AI triage in the emergency room: what the studies say

Artificial intelligence is not yet ready to handle emergency room triage on its own. That is the conclusion of a meta-analysis published in 2026 in the journal BMC Emergency Medicine: language models did well at identifying who was not severe, but failed to recognize about 39% of the most severe patients. At the same time, AI has entered government plans for the lines in the SUS (Brazil's public health system), and BBC News Brasil heard experts on what that requires.

What happened

Researchers at Zhejiang University, in China, pooled 11 studies, with 3,088 real or simulated emergency room cases, evaluated between 2023 and 2025. They compared the classification made by large language models with the reference standard used in the studies.

The numbers show both sides. For identifying patients in the highest severity category, the pooled sensitivity was 61% and the specificity 97%. In practice, when the AI said a case was severe, it was almost always right. But of every ten truly severe patients, about four were not recognized as such. The authors conclude that current models are not ready for autonomous use in triage and may play a supporting role, provided they are validated in real-world, prospective use.

On September 25, 2026, BBC News Brasil showed that candidates from different political camps had proposed using AI in the SUS, including to replace first-come, first-served lines with risk-based triage. The report heard from researchers. Alexandre Chiavegatto Filho, of the University of São Paulo's School of Public Health, sees great potential, especially for bringing support to regions without specialists. Wagner Meira Jr., of the Federal University of Minas Gerais (UFMG), pointed to the central limit: "The problem is the false negative." For him, a false positive generates cost; a false negative may leave without care precisely those who needed it most.

The report also raised two practical obstacles. Changing the order of the line does not help if care capacity stays the same. And, according to the Ministry of Health, there is still no consolidated national figure for how many people are waiting for care in the SUS; a 2025 ordinance began requiring the standardized submission of these data to the National Health Data Network, which is being implemented.

Why it matters

Triage decides who is seen first. An error on the high side spends resources; an error on the low side can cost a life. That is why the right question is not whether AI is good, but what it is for and who answers for the result.

Brazilian regulation has already given part of the answer. CFM Resolution 2,454/2026, in force since August, treats AI as support, requires medical supervision and prohibits delegating to the system the communication of a diagnosis or therapeutic decision. The studies and the rule point in the same direction: the machine organizes, the professional decides.

In Bairo's view

For Bairo Leandro Gonzalez Martinez, the meta-analysis is an argument in favor of the design he defends, not against the technology. In his reading, AI delivers more when it does what it does well, listening patiently, organizing the history, remembering what changed, and leaves the decision to those with the training and responsibility to make it.

That is how he thinks about Hospital Triage: the clinical assistant talks with the person before the consultation and hands everything over, organized, to the doctor or nurse, who confirms, corrects or discards it. When faced with a warning sign, the conversation stops and the person is directed to 192 (Brazil's emergency ambulance number). Bairo sees in this a way to broaden access for those who today do not know which door to knock on, without putting anyone at risk through a false negative. At Xperienc Global Labs, the same logic applies to the clinical second opinion: an extra set of eyes for the doctor, not a diagnostic report.

And when the emergency is of the soul, he recalls the human door: CVV (Brazil's emotional support service) answers at 188, and the pages on pain, faith and reconnection show when to seek professional help.

Where this meets the ecosystem

Hospital Triage does not give a diagnosis, does not show a severity classification to the patient and does not decide anything on its own. The suggested priority goes to the professional, who confirms or corrects it. The system was created with CLAIN, the ecosystem's agent orchestrator, and exists as a prototype in a test environment, with no commercial operation. Before any real-world use, it will have to go through the risk assessment the CFM rule requires and through prospective validation, as the authors of the meta-analysis call for.

Sources

• Cui L, Wang M, Xu Y et al., "Diagnostic accuracy of large language models for emergency department triage: a systematic review and meta-analysis", BMC Emergency Medicine, 2026: pmc.ncbi.nlm.nih.gov
• BBC News Brasil, "IA vai decidir sua vez na fila do SUS? Propostas de candidatos da esquerda à direita não são simples de implementar" (Will AI decide your turn in the SUS line? Proposals from candidates from left to right are not simple to implement), 9/25/2026: bbc.com/portuguese
• Federal Council of Medicine, "CFM normatiza uso da IA na medicina" (CFM regulates the use of AI in medicine), Feb. 2026: portal.cfm.org.br

Read also

• CFM Resolution 2,454/2026: AI in medicine has rules
• WHO: the world will be short 11.1 million health workers
• The future of triage