AI and the Future of Emergency Call Response: From Human Triage to Augmented Triage

Imagine this scene. It's 2:47 AM. A woman calls 118 for her husband: "he's not feeling well, he has some nausea." Her voice is calm, almost apologetic for the disturbance. The operator assigns a medium priority code. The ambulance arrives in 18 minutes. It's a heart attack.

This isn't a made-up case. It's the type of scenario that repeats itself every day in dispatch centers around the world — and what research calls undertriage: the critical patient who is underestimated because the signs aren't obvious enough, or because the caller doesn't know how to describe them.

The problem isn't the operator. It's structural. In those 90-120 seconds of a call, a human being must decipher panic, accents, background noise, symptoms described in fragments — and translate everything into a priority code, under pressure, often alone [1]. It's an extraordinarily difficult job. And it's exactly the point where artificial intelligence can make a difference — not by replacing human judgment, but by supporting it.

A paramedic, three hours, an idea worth reflecting on

James Oswald works as a Clinical Practice Guideline Specialist for Ambulance Victoria, in Australia. A few weeks ago he shared something unusual on LinkedIn: a prototype built in his spare time, in about three hours [2].

The system is simple in its design. It transcribes the emergency call, extracts 20 clinical parameters from the conversation using a language model, and then estimates the real clinical severity of the situation by comparing it with the priority code assigned by the AMPDS protocol. When the two judgments diverge significantly, the call is flagged for clinical review, with a graph showing the physician which parameters weighed most in the estimate — and by how much.

The most instructive result Oswald shows: a call coded as high priority by the standard protocol receives from the system an estimate of 5% probability of actually being a serious emergency. Not to cancel the code — but to say: this one is worth being rechecked by a clinician.

> "They don't have to be enterprise megasolutions. We can work within our constraints while still leveraging the benefits of AI." > — James Oswald, Ambulance Victoria [2]

Three hours. No dedicated budget. No special infrastructure. Just clarity about a real problem and the willingness to try.

What the scientific literature says

Before even talking about operational applications, it's worth understanding what the empirical basis for all this rests on.

A retrospective study presented at ESOC 2023 — the European reference congress for cerebrovascular diseases — trained a deep learning model on over 1.5 million calls to Copenhagen Emergency Medical Services between 2015 and 2020, of which more than 7,000 were related to stroke. The system transcribes the call audio and analyzes the text to predict stroke risk. The results are clear: sensitivity of 63.0% versus 52.7% for human operators, with an F1-score of 35.7 vs 25.8. The model outperformed human dispatchers in every analyzed subgroup — by sex, age, and type of stroke — without exception [3].

This isn't an isolated case. A narrative review published in Cureus in September 2025, which analyzed the literature on PubMed and PMC through August 2025, concludes that machine learning models consistently outperform traditional clinical scores in prehospital triage for trauma, stroke, and cardiac emergencies. In particular, ML models reduce undertriage below 10% — a significant result in contexts where in the USA undertriage rates for severe trauma exceed 30% [4].

A PRISMA-guided scoping review published in Big Data and Cognitive Computing (August 2025), based on 1,181 articles from the 2018–2025 period, identifies triage and prioritization as the fastest-growing areas in prehospital AI, with AUROC above 0.85 in multiple independent clinical contexts [5].

These data don't validate specific solutions — they validate an approach. And Oswald's approach is a practical, low-cost implementation with the same logical framework as peer-reviewed studies.

What's already happening in Italy

Oswald's case isn't isolated, even in our context.

ASL Benevento has launched AI 118 SmartPlanner, a generative artificial intelligence platform with mathematical models for real-time resource allocation, designed specifically for mountainous inland areas where the relationship between distances, tourist seasonality, and isolation makes every second even more critical [6].

Beta80 Group is experimenting with NLP for automatic telephone triage in 112 dispatch centers: assistance in form completion, management of non-emergency calls, construction of AI-guided decision trees. The protocols were validated by EENA in Portugal as early as 2019 [1].

The international landscape: hybrid systems already operational

Abroad, AI-dispatch integration is already an operational reality on a large scale.

Carbyne AI Triage prioritizes calls during volume peaks by analyzing audio in real time, while the operator maintains total control of the situation [7]. The system doesn't decide — it flags, always leaving the final word to the human.

The London Ambulance Service has integrated AI with MPDS on 2.4 million annual calls: it detects stroke, sepsis, and mental health crises in real time, with an 11% improvement in clinical accuracy without disrupting established protocols [8].

In the USA, RapidSOS leverages APIs to transmit GPS geolocation, scene context, and prioritization to 911 dispatchers, combining structured data with voice analysis for more precise dispatch even before the ambulance departs [9]. The impact of these technologies is also recognized at the federal level: the American NTIA has identified AI applied to 911 centers as a priority lever for reducing response times on a national scale [10].

---

Overview of main systems

---

The challenges we cannot ignore

It would be dishonest to present this scenario as free of obstacles. The real barriers are not primarily technical.

Regulatory and privacy. GDPR requires strict anonymization of voice data from emergency calls. The EU AI Act classifies decision support systems in high-risk contexts — such as health dispatch — in categories that require explicit validation and guaranteed human supervision [10].

Model transparency. Black-box systems are not acceptable in emergency settings. The clinician must be able to understand why the system produced that output, not just what it produced — which is why Oswald's prototype includes a graph showing the weight of each parameter in the decision [2]. The same Cureus review emphasizes how model transparency is one of the non-negotiable requirements for clinical deployment [4].

Bias and generalizability. A model trained on calls from an Australian or Danish EMS system might perform differently in an Italian context — different language, different call patterns, different prevalent pathologies. The 2025 PRISMA scoping review highlights that most available studies come from Europe and North America, with limited representation of other contexts [5]. Local clinical validation is essential, not optional.

Integration into workflows. Adding an AI system to an operator already under pressure can increase cognitive load instead of reducing it, if the interface is not carefully designed [1]. Training must evolve toward a new role: no longer just "protocol operator," but "AI supervisor."

The role of the prehospital professional

There's a question many ask, often quietly: will AI make the judgment of the responder or dispatch physician obsolete?

The answer, at least in the medium term, is no — but with a condition. The most effective systems are not those that try to replace clinical experience, but those that amplify it. A system that flags "this call coded HIGH has only a 5% probability of being a serious emergency" is not deciding in place of the clinician: it's delivering a quantified second opinion, built on thousands of cases, in a few seconds [2].

The professional who can critically read the output of these systems — understanding when to trust and when to override — will be significantly more effective than those who ignore them and those who follow them blindly. An additional advantage, often underestimated: these systems enable automated audits on thousands of calls and training on synthetic transcripts to simulate rare scenarios — silent strokes, CO poisoning in enclosed spaces — that rarely emerge in traditional training pathways [1].

Doing something, now

The most important message from Oswald's work — and from the literature supporting it — is not technological. It's operational: mega corporate solutions worth millions of euros are not needed. Ethical, transparent, low-cost prototypes are enough, built with clarity about the clinical purpose [2].

In Italy, the 118 and territorial emergency systems still face enormous structural challenges: resources, personnel, regional fragmentation. But ignoring this transformation is not a neutral choice — it's a choice with consequences.

The future of dispatch centers is not science fiction. It's a call that's already ringing. It's up to us to answer with operational vision.

---

References

Scientific literature

[3] Havtorn JD, Wenstrup J, Borgholt L, et al. A retrospective study on deep learning-enabled stroke recognition for a medical help line. Presented at the European Stroke Organisation Conference (ESOC) 2023, Munich, May 24, 2023. Press release: https://www.eurekalert.org/news-releases/989994 Medscape coverage: https://www.medscape.com/viewarticle/992515

[4] Artificial Intelligence in Prehospital Emergency Care: Advancing Triage and Destination Decisions for Time-Critical Conditions. Cureus, September 2025. https://www.cureus.com/articles/409861

[5] A Systematic Literature Review of Artificial Intelligence in Prehospital Emergency Care. Big Data and Cognitive Computing, 9(9):219, August 2025. https://doi.org/10.3390/bdcc9090219

Operational and industry sources

[1] Beta80 Group — Emergency prevention tools: leveraging AI in the 112 dispatch center https://news.beta80group.it/strumenti-di-prevenzione-emergenza-sfruttare-ai-nella-centrale-112 See also: How AI improves call handling in dispatch centers https://news.beta80group.it/come-intelligenza-artificiale-migliora-call-handling-nelle-centrali-operative

[2] James Oswald, Ambulance Victoria — LinkedIn post on AI AMPDS Emergency Call Prioritiser prototype https://www.linkedin.com/posts/james-michael-oswald_i-recently-built-this-tool-an-ai-ampds-activity-7449378183096987648-yhGm

[6] NTR24 — ASL Benevento focuses on AI to revolutionize 118 https://www.ntr24.tv/2026/03/05/lasl-benevento-punta-sullia-per-rivoluzionare-118-cronicita-nelle-aree-interne-e-inclusione-attraverso-la-musica/

[7] Carbyne — Emergency Call Triage https://carbyne.com/solutions/emergency-call-triage/

[8] LinkedIn / Andre — AI to improve emergency dispatch accuracy (London Ambulance Service) https://www.linkedin.com/pulse/ai-improve-emergency-dispatch-accuracy-technology-andre-p8oje

[9] RapidSOS — 911 with artificial intelligence (demonstration video) https://www.youtube.com/watch?v=34zBoKjGPIM

[10] NTIA — Improving 911 Operations with Artificial Intelligence https://www.ntia.gov/category/next-generation-911/improving-911-operations-with-artificial-intelligence