Faculty Bibliography

OBJECTIVES/UNASSIGNED:This study aimed to assess the current utilization of artificial intelligence (AI) tools among emergency physicians, their attitudes toward AI in clinical practice, and how a national physician professional organization could best support its members regarding AI. METHODS/UNASSIGNED:A cross-sectional survey was emailed to American College of Emergency Physicians members and made available to conference attendees at a national symposium. The survey collected demographic information, details on the use of noninstitutional and institutional AI tools, attitudes toward AI, and desired forms of support. Descriptive statistics were used to summarize the data. RESULTS/UNASSIGNED:A total of 658 physicians responded, primarily practicing attendings (78%) and residents (9%), with 60% aged 35 to 54 years and 67% identifying as male; these respondents represented 2% of the membership of American College of Emergency Physicians. Noninstitutional AI tool use (eg, ChatGPT and independent electrocardiogram interpretation) was reported by 31% of respondents. Institutional AI was integrated into 52% of respondents' practices, with 18% regularly using ambient AI documentation and 22% using AI-assisted clinical decision support. AI tools for point-of-care ultrasound were available to 10%, and AI-assisted radiology interpretation was used by 14%, mainly for X-rays and computed tomography. Operational AI for triage, capacity management, and staff optimization were reported by 15%, while 9% used AI-assisted coding and billing. Moreover, 75% believed AI improves clinical efficiency, 57% felt it enhanced care quality, but 12% expressed concern about job displacement, 16% are unsure whether AI tools would adequately comply with Health Insurance Portability and Accountability Act regulations, and 38% noted potential biases. About half desire educational support and guidelines. CONCLUSION/UNASSIGNED:In this nonrepresentative emergency physician survey, respondents reported moderate rates of adoption of AI tools and generally positive attitudes toward AI's impact on efficiency and care quality. However, respondents also reported important concerns about job displacement, Health Insurance Portability and Accountability Act regulation compliance, and potential biases. Larger studies are needed to more fully understand emergency physician views on AI.

Artificial intelligence (AI) tools and technologies are increasingly being integrated into emergency medicine (EM) practice, not only offering potential benefits such as improved efficiency, better patient experience, and increased safety, but also resulting in potential risks including exacerbation of biases. These biases, inadvertently embedded in AI algorithms or training data, can adversely affect clinical decision making for diverse patient populations. Bias is a universal human attribute, subject to introduction into any human interaction. The risk with AI is magnification of, or even normalization of, patterns of biases across the health care ecosystem within tools that in time may be considered authoritative. This article, the work of members of the American College of Emergency Physicians (ACEP) AI Task Force, aims to equip emergency physicians (EPs) with a practical framework for understanding, identifying, and addressing bias in clinical and operational AI tools encountered in the emergency department (ED). For this publication, we have defined bias as a systematic flaw in a decision-making process that results in unfair or unintended outcomes that can be inadvertently embedded in AI algorithms or training data. This can result in adverse effects on clinical decision making for diverse patient populations. We begin by reviewing common sources of AI bias relevant to EM, including data, algorithmic, measurement, and human-interaction factors, and then, we discuss the potential pitfalls. Following this, we use illustrative examples from EM practice (eg, triage tools, risk stratification, and medical devices) to demonstrate how bias can manifest. We subsequently discuss the evolving regulatory landscape, structured assessment frameworks (including predeployment, continuous monitoring, and postdeployment steps), key principles (like sociotechnical perspectives and stakeholder engagement), and specific tools. Finally, this review outlines the EP's vital role in mitigation of AI-related biases through advocacy, local validation, clinical feedback, demanding transparency, and maintaining clinical judgment over automation.

BACKGROUND:Endotracheal intubation in the emergency department (ED) is a critical and time-sensitive procedure requiring both technical skills and cognitive-based reasoning. Evidence on supervised resident-attending dyads with differing years of seniority on decision making during clinical encounters with endotracheal intubations is nascent. OBJECTIVE:To investigate the intersection of postgraduate years in clinical practice between resident and attending supervisor dyads and its associations for clinician choice of laryngoscopy technique and paralytic agent during ED intubations. METHODS:We conducted a retrospective analysis of intubations performed at a multi-site, urban, academic emergency medicine training program, analyzing institutional airway registry data from 2013 to 2023. Using a standardized predictor that accounted for similarity in years of clinical experience within a dyad between a resident and their supervising attending, we performed adjusted mixed-effects logistic regression examining the association of this dyad on two primary outcomes in endotracheal intubation decision making. Our primary outcome measures were the selection of a laryngoscopy technique (either DL or VL), and of a paralytic agent (either short-acting or long-acting) analyzed as categorical variables with a linear mixed effects model, using a binomial response distribution. RESULTS:We examined 2969 intubations for choice of laryngoscopy technique (n = 1117, 37.6 %) and paralytic agent (n = 967, 32.6 %). Higher adjusted odds (aOR) were associated with resident choice of DL over VL when years of experience between residents and supervising attendings were more concordant (aOR 3.05, 95 % CI: 1.1-8.2). Choice of paralytic agent was not associated with differing years of experience. CONCLUSION/CONCLUSIONS:Concordant years of experience between residents and their attendings were associated with technical skill-based laryngoscopy technique choice but not for cognitive-based reasoning in paralytic agent choice among ED intubations, suggesting supervising attending's years in clinical practice may influence decision making during time-sensitive procedures.

Background Emergency Department (ED) handoff to inpatient teams is a potential source of error. Generative Artificial Intelligence (AI) has shown promise in succinctly summarizing large quantities of clinical data and may help improve Emergency Department (ED) handoff. Objectives Our objectives were to: 1) evaluate the accuracy, clinical utility, and safety of AI-generated ED-to-inpatient handoff summaries; 2) identify patient and visit characteristics influencing summary effectiveness; and 3) characterize potential error patterns to inform implementation strategies. Methods This exploratory study evaluated AI-generated handoff summaries at an urban academic ED (February-April 2024). A HIPAA-compliant GPT-4 model generated summaries aligned with the IPASS framework; ED providers assessed summary accuracy, usefulness, and safety through on-shift surveys. Results Among 50 cases, median quality and usefulness scores were 4/5 (SE = 0.13). Safety concerns arose in 6% of cases, with issues including data omissions and mischaracterizations. Consultation status significantly affected usefulness scores (p < 0.05). Omissions of relevant medications, laboratory results, and other essential detailss were noted (n=6), and EM clinicians disagreed with some AI characterizations of patient stability, vitals and workup (n=8). The most common response was positive impressions of the technology incorporated into the handoff process (n=11). Conclusions This exploratory provider-in-the-loop model demonstrated clinical acceptability and highlighted areas for refinement. Future studies should incorporate recipient perspectives and examine clinical outcomes to scale and optimize AI implementation.

BACKGROUND:High-sensitivity cardiac troponin (hs-cTnI) assays can quantify troponin concentrations with low limits of detection, potentially expediting and enhancing myocardial infarction diagnoses. This study investigates the real-world impact of hs-cTnI implementation on operational metrics and downstream cardiac services in patients presenting to the emergency department with chest pain. METHODS AND RESULTS/RESULTS:[lt] 0.001) during the index encounter. CONCLUSION/CONCLUSIONS:Implementation of the hs-cTnI assay was associated with reduced hospital admissions, shorter length of stay, and decreases in most downstream cardiac testing.

INTRODUCTION/UNASSIGNED:With the move toward competency-based medical education (CBME), data from the electronic health record (EHR) for informed self-improvement may be valuable as a part of programmatic assessment. Personalized dashboards are one way to view these clinical data. The purpose of this concept paper is to summarize the current state of clinical dashboards as they can be utilized by emergency medicine (EM) residency programs. METHODS/UNASSIGNED:The author group consisted of EM physicians from multiple institutions with medical education and informatics backgrounds and was identified by querying faculty presenting on resident clinical dashboards at the 2024 Society for Academic Emergency Medicine conference. Additional authors were identified by members of the initial group. Best practice literature was referenced; if none was available, group consensus was used. CATEGORIES OF METRICS/UNASSIGNED:Clinical exposures as well as efficiency, quality, documentation, and diversity metrics may be included in a resident dashboard. Resident dashboard metrics should focus on resident-sensitive measures rather than those primarily affected by attendings or systems-based factors. CONSIDERATIONS FOR IMPLEMENTATION/UNASSIGNED:Implementation of these dashboards requires the technical expertise to turn EHR data into actionable data, a process called EHR phenotyping. The dashboard can be housed directly in the EHR or on a separate platform. Dashboard developers should consider how their implementation plan will affect how often dashboard data will be refreshed and how to best display the data for ease of understanding. IMPLICATIONS FOR EDUCATION & TRAINING/UNASSIGNED:Dashboards can provide objective data to residents, residency leadership and clinical competency committees as they identify areas of strength, growth areas, and set specific and actionable goals. The success of resident dashboards is reliant on resident buy-in and creating a culture of psychological safety through thoughtful implementation, coaching, and regular feedback. . CONCLUSION/UNASSIGNED:Personalized clinical dashboards can play a crucial role in programmatic assessment within CBME, helping EM residents focus their efforts as they advance and refine their skills during training.

IMPORTANCE/UNASSIGNED:The emergency department (ED) offers an opportunity to initiate palliative care for older adults with serious, life-limiting illness. OBJECTIVE/UNASSIGNED:To assess the effect of a multicomponent intervention to initiate palliative care in the ED on hospital admission, subsequent health care use, and survival in older adults with serious, life-limiting illness. DESIGN, SETTING, AND PARTICIPANTS/UNASSIGNED:Cluster randomized, stepped-wedge, clinical trial including patients aged 66 years or older who visited 1 of 29 EDs across the US between May 1, 2018, and December 31, 2022, had 12 months of prior Medicare enrollment, and a Gagne comorbidity score greater than 6, representing a risk of short-term mortality greater than 30%. Nursing home patients were excluded. INTERVENTION/UNASSIGNED:A multicomponent intervention (the Primary Palliative Care for Emergency Medicine intervention) included (1) evidence-based multidisciplinary education; (2) simulation-based workshops on serious illness communication; (3) clinical decision support; and (4) audit and feedback for ED clinical staff. MAIN OUTCOME AND MEASURES/UNASSIGNED:The primary outcome was hospital admission. The secondary outcomes included subsequent health care use and survival at 6 months. RESULTS/UNASSIGNED:There were 98 922 initial ED visits during the study period (median age, 77 years [IQR, 71-84 years]; 50% were female; 13% were Black and 78% were White; and the median Gagne comorbidity score was 8 [IQR, 7-10]). The rate of hospital admission was 64.4% during the preintervention period vs 61.3% during the postintervention period (absolute difference, -3.1% [95% CI, -3.7% to -2.5%]; adjusted odds ratio [OR], 1.03 [95% CI, 0.93 to 1.14]). There was no difference in the secondary outcomes before vs after the intervention. The rate of admission to an intensive care unit was 7.8% during the preintervention period vs 6.7% during the postintervention period (adjusted OR, 0.98 [95% CI, 0.83 to 1.15]). The rate of at least 1 revisit to the ED was 34.2% during the preintervention period vs 32.2% during the postintervention period (adjusted OR, 1.00 [95% CI, 0.91 to 1.09]). The rate of hospice use was 17.7% during the preintervention period vs 17.2% during the postintervention period (adjusted OR, 1.04 [95% CI, 0.93 to 1.16]). The rate of home health use was 42.0% during the preintervention period vs 38.1% during the postintervention period (adjusted OR, 1.01 [95% CI, 0.92 to 1.10]). The rate of at least 1 hospital readmission was 41.0% during the preintervention period vs 36.6% during the postintervention period (adjusted OR, 1.01 [95% CI, 0.92 to 1.10]). The rate of death was 28.1% during the preintervention period vs 28.7% during the postintervention period (adjusted OR, 1.07 [95% CI, 0.98 to 1.18]). CONCLUSIONS AND RELEVANCE/UNASSIGNED:This multicomponent intervention to initiate palliative care in the ED did not have an effect on hospital admission, subsequent health care use, or short-term mortality in older adults with serious, life-limiting illness. TRIAL REGISTRATION/UNASSIGNED:ClinicalTrials.gov Identifier: NCT03424109.

Clinical documentation in the United States has grown longer and more difficult to read, a phenomenon described as "note bloat." This issue is especially pronounced in emergency medicine, where high diagnostic uncertainty and brief evaluations demand focused, efficient chart review to inform decision-making. Note bloat arises from multiple factors: efforts to enhance billing, mitigate malpractice risk, and leverage electronic health record tools that improve speed and completeness. We discuss best practices based on available evidence and expert opinion to improve note clarity and concision. Recent E/M coding reforms aim to streamline documentation by prioritizing medical decision-making over details of historical and physical examination, though implementation varies. New technologies such as generative artificial intelligence present opportunities and challenges for documentation practices. Addressing note bloat will require ongoing effort from clinical leadership, electronic health record vendors, and professional organizations.

Ensuring reliability of Large Language Models (LLMs) in clinical tasks is crucial. Our study assesses two state-of-the-art LLMs (ChatGPT and LlaMA-2) for extracting clinical information, focusing on cognitive tests like MMSE and CDR. Our data consisted of 135,307 clinical notes (Jan 12th, 2010 to May 24th, 2023) mentioning MMSE, CDR, or MoCA. After applying inclusion criteria 34,465 notes remained, of which 765 underwent ChatGPT (GPT-4) and LlaMA-2, and 22 experts reviewed the responses. ChatGPT successfully extracted MMSE and CDR instances with dates from 742 notes. We used 20 notes for fine-tuning and training the reviewers. The remaining 722 were assigned to reviewers, with 309 each assigned to two reviewers simultaneously. Inter-rater-agreement (Fleiss' Kappa), precision, recall, true/false negative rates, and accuracy were calculated. Our study follows TRIPOD reporting guidelines for model validation. For MMSE information extraction, ChatGPT (vs. LlaMA-2) achieved accuracy of 83% (vs. 66.4%), sensitivity of 89.7% (vs. 69.9%), true-negative rates of 96% (vs 60.0%), and precision of 82.7% (vs 62.2%). For CDR the results were lower overall, with accuracy of 87.1% (vs. 74.5%), sensitivity of 84.3% (vs. 39.7%), true-negative rates of 99.8% (98.4%), and precision of 48.3% (vs. 16.1%). We qualitatively evaluated the MMSE errors of ChatGPT and LlaMA-2 on double-reviewed notes. LlaMA-2 errors included 27 cases of total hallucination, 19 cases of reporting other scores instead of MMSE, 25 missed scores, and 23 cases of reporting only the wrong date. In comparison, ChatGPT's errors included only 3 cases of total hallucination, 17 cases of wrong test reported instead of MMSE, and 19 cases of reporting a wrong date. In this diagnostic/prognostic study of ChatGPT and LlaMA-2 for extracting cognitive exam dates and scores from clinical notes, ChatGPT exhibited high accuracy, with better performance compared to LlaMA-2. The use of LLMs could benefit dementia research and clinical care, by identifying eligible patients for treatments initialization or clinical trial enrollments. Rigorous evaluation of LLMs is crucial to understanding their capabilities and limitations.