AI could help trauma teams make more accurate triage decisions before severely injured patients even reach the emergency department. Researchers at the University at Buffalo have shown that a large language model (LLM) can extract clinically relevant information from emergency medical services (EMS) calls, reducing communication errors and improving trauma team preparedness. Their findings were based on pediatric trauma cases.
According to the researchers, AI is not intended to replace clinicians but to function as a decision-support tool that filters critical information from often chaotic prehospital communications.
High stakes in trauma care
Trauma triage begins long before a patient arrives at the hospital. Emergency physicians often rely solely on a brief radio or telephone report from paramedics at the accident scene to determine how many specialists and resources should be mobilized. This process is vulnerable to errors. Critical details may be omitted, vital signs described imprecisely, or information misunderstood because of poor radio connections, background noise or time pressure. Such communication failures can lead to undertriage, where seriously injured patients do not receive sufficient emergency resources, or overtriage, where more personnel and interventions are activated than necessary.
"In most medical settings, the danger lies in making the wrong diagnosis," says lead author Ascharya K. Balaji, now a surgical resident at Tripler Army Medical Center. "In prehospital trauma, the bigger danger is that critical information never reaches the right people in time." The researchers evaluated the technology using 133 pediatric trauma activations. Pediatric trauma presents unique challenges because children often compensate physiologically for severe injuries longer than adults, making life-threatening conditions more difficult to recognize during the initial assessment.
Filtering signal from noise
Large language models are particularly well suited to processing unstructured language. The research team investigated whether an LLM could transform lengthy EMS reports into concise, clinically meaningful summaries. The AI system analyzed call transcripts, extracted essential medical information, including injury mechanism, vital signs, level of consciousness and signs of bleeding, and generated a structured summary together with a recommended trauma activation level.
The results showed that the LLM reduced transcript length by approximately 80 percent while preserving clinically relevant information. Researchers also found that more than 98 percent of the words spoken during EMS calls were not directly related to medical decision-making, highlighting the challenge clinicians face when identifying critical details under time pressure. "Our goal was to provide clinicians with a cleaner, more actionable signal from the same noisy input," the researchers note.
Supporting better clinical decisions
The study found that the LLM achieved triage accuracy comparable to experienced trauma clinicians. While that alone represents only a modest improvement, the researchers identified a more significant benefit. When physicians first made an incorrect triage decision and were subsequently shown the AI-generated recommendation, they were three times more likely to revise their decision correctly.
According to senior author Peter C. W. Kim, vice chair for research and innovation in the Department of Surgery at the Jacobs School of Medicine, this demonstrates the greatest potential of LLMs in trauma care. "LLMs are promising cognitive aids, not replacements for clinical judgment," Kim says. "They can match or slightly exceed human accuracy in interpreting EMS communications and, more importantly, help clinicians make better decisions when used alongside their own assessment."
The researchers envision the technology functioning as a real-time communication assistant. During an EMS call, the AI would automatically process the incoming information, produce a structured clinical summary and recommend an appropriate trauma activation level. The attending physician would then review, modify or reject that recommendation.
Toward safer emergency care
Although the study focused on pediatric trauma, the researchers believe the concept could be applied more broadly across emergency medicine. By reducing communication errors during one of the most time-critical phases of care, AI could help hospitals prepare more effectively before patients arrive.
The team emphasizes that human oversight remains essential. Rather than automating triage decisions, future AI systems are expected to support clinicians by handling one of the most difficult aspects of emergency medicine: rapidly separating clinically relevant information from the large amount of nonessential communication generated during emergency calls.
If validated in larger clinical studies, communication-aware AI assistants could become an important addition to trauma workflows, improving both patient safety and the efficiency of emergency departments.
AI triage helps breastcancer diagnosis
A couple of weeks ago researchers from the University of California, San Francisco (UCSF) and UC Berkeley showed that an AI triage tool can also dramatically shorten the time women with abnormal mammograms wait for follow-up care. Using the open-source AI model Mirai, trained on hundreds of thousands of mammograms, the team identified women at the highest risk of breast cancer and prioritized them for immediate assessment.
Among more than 4,100 screening mammograms, the system flagged 525 high-risk patients, enabling many to receive same-day diagnostic imaging and, when needed, biopsy. As a result, diagnostic evaluation times dropped from several weeks to around one hour, while the average wait for a biopsy among women diagnosed with breast cancer fell from more than two months to fewer than ten days. Rather than replacing radiologists, the AI acts as a triage tool, supporting personalized, risk-based screening pathways that accelerate diagnosis without overwhelming clinical capacity.
