Trauma triage is challenging: A UB study assesses how AI might help improve accuracy

BUFFALO, N.Y. – Making triage decisions in a busy emergency room is tough enough; making triage decisions based on a hurried call with emergency medical staff responding at the scene of an accident is all the more challenging.

Surgeons and trainees in the Jacobs School of Medicine and Biomedical Sciences at the University at Buffalo wondered if artificial intelligence, specifically large language models (LLMs), could help improve the accuracy of the initial emergency medical services (EMS) information, which would, in turn, better prepare the hospital trauma team. Increased accuracy could reduce the risks of both undertriage — when a patient with severe injuries doesn’t receive sufficient emergency intervention — and overtriage — when a patient receives more intervention than they require.

They put an LLM to the test, using 133 pediatric emergency department activations. Their results were published online June 12 in the Journal of the American College of Surgeons. 

The importance of accurate information is especially critical in the prehospital setting, says Peter C. W. Kim, MD, PhD, corresponding author, vice chair for research and innovation in the Department of Surgery in the Jacobs School and head of its Data Augmented Research Technology in Surgery (DARTS) laboratory, which connects surgeon innovators with machine learning scientists.

Lead co-authors on the paper are Ascharya K. Balaji, MD, a 2025 Jacobs School alumnus who is now a surgery resident at the Tripler Army Medical Center, and Brendan T. Fox, MD candidate in the Class of 2028.

“In most medical settings, the danger lies in making the wrong diagnosis,” says Balaji. “In prehospital trauma, the bigger danger is that critical information never reaches the right people in time.”

Kim explains that when a paramedic calls ahead to the emergency department, that brief, often chaotic radio or phone report is the only data the receiving team has. “If the paramedic forgets to mention a dangerous mechanism of injury or vital sign, uses vague language like ‘vitals look okay,’ or if the physician mishears something, the trauma team may not mobilize the right resources before the patient arrives,” he says.

The researchers add that at the often chaotic scene of an accident, paramedics may make instinctive calls rather than systematically checking every clinical box. That subjectivity, compounded by noisy radio connections, speech disfluencies, and time pressure, creates fertile ground for miscommunication.

“By the time a patient arrives undertriaged — meaning the team wasn’t adequately prepared — precious minutes are already lost,” says Fox.

That reality is what drove the DARTS team to see if using an LLM could help improve the accuracy of prehospital triage.

The researchers decided to focus their LLM project on pediatric emergencies, which are especially challenging because children’s bodies respond to injury differently from adults. Children’s responses also differ as they age.

“What looks like stable vital signs in a child can mask serious internal bleeding far longer than in an adult,” Kim explains. “Children’s physiology compensates more aggressively before crashing suddenly. This means that the standard warning signs paramedics are trained to recognize in adults can be misleading or absent in children. Adding to this, paramedics have far less exposure to pediatric trauma cases simply because children are injured less often than adults, so their pattern recognition for kids is less practiced.”

LLMs are essentially sophisticated language processors.  They excel at reading, interpreting and summarizing messy, unstructured text or speech, which, the researchers say, is basically the definition of what a prehospital trauma phone call produces.

“A human physician listening to that call in real time has to mentally filter all of that noise while simultaneously preparing the trauma team,” Kim explains. “An LLM can rapidly process the transcript of that same call, strip away the nonessential content, extract the clinically important elements — mechanism of injury, vital signs, mental status, bleeding indicators — and deliver a structured summary with a recommended triage level.”

The paper shows the LLM compressed transcripts by about 80% while preserving accuracy, providing clinicians with what the researchers say is “a cleaner, more actionable signal from the same noisy input.” Interestingly, the paper reports that more than 98% of the words in call transcripts were nonmedical.

The LLM achieved triage accuracy that was comparable to trauma staff. That may not sound like a dramatic improvement, but the research revealed that when trauma staff were exposed to LLM output after making an incorrect triage decision, it tripled their odds of correcting that decision.

“Right now, LLMs are promising cognitive aids, not replacements for clinical judgment,” Kim says. “What the results do show is that LLMs can match or slightly exceed human accuracy in interpreting EMS communications, and — perhaps more importantly — that when human clinicians see an LLM recommendation alongside their own thinking, they make better decisions.”

Currently, the researchers say the value of the LLM is as a “communication-aware” triage assistant that listens to or reads the EMS report in real time, produces a structured clinical summary, and offers a recommended activation level that the physician then accepts, modifies or overrides.

“Human oversight stays in the loop and the LLM handles the signal-from-noise problem that humans struggle with under time pressure,” Kim concludes.

UB co-authors are Akhil Gorugantu, a graduate student majoring in computer science; Philip Seger, DARTS Innovation Fellow; Andrew Nordin, MD, clinical assistant professor, and Steven D. Schwaitzberg, MD, SUNY Distinguished Service Professor and Chair, all of the Department of Surgery.

Additional co-authors are Tiffany Fabiano and Gene Yang of the Golisano Children’s Hospital of Buffalo, and Sharifa Himidan of the Rainbow Babies and Children’s University Hospitals and Case Western Reserve University.