A study published in NEJM AI is giving hospital administrators one of the clearer numbers they’ve had yet on whether AI actually speeds up emergency care: after three emergency departments rolled out an AI-informed triage clinical decision support tool, the median time from patient arrival to being placed in an initial care area dropped 33%, from 12 minutes down to 8. The analysis covered 174,648 total ED visits — 83,404 before the tool went live and 91,244 after.
The Basic Problem the Tool Is Trying to Solve
Emergency department triage has historically depended on a nurse’s five-to-ten-minute assessment using scoring frameworks like the Emergency Severity Index (ESI), a system that assigns patients a 1-through-5 acuity level largely based on vital signs, chief complaint and clinical judgment under time pressure. Studies have long shown meaningful variability between nurses using the same framework on the same patient, and missed high-acuity patients sitting in overcrowded waiting rooms have been tied to preventable harm in emergency medicine literature for years.
What the Numbers Showed
Beyond the arrival-to-care-area time drop, the study found the proportion of patients who actually required critical care and were correctly assigned a high-acuity triage level (ESI level 1 or 2) rose from 78.8% to 83.1% after the AI tool was introduced. Median time to ED disposition — the point when a decision is made to admit, discharge or transfer a patient — fell 4.2%, from 190 to 182 minutes, and time to ED departure dropped 6.1%. Those are modest percentage shifts, but at the scale of nearly 175,000 visits, they translate into thousands of patients moved through the system meaningfully faster.
How the Tool Actually Works Alongside Nurses
The AI system doesn’t replace the triage nurse; it functions as a clinical decision support layer that reviews the same intake information — vital signs, presenting complaint, medical history flags — and generates a suggested acuity level in real time, which the nurse can accept, override or use as a second opinion. That design mirrors a broader pattern in AI-assisted triage research: models consistently outperform traditional ESI scoring on predictive accuracy for hospital admission, ICU transfer and mortality risk, according to a body of comparative studies reviewed in emergency medicine journals over the past two years.
Where Nurses and the Algorithm Disagreed
The study’s authors flagged one complication that keeps this from being a clean win: variability in how often nurses agreed with the AI’s suggested triage level. In cases of disagreement, it wasn’t always clear whether the nurse’s judgment or the algorithm’s output was the more accurate call, and the researchers said harmonizing that gap remains an open challenge. Skeptics of AI triage point to this exact issue as the crux of their concern — that even a statistically improved system will occasionally clash with an experienced nurse’s read on a patient, and hospitals need clear escalation protocols for those moments rather than defaulting to either the human or the machine automatically.
The Case for Moving Fast Anyway
Proponents argue that emergency departments are exactly the setting where marginal speed improvements save lives, since delayed recognition of sepsis, stroke or cardiac events is one of the most common root causes of ED-related preventable deaths. A 33% cut in time to initial care, if it holds up across more hospital systems, could be one of the more consequential AI deployments in emergency medicine simply because it touches nearly every patient who walks through the door, not just a narrow diagnostic subgroup.
What’s Next for AI in the ER
The three-site study is likely to accelerate interest from health systems that have been cautious about triage automation, but it also raises the obvious follow-up questions: does the 33% time reduction hold in under-resourced or rural EDs with different staffing ratios, and does faster triage actually change downstream clinical outcomes like mortality or readmission, not just process metrics like wait times? Expect more multi-site trials over the next year aimed at answering exactly that, along with pressure from hospital administrators to deploy triage AI faster given how directly it appears to relieve overcrowding, one of the most persistent operational headaches in American emergency medicine.
Photo: ArtisticOperations / PIXABAY via Pixabay
