a standards body or public institution.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.