Editorial
May 2, 2025
Triage, the Tip of the Shears
N. Seth Trueger, Ari B. Friedman
JAMA Netw Open. 2025;8(5):e258508. doi:10.1001/jamanetworkopen.2025.8508
In this issue of JAMA Network Open, Sax and colleagues1 analyze the association of triage accuracy for emergency department (ED) patients with 3 critical diagnoses with time-dependent interventions: ST-elevation myocardial infarction (STEMI), subarachnoid hemorrhage (SAH), and aortic dissection (AD), finding that undertriage is associated with delays in some key elements of care.1 Orders were delayed compared with correctly triaged patients, suggesting undertriage leads to worse care processes, and presumably, worse outcomes. Fortunately, these delays were fairly small: computed tomography orders for undertriaged patients with SAH or AD were delayed an average 2.4 and 8.9 minutes, respectively. By contrast, electrocardiogram (ECG) and troponin orders for undertriaged patients with STEMI were effectively identical to accurately triaged patients.
Emergency medicine has been described as “looking for needles in a haystack,” or even, “looking for needles in a stack of needles.”2 Not all severe diagnoses are obvious: there is considerable overlap in signs and symptoms in patients with serious and benign diagnoses, and serious and other serious diagnoses. The core challenge is the “limited concordance between presenting complaints and ED discharge diagnoses.”3 Compounding this diagnostic complexity is crowding, which makes finding time-sensitive diagnoses critical to the functioning of an emergency care system often operating over capacity.
Triage plays an essential role in mitigating the harms of the delays experienced in ED waiting rooms.4 Triage is typically conducted by knowledgeable and experienced ED nurses. The triage nurse applies the Emergency Severity Index (ESI) in the US. In the ESI, triage nurses sort patients into 5 levels: level 1 patients require immediate life-saving care; level 2 patients are in distress, have abnormal vital signs, or concerning ECGs; the remaining patients are divided by predicted resource use (ie, laboratory tests, imaging, and/or procedures) into levels 3 to 5.5
The ESI is not intended to and cannot be perfect. Instead, triage relies on the clinical expertise of emergency nurses. In hospitals operating over capacity, the difference between being triaged as an ESI 3 vs 2 can mean many more hours of waiting. It is common for the number of patients assigned an ESI 3 to be higher than an ED can get through in a reasonable time, as many EDs triage more than half of patients as level 3.6 Mistriage may account for some of this, but it more likely represents the distribution of needs of the patients and the community. This resonates with our experience working in busy urban EDs in crowded hospitals, where the waiting room is often full of patients assigned ESI 3 waiting for several more hours than anyone should.
The lack of delays in time to ECG and troponin found by Sax and colleagues1 is reassuring, intuitive, and instructive. Time to ECG in patients with chest pain is a hospital quality metric that is specifically mentioned in the ESI handbook,5 and many EDs have standing orders for laboratory tests that nurses can draw in triage. ECGs and troponins are relatively noninvasive, non–resource-intensive steps that can be fairly easily applied to all ED patients with chest pain.
SAH and AD are trickier; first, as Sax and colleagues1 report, their incidence is far lower than STEMI (patients with STEMI made up more than 75% of their cohort). Furthermore, signs and symptoms of SAH and AD are generally unreliable, and the key diagnostic step is computed tomography (CT) or CT angiography, a resource-intensive tool. Near-perfect sensitivity for SAH could be achieved by scanning every ED patient with a headache within 6 hours of presentation, but the costs and opportunity costs in CT scanner, technician, and radiologist time (and even further gridlocked ED throughput) harm all the other patients assigned ESI 3 in the waiting room, particularly when a small fraction of ED patients with headaches have an SAH. Furthermore, less than half of patients diagnosed with SAH or AD presented with headache or chest pain, respectively, whereas more than 80% of patients with STEMI had chest pain.1 Reassuringly, the delays in key orders for undertriaged patients with AD and SAH reported here were generally short, suggesting that perhaps our current system is appropriately balancing diagnosis and opportunity costs. Mistriage may reflect less straightforward presentations of less common diseases being more difficult to diagnose.
In addition to readily available, inexpensive, noninvasive testing, a key factor differentiating STEMI from AD is that the former has a dedicated acute activation pathway (similar to trauma, stroke, and at some institutions, sepsis). As we prioritize specific disease processes, every other patient presentation is deprioritized. Diseases such as AD and syndromic presentations such as geriatric abdominal pain, which require detailed, clinical judgments, may be the first to experience delays due to this deprioritization. Prioritized pathways with strict timelines (eg, 5-minute ECG reads) can also disrupt clinicians’ thought processes, contributing to heuristic decision-making at the expense of deeper thought, as well as fostering premature closure with diagnostic and therapeutic momentum. Balancing the clear benefits of activation pathways against the diffuse harms needs to be explicit in any proposal to prioritize a particular disease.
The authors1 make an important contribution to the literature on the potential ramifications of mistriage, and leave other avenues to explore in future work. In this study, the authors appropriately focus on undertriage; however, overtriage can also harm through deprioritization of other patients in a resource-constrained setting and through overtesting, where every choice has opportunity costs.
Many EDs have implemented various types of split-flow models to help address wait times, eg, more traditional fast track or minor care areas, or physician-in-triage models. More research is needed on the interaction of these flow models and zones of the ED with triage, and whether these innovations explain the differences seen in triage accuracy and times to key orders, or if they serve primarily to improve time to clinician and other metrics without clear clinical improvements.
Ultimately, perfect triage would require perfect clinical decision-making—at which point the whole ED visit might as well be conducted. Studies of the increasingly prevalent physician-in-triage models, which are equally time-pressured but bring even more advanced decision-making relative to an experienced triage nurse, may shed some light on the value of more accurate triage.
In the future, artificial intelligence (AI) systems built on large language models (LLMs) may help balance how much history, physical examination, and cognitive effort is needed for sufficiently accurate triage. However, the performance of these systems is not yet sufficient to avoid clinical harm, with accuracy below 40% when differentiating between ESI 2 and 4 at the distributions seen in EDs.7 LLMs also risk AI “hallucinations,” as well as magnifying biases in their training data. Ambient learning systems (eg, Abridge, DAX) may improve AI models rapidly, because they are placed in the examination room and incorporate precisely the kind of conversations which occur in triage. These models should be rigorously assessed for potential harm, using the standards any diagnostic test would be evaluated with.
Accurately triaging ED patients is a persistent challenge given the complexities of human physiology, pathologies, and behaviors, particularly in resource-strained EDs. While this Sisyphean challenge could foster nihilism, patients need our help and we can and should work to improve our systems to take better care of people. At the same time, we should take comfort that the delays induced by mistriage for these high-risk conditions seem to be minimal.
