Washington University Emergency Medicine Journal Club – May 21, 2026
Dr. Brian Cohn
Hello all,
This month we will look at the use of AI in Emergency Medicine, specifically focusing on ECG and CXR interpretation.
The PGY-1, PGY-2, and PGY-4 articles will be appraised using the Diagnostic Test form.
The PGY-3 paper will be appraised using the Therapy form.
Vignette
You’re working an evening shift in a busy community emergency department when EMS brings in a 62-year-old male with chest pain. He reports 2 hours of substernal pressure radiating to his left arm, associated with diaphoresis and nausea. His vital signs are notable for a blood pressure of 148/92 and a heart rate of 104.
An ECG is obtained within minutes of arrival and is interpreted by the computer as “normal sinus rhythm, no acute ST-segment elevation.” You review the tracing yourself—there are no obvious STEMI criteria, but something about the morphology gives you pause. Your department recently implemented an AI-based ECG interpretation tool that flags subtle patterns concerning for occlusive myocardial infarction (OMI), even in the absence of classic STEMI findings. The AI algorithm flags this ECG as “high probability for acute coronary occlusion.”
While you’re considering whether to activate the cath lab, a chest X-ray is performed to evaluate for alternative diagnoses. Your preliminary interpretation is that the x-ray is unremarkable, but the AI-assisted imaging software highlights a possible subtle right lower lobe opacity. You are now faced with two decisions: should the AI-flagged ECG prompt emergent cardiology activation despite a “non-diagnostic” ECG, and how much weight should you give to AI-assisted chest X-ray findings that differ from your initial interpretation?
As you consider your next steps, you begin to wonder if incorporating AI into ECG and imaging interpretation actually improves diagnostic accuracy and patient outcomes, or does it risk overdiagnosis and unnecessary interventions?
PICO Question #1
Population: Adult emergency department patients presenting with symptoms
concerning for acute coronary syndrome
Intervention: Artificial intelligence–assisted ECG interpretation
Comparison: Standard ECG interpretation using STEMI criteria and/or clinician
interpretation
Outcome: Accuracy for detection of occlusive myocardial infarction (OMI), including
sensitivity, specificity, and time to diagnosis/reperfusion
PICO Question #2
Population: Adult emergency department patients undergoing chest radiography
for acute cardiopulmonary symptoms
Intervention: Artificial intelligence–assisted chest X-ray interpretation
Comparison: Standard interpretation by emergency physicians and/or radiologists
Outcome: Diagnostic accuracy, clinical decision-making, and downstream outcomes
(e.g., appropriate treatment, disposition, or time to diagnosis.
Article 1: Carvalho PEP, Belzer W, Pollmann DL, et al. AI-Enhanced
Electrocardiogram for Detection of Occlusive Myocardial Infarction in High-Risk
Non-ST-Segment Elevation Acute Coronary Syndrome. JACC Adv. 2026
Apr;5(4):102663. [Answer Key].
Article 2: Herman R, Meyers HP, Smith SW, et al. International evaluation of an
artificial intelligence–powered electrocardiogram model detecting acute coronary
occlusion myocardial infarction. Eur Heart J Digit Health. 2024;5(2):123-133.
[Answer Key].
Article 3: Hwang EJ, Goo JM, Nam JG, et al. Conventional Versus Artificial
Intelligence-Assisted Interpretation of Chest Radiographs in Patients With Acute
Respiratory Symptoms in Emergency Department: A Pragmatic Randomized Clinical
Trial. Korean J Radiol. 2023 Mar;24(3):259-270. [Answer Key].
Article 4: Ló pez Alcolea J, Ferná ndez Alfonso A, Cano Alonso R, et al. Diagnostic
Performance of Artificial Intelligence in Chest Radiographs Referred from the
Emergency Department. Diagnostics (Basel). 2024 Nov 18;14(22):2592. [Answer
Key].
Bottom Line
Given the rising use of artificial intelligence (AI) in general, and its particular efficacy in pattern recognition, there are areas in medicine where AI could be useful tool. Electrocardiographic and radiologic interpretation involve pattern recognition on the part of the clinician, and are two areas in which AI could prove beneficial to the emergency physician. We therefore reviewed four articles in which AI accuracy was evaluated in ECG interpretation and plain chest radiograph interpretation.
The most promising results came from an international AI-ECG study involving more than 2,200 patients, where the AI model achieved an area under the curve (AUC) of 0.94 for detecting occlusive myocardial infarction (OMI), compared with substantially lower sensitivity for traditional STEMI criteria (80.6% vs. 32.5%). The AI system identified many OMI cases missed by STEMI criteria and achieved diagnostic performance comparable to expert ECG interpreters, suggesting a potential role as a clinical decision-support tool.
In contrast, a separate single-center study of high-risk non-STEMI patients found much weaker performance. AI-enhanced ECG interpretation demonstrated only 58% sensitivity and 78% specificity for OMI on the initial ECG, with a positive likelihood ratio of 2.6 and negative likelihood ratio of 0.54, values unlikely to meaningfully change disease probability in clinical practice. While the AI-assisted pathway reduced false-positive catheterization activations from 42% to 22%, it still missed a substantial number of patients with OMI.
Results were similarly mixed for chest radiograph interpretation. In a randomized trial of 3,576 emergency department patients in South Korea, AI-assisted interpretation did not improve trainee radiologists’ sensitivity for acute thoracic disease (67.2% with AI vs. 66.0% without AI) or reduce false-positive rates (19.3% vs. 18.5%). AI assistance also failed to improve important clinical outcomes such as CT utilization, antibiotic use, emergency department length of stay, or 30-day revisit rates. As a stand-alone tool, however, the AI system demonstrated very high sensitivity (95.3%) but an extremely high false-positive rate (61.7%).
A final study comparing AI with radiology residents at a single center in Madrid, Spain found that AI performance varied substantially by pathology. The system achieved 100% sensitivity for pneumothorax and fractures, 75.6% sensitivity for pulmonary opacities, 59.7% for pleural effusions, and only 33.3% for pulmonary nodules. Overall, radiology residents matched or exceeded AI performance for most findings.
Across these four studies, AI demonstrated mixed performance when applied to ECG and chest radiograph interpretation in emergency medicine. Taken together, these studies suggest that AI can approach expert-level performance in selected ECG applications and may serve as a useful adjunct to clinician interpretation, but current evidence does not support replacing physician expertise or substantially changing emergency department practice based on AI alone.