Harnessing the Power of AI in Emergency Triage: A Paradigm Shift
By Trisha Chakraborty, Global High School Fellow (South Pasadena High School ‘25)
Each year there are over a staggering 130 million Americans visiting emergency rooms across the nation. The first step in every person’s path at the emergency department is triage, or the preliminary assessment of a patient’s need for care. Fundamentally, it’s a system designed to maximize patient safety by prioritizing those who need the most urgent care. As emergency departments continually strive to improve patient safety and well-being, the development of artificial intelligence is paving the way for major leaps in emergency triage. Through extensive research and effective implementation, AI-powered triage models have demonstrated an incredible capacity for improving the way emergency departments are able to provide their patients with care. They prove useful in improving patient flow, enhancing risk stratification, and increasing the consistency by which patients are triaged
Performed by experienced healthcare professionals known as triage nurses, the stratification process is fairly quick, on average taking about 3 to 5 minutes. In the United States, the most commonly used system is the Emergency Severity Index (ESI), which differentiates patients into various acuity levels in accordance with hospital resources. In this five-tier system, level 1 is for the patients who require the most immediate care, and level 5 is for those whose lives would not be in danger if they waited. For instance, a patient arriving with complaints of abdominal pain might require blood work, medication, or a scan of their belly. That’s a total of about three resources. Whereas, a patient arriving with a twisted ankle would mainly require an X-ray, which is only one resource. In this instance, the patient with abdominal pain would be triaged higher than the patient with a twisted ankle because of their increased need for hospital resources.
As a concept, ESI seems rather straightforward. However in practice, ESI level 3 accounts for nearly 50–70% of all triage visits. According to Danielle Krupa, an ER nurse from LA County, “threes do tend to be a kind of catch-all”. Threes consist of varying conditions, and often the reasons for patients being placed at that level are due to natural subjectivities. Every nurse is different, as triage is heavily reliant on experience. Many of triage’s challenges include “having to have the clinical experience to appropriately assess patients as they come in, and being able to ask the right questions to narrow down a patient’s chief complaint,” says Krupa. Oftentimes, patients come in with trouble articulating their complaints and nurses face the challenge of narrowing it down to a specific chief complaint. Coupled with time restraint and the limited information nurses are presented with, emergency triage is highly variable and can result in inconsistent stratification.
With artificial intelligence brought into the picture, machine learning triage models can provide assistive tools for nurses to more accurately and consistently stratify patients. As patients are being treated on care pathways more aligned with their health concerns, patient flow is expedited and wait times decrease. Ultimately, artificial intelligence is aiding emergency departments in their mission to enhance patient safety and provide them with the best possible care.
Designed by researchers at Johns Hopkins, TriageGO is an artificial intelligence model under the biomedical tech company, Beckman Coulter, that serves as an auxiliary tool for nurses during triage. Its process begins with nurses collecting pertinent information on a patient and entering it into a computer in front of them. With the logged information, TriageGO is able to analyze a patient’s vital signs, complaints, symptoms, demographics, and medical history from their electronic medical health record. It then is able to leverage that data to determine a patient’s risk of needing critical care, having an emergent procedure, or being admitted overnight at the hospital. The compounded risk of these three outcomes is what ultimately fuels TriageGO’s recommendation, along with a brief explanation as to why it’s triaging a patient at that level. In a matter of seconds, nurses are provided with a valuable second opinion to aid their triage assignment.
TriageGO uses a subtype of artificial intelligence known as machine learning, in which a computer is able to provide descriptions, predictions, or suggestions without being explicitly programmed to do so. A very basic level example could be that you feed a machine learning algorithm various labeled pictures of plants, and it would eventually learn to distinguish and name them independently of the provided label. This form of machine learning, where the program is fed labeled datasets with given input-output pairs, is known as Supervised Machine Learning. TriageGO is powered by Supervised Learning in order to predict a patient’s risk of needing urgent care and provides suggestions regarding the triage level they should be assigned.
As of now, TriageGO has been implemented at multiple hospitals such as the Johns Hopkins Hospital, Johns Hopkins Bayview Medical Center, Howard County General Hospital, and many more. Post-implementation, the new triage process resulted in significant changes in the ESI level distribution. Although high-acuity levels 1 and 2 largely remained the same, it was changes in the mid and lower acuity levels that stood out. Level 3 patients decreased by 15%, and level 4 and 5 patients together increased by 56%. “If you can risk stratifying patients better, there are benefits in patient flow,” says Dr. Scott Levin, TriageGO’s co-founder and associate professor of emergency medicine at Johns Hopkins. “You can leverage more efficient care pathways for the non-sick patients.” Those benefits can be seen through a 35-minute reduction in door-to-admit-decision times for hospitalized patients and on average, a 61–82 minute acceleration of ICU transfers for high-risk patients. Improved patient flow not only results in lowered wait times but inevitably leads them out of the emergency department faster as well.
“Really the goal, at least in our context, is to partner the clinician with the AI, that the clinician and the AI together, is better than the clinician would be alone or with the tool alone — we’re trying to achieve this partnership,” explains Levin. “In large part we use data to show them why it is we think harmonizing nurses with the tools can be an improvement at triage, compared to the status quo.”
This partnership can be found in an additional instance of machine learning triage: KATE. Created under the company Mednition, KATE is an artificial intelligence triage tool with a mission to empower emergency triage nurses by drawing a second look at any anomalies in patient care. Similar to TriageGO, it’s integrated into an emergency department’s natural workflow and provides ESI-level recommendations for incoming patients.
KATE has the incredible power of providing early sepsis detection, an incredibly life-saving feature. Sepsis is a deadly medical emergency that is the result of the body’s extreme response to an infection. In the United States, there are over 1.7 million people diagnosed with sepsis yearly and nearly 270,000 die as a result of it. Sepsis is difficult to diagnose due to its high variety of symptoms, making it a lethal culprit once discovered. Therefore, early detection is imperative for increasing a patient’s chances of survival. After implementation, KATE’s machine learning model for early sepsis detection achieved a 50% improvement in comparison to the prior best practice screening tool, allowing patients to have improved odds against their battle with sepsis.
Currently being used by the emergency department at Adventist Health White Memorial Hospital, their medical director, Dr. Stephen Liu has high praise for the machine learning technology, “KATE is catching patients with sepsis at the door, without using lab results, that would have been otherwise missed.” He states, “Our nurses are more accurately identifying and initiating care for patients with Sepsis by using KATE than our prior screening protocol.”
TriageGO and KATE are both powerful tools, but their key to success lies in implementation. Before either product began playing a role in the triage process, the nurses that would use them went through a carefully orchestrated training process and their interactions with the AI triage tools underwent heavy observation. Like any other change management process, there were challenges that presented themselves. At the core of overcoming those hurdles was a strong emphasis on the partnership between the artificial intelligence model and the nurses that harnessed it.
Another way of understanding this stress on AI partnership is a paradigm shift from artificial intelligence to augmented intelligence. A term snowballing in popularity, augmented intelligence refers to the idea in which AI models enhance, or augment, human cognitive performance. In a field so heavily reliant on the human ability to be empathetic, caring, and responsive to gut feelings and instincts, jobs in healthcare and medicine will continue to be piloted by people. “For emergency triage applications, an AI algorithm can’t see the patient. So there are things that healthcare providers sense, that I don’t think can be replaced or will ever be replaced,” says Levin. “This will continue to be assistive technology that helps people make better decisions, leverage data better, make decisions faster, make them more efficient, but not necessarily take their jobs.” Augmented intelligence is a more mindful way of referring to artificial intelligence in the case when its goal is to be of assistance, not automation. TriageGO and KATE are remarkable instances of how a field can be enhanced with AI.
From improvements in patient flow and risk stratification via TriageGO to KATE’s life-saving capacity for detecting sepsis patients early, electronic triage is transforming the way patient care is being delivered in the emergency department. Machine learning models bring power in exposing patterns that are extremely valuable in clinical decision support, and they allow nurses to harness those recommendations in order to lead patients down safer and more expedited healthcare pathways.
