Development of emergency department (ED) triage systems that accurately differentiate and prioritize critically ill from stable patients remains challenging. We used machine learning models to predict clinical outcomes, and then compared their performance with that of a conventional approach—the Emergency Severity Index (ESI). Using National Hospital and Ambulatory Medical Care Survey (NHAMCS) ED data, from 2007 through 2015, we identified all adult patients (aged ≥ 18 years). In the randomly sampled training set (70%), using routinely available triage data as predictors (e.g., demographics, triage vital signs, chief complaints, comorbidities), we developed four machine learning models: Lasso regression, random forest, gradient boosted decision tree, and deep neural network. As the reference model, we constructed a logistic regression model using the five-level ESI data. The clinical outcomes were critical care (admission to intensive care unit or in-hospital death) and hospitalization (direct hospital admission or transfer). In the test set (the remaining 30%), we measured the predictive performance, including area under the receiver-operating-characteristics curve (AUC) and net benefit (decision curves) for each model. Of 135,470 eligible ED visits, 2.1% had critical care outcome and 16.2% had hospitalization outcome. In the critical care outcome prediction, all four machine learning models outperformed the reference model (e.g., AUC, 0.86 [95%CI 0.85–0.87] in the deep neural network vs 0.74 [95%CI 0.72–0.75] in the reference model), with less under-triaged patients in ESI triage levels 3 to 5 (urgent to non-urgent). Likewise, in the hospitalization outcome prediction, all machine learning models outperformed the reference model (e.g., AUC, 0.82 [95%CI 0.82–0.83] in the deep neural network vs 0.69 [95%CI 0.68–0.69] in the reference model) with less over-triages in ESI triage levels 1 to 3 (immediate to urgent). In the decision curve analysis, all machine learning models consistently achieved a greater net benefit—a larger number of appropriate triages considering a trade-off with over-triages—across the range of clinical thresholds. Compared to the conventional approach, the machine learning models demonstrated a superior performance to predict critical care and hospitalization outcomes. The application of modern machine learning models may enhance clinicians’ triage decision making, thereby achieving better clinical care and optimal resource utilization.

Of 137 million annual emergency department (ED) visits in the United States, 30 million visits are made by children.^1-3 With the steady increase in the volume and acuity of patient visits to EDs,⁴ accurate differentiation and prioritization of patients at the ED triage is important. However, current triage systems have suboptimal ability to differentiate critically ill children,^5-7 and the proportion of children seen by a physician within the time recommended by triage has been declining because of pervasive ED crowding.⁸ Therefore, it is essential to optimize triage systems to not only avoid undertriaging critically ill children but also reduce overtriaging in order to provide high-quality and timely care and to achieve efficient resource allocation in the ED.

Background: Despite the increasing availability of clinical decision support systems (CDSSs) and rising expectation for CDSSs based on artificial intelligence (AI), little is known about the acceptance of AI-based CDSS by physicians and its barriers and facilitators in emergency care settings. Objective: We aimed to evaluate the acceptance, barriers, and facilitators to implementing AI-based CDSSs in the emergency care setting through the opinions of physicians on our newly developed, real-time AI-based CDSS, which alerts ED physicians by predicting aortic dissection based on numeric and text information from medical charts, by using the Unified Theory of Acceptance and Use of Technology (UTAUT; for quantitative evaluation) and the Consolidated Framework for Implementation Research (CFIR; for qualitative evaluation) frameworks. Methods: This mixed methods study was performed from March to April 2021. Transitional year residents (n=6), emergency medicine residents (n=5), and emergency physicians (n=3) from two community, tertiary care hospitals in Japan were included. We first developed a real-time CDSS for predicting aortic dissection based on numeric and text information from medical charts (eg, chief complaints, medical history, vital signs) with natural language processing. This system was deployed on the internet, and the participants used the system with clinical vignettes of model cases. Participants were then involved in a mixed methods evaluation consisting of a UTAUT-based questionnaire with a 5-point Likert scale (quantitative) and a CFIR-based semistructured interview (qualitative). Cronbach α was calculated as a reliability estimate for UTAUT subconstructs. Interviews were sampled, transcribed, and analyzed using the MaxQDA software. The framework analysis approach was used during the study to determine the relevance of the CFIR constructs. Results: All 14 participants completed the questionnaires and interviews. Quantitative analysis revealed generally positive responses for user acceptance with all scores above the neutral score of 3.0. In addition, the mixed methods analysis identified two significant barriers (System Performance, Compatibility) and two major facilitators (Evidence Strength, Design Quality) for implementation of AI-based CDSSs in emergency care settings. Conclusions: Our mixed methods evaluation based on theoretically grounded frameworks revealed the acceptance, barriers, and facilitators of implementation of AI-based CDSS. Although the concern of system failure and overtrusting of the system could be barriers to implementation, the locality of the system and designing an intuitive user interface could likely facilitate the use of optimal AI-based CDSS. Alleviating and resolving these factors should be key to achieving good user acceptance of AI-based CDSS.

Background: Whether GPT-4, the conversational artificial intelligence, can accurately diagnose and triage health conditions and whether it presents racial and ethnic biases in its decisions remain unclear. Objective: We aim to assess the accuracy of GPT-4 in the diagnosis and triage of health conditions and whether its performance varies by patient race and ethnicity. Methods: We compared the performance of GPT-4 and physicians, using 45 typical clinical vignettes, each with a correct diagnosis and triage level, in February and March 2023. For each of the 45 clinical vignettes, GPT-4 and 3 board-certified physicians provided the most likely primary diagnosis and triage level (emergency, nonemergency, or self-care). Independent reviewers evaluated the diagnoses as “correct” or “incorrect.” Physician diagnosis was defined as the consensus of the 3 physicians. We evaluated whether the performance of GPT-4 varies by patient race and ethnicity, by adding the information on patient race and ethnicity to the clinical vignettes. Results: The accuracy of diagnosis was comparable between GPT-4 and physicians (the percentage of correct diagnosis was 97.8% (44/45; 95% CI 88.2%-99.9%) for GPT-4 and 91.1% (41/45; 95% CI 78.8%-97.5%) for physicians; P=.38). GPT-4 provided appropriate reasoning for 97.8% (44/45) of the vignettes. The appropriateness of triage was comparable between GPT-4 and physicians (GPT-4: 30/45, 66.7%; 95% CI 51.0%-80.0%; physicians: 30/45, 66.7%; 95% CI 51.0%-80.0%; P=.99). The performance of GPT-4 in diagnosing health conditions did not vary among different races and ethnicities (Black, White, Asian, and Hispanic), with an accuracy of 100% (95% CI 78.2%-100%). P values, compared to the GPT-4 output without incorporating race and ethnicity information, were all .99. The accuracy of triage was not significantly different even if patients’ race and ethnicity information was added. The accuracy of triage was 62.2% (95% CI 46.5%-76.2%; P=.50) for Black patients; 66.7% (95% CI 51.0%-80.0%; P=.99) for White patients; 66.7% (95% CI 51.0%-80.0%; P=.99) for Asian patients, and 62.2% (95% CI 46.5%-76.2%; P=.69) for Hispanic patients. P values were calculated by comparing the outputs with and without conditioning on race and ethnicity. Conclusions: GPT-4’s ability to diagnose and triage typical clinical vignettes was comparable to that of board-certified physicians. The performance of GPT-4 did not vary by patient race and ethnicity. These findings should be informative for health systems looking to introduce conversational artificial intelligence to improve the efficiency of patient diagnosis and triage.