AI in Emergency Triage: How Hospitals Use Algorithms to Prioritize Care

Last updated: October 2026

Every emergency department runs on a quick, high-stakes judgement made at the door. Within minutes of arrival, a triage nurse has to decide who can safely wait and who needs a doctor now, often with little more than a set of vital signs, a short conversation and a busy waiting room in view. Most of the time that judgement is excellent. But triage scales leave room for disagreement, and in crowded departments small misjudgements can turn into long, risky waits.

That is the gap hospitals are trying to fill with algorithms. Some predict, at the moment of triage, how likely a patient is to need intensive care or admission. Others read CT scans and push suspected strokes to the top of the radiologist’s list. Others watch the electronic health record for early signs of sepsis. This article looks at each of those uses, what the strongest studies show, and why some widely deployed tools have performed far worse than their developers expected.

Why triage is a natural target for algorithms

Most emergency departments sort patients with a structured scale. In the United States the five-level Emergency Severity Index (ESI) is common; much of Europe, including many Spanish hospitals, uses the Manchester Triage System or the Spanish Triage System (Sistema Español de Triaje, SET). These scales are reliable at the extremes. A patient in cardiac arrest and a patient needing a prescription refill are easy to place.

The difficulty lies in the middle. In the Johns Hopkins study described below, more than 65% of visits were assigned to ESI level 3, a broad category that mixes patients who will go home after a few tests with patients who will end up in intensive care. That compression is exactly where a statistical model, trained on what happened to hundreds of thousands of earlier patients, can add information a single human glance cannot.

Four places algorithms enter the emergency pathway

These tools usually sit inside the hospital’s electronic health record or imaging system. Patients rarely see them directly, which is one reason transparency about their use is an ongoing debate.

What goes into a triage prediction model

Front-door models work with the information that exists in the first minutes of a visit. Their power comes less from any single variable than from learning how combinations of ordinary data points relate to outcomes across very large numbers of past visits.

Systematic reviews of the field find that relatively familiar methods dominate. Logistic regression, random forests and gradient-boosting models are the most common, with deep learning and natural language processing increasingly used to make sense of free-text complaints and nursing notes. The usual target is not the triage level itself but a hard outcome, such as death, intensive care admission, an emergency procedure or hospitalization, which the model then translates into a suggested priority.

Case study: electronic triage at Johns Hopkins

One of the most cited examples was built at Johns Hopkins and published in the Annals of Emergency Medicine. Researchers trained a machine learning model on vital signs, chief complaints and medical history, then compared its classifications with nurse-assigned ESI levels across 172,726 visits to urban and community emergency departments.

Its main contribution was splitting the crowded middle category. Analysing past visits, the algorithm retrospectively identified more than 10% of patients assigned to ESI level 3, 14,326 visits in total, as candidates for a higher triage priority. These patients had higher observed rates of hospitalization and of critical care admission or emergency procedures than the broader ESI level 3 group.

These findings show that the model could identify meaningful risk differences within a broad triage category. They do not, by themselves, prove that deploying the algorithm improves patient outcomes.

The model also identified patients who could potentially be assigned a lower priority. In principle, better risk stratification could help departments allocate attention and resources more efficiently. However, any claim about shorter waits, reduced bed use or improved outcomes needs implementation data from the clinical setting concerned. The tool was later deployed at The Johns Hopkins Hospital and Howard County General Hospital.

Imaging triage: reordering the radiologist’s list

A different kind of triage happens after the scan. In a large hospital, dozens of CT studies may be waiting to be read. Computer-aided triage and notification software (often abbreviated CADt) reviews images as soon as they are acquired and flags those likely to show a time-critical problem, so a specialist sees them sooner.

In February 2018 the FDA authorized the Viz.AI Contact application through its De Novo pathway. The software analyses CT images of the brain and sends a text notification to a neurovascular specialist when it detects a suspected large vessel blockage. According to the FDA, the company submitted a retrospective study of 300 CT images comparing the software’s performance with that of two trained neuroradiologists. The FDA also noted that the decision created a new device category, allowing later triage software with the same imaging purpose to follow the 510(k) route.

At the time of authorization, trade press reported sensitivity and specificity of about 90% for that study. These figures describe one retrospective dataset; the full regulatory summary (De Novo DEN170073) is the reference for the exact results, and performance in a given hospital depends on its scanners, protocols and patient population.

The key design principle is that triage software does not make the diagnosis. The radiologist still reads every scan, including those the software does not flag. Its job is to change the order of the queue and shorten the time to the right specialist for the patients who need it most. Similar tools now exist for intracranial haemorrhage, pulmonary embolism and other emergencies.

Sepsis alerts: a lesson in why validation matters

Sepsis is one of the most important conditions to catch early and one of the hardest to recognize, because its first signs overlap with many less serious illnesses. Automated alerts are now widespread, and two well-known systems show how differently such tools can perform.

These figures describe the Epic Sepsis Model at the alert threshold evaluated in this external validation (a score of 6, within the range the developer recommends). The 18% refers to hospitalizations that triggered an alert, not to patients with sepsis, who made up 7% of hospitalizations. Performance can vary with the threshold, patient population, clinical setting and definition of sepsis, so the results are not a universal estimate for every hospital or version of the model.

The Michigan study, published in JAMA Internal Medicine in 2021, became a landmark because the model was already in use at many hospitals. It found the tool identified only 7% of sepsis patients whom clinicians had not already treated promptly, while generating enough alerts to create a serious risk of alert fatigue. The developer disputed aspects of the analysis, including how sepsis onset was defined, but the study made external validation a central expectation for clinical prediction tools.

The TREWS results are more encouraging, but they need careful reading. The 3.3-point absolute reduction (95% CI 1.7–5.1) and the 18.7% relative reduction describe the same effect from two angles: the first is the drop in the mortality rate, the second is that drop as a share of the baseline risk. The findings support the potential value of timely clinical responses to sepsis alerts, but they do not establish that the alert system alone caused the reduction in mortality. Differences between patients whose alerts were confirmed promptly and those whose alerts were not may have influenced the results, despite statistical adjustment. Randomized evidence would be the stronger test.

Why more alerts do not always mean better care

A useful early-warning system must do more than identify patients who may deteriorate. It must also produce alerts at a manageable rate and help clinicians tell apart the patients who need immediate assessment from those who do not. If too many alerts fire without a meaningful clinical finding, staff may become less responsive to later warnings. Hospitals therefore need to assess sensitivity, positive predictive value (the share of alerts that turn out to be real cases), alert burden and clinical outcomes together, rather than judging a model by a single accuracy figure.

How regulators treat these tools

In the United States, software that analyses medical images, such as stroke triage tools, is regulated as a medical device by the FDA. Some prediction tools embedded in health records may fall under the FDA’s clinical decision support framework, where the level of oversight depends on what the software does and whether clinicians can independently review the basis for its recommendations.

In the European Union, emergency triage software intended for medical purposes may be subject to the Medical Device Regulation (MDR). Depending on its intended purpose and classification, it may also fall under the high-risk AI provisions of the EU AI Act. Requirements and implementation dates depend on the applicable legal framework, so hospitals should verify each tool’s regulatory status and intended use rather than relying on a general claim of compliance.

What this means for patients in the emergency department

If a hospital uses a triage algorithm, a nurse still assesses you and still decides your priority. The software may suggest a higher or lower level, and the nurse can accept or override it. Algorithms only work with the information they receive, so the most useful things a patient can do are the same as always.

• Describe your main symptom and when it started as clearly as you can.
• Mention medicines you take, especially blood thinners, and relevant conditions such as diabetes or heart disease.
• Tell staff if you are pregnant or recently had surgery.
• Report it straight away if your symptoms get worse while you are waiting.
• Ask if you are unsure why you have been given a particular priority.

For hospitals evaluating a triage tool

The evidence above points to a short set of questions any emergency department should be able to answer before going live: Has the model been validated on data from a hospital like ours, by people independent of the developer? How often will it alert, and who is expected to respond? How will we monitor its performance after launch, including across age, sex and ethnic groups? And what happens to the workflow when clinicians disagree with it?

Limitations of the current evidence

Most triage prediction studies are retrospective and test models on past records from the same hospitals that produced the training data, which tends to overstate performance. Area under the curve is the most commonly reported metric, but it says little about what happens at the specific alert threshold a hospital chooses, or about workload. Few studies are randomized, and fewer still measure patient outcomes such as mortality or harm from delayed care. Models can also drift as patient populations, documentation habits and treatment practices change, so a tool that worked well at launch may need recalibration. Finally, historical data reflect historical decisions; if some groups were under-triaged in the past, a model trained on that history can learn the same pattern unless it is specifically checked.

The Bottom Line

FAQ

Scientific Studies and Regulatory Sources

Peer-reviewed research

• Levin S, Toerper M, Hamrock E, et al. Machine-learning-based electronic triage more accurately differentiates patients with respect to clinical outcomes compared with the Emergency Severity Index. Annals of Emergency Medicine, 2018.
• Wong A, et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine, 2021. See also the accompanying commentary: The Epic Sepsis Model Falls Short—The Importance of External Validation.
• Adams R, et al. Prospective, multi-site study of patient outcomes after implementation of the TREWS machine learning-based early warning system for sepsis. Nature Medicine, 2022.
• Miles J, Turner J, Jacques R, et al. Using machine-learning risk prediction models to triage the acuity of undifferentiated patients entering the emergency care system: a systematic review. Diagnostic and Prognostic Research, 2020.

Regulatory information (accessed October 2026)

• U.S. Food and Drug Administration. FDA permits marketing of clinical decision support software for alerting providers of a potential stroke in patients, February 2018.
• U.S. Food and Drug Administration. De Novo classification request DEN170073: ContaCT (Viz.ai).

This article is for general information only and does not constitute medical advice. It does not describe how any specific hospital triages patients. In an emergency, call your local emergency number (112 in Spain and across the EU). Study results come from specific hospitals and time periods and may not reflect performance elsewhere. Helathora is not affiliated with any of the companies or institutions mentioned.