In the face of confounders: Atrial fibrillation detection - Practitioners vs. ChatGPT.

Introduction: Atrial fibrillation (AF) is the most common arrhythmia in clinical practice, yet interpretation concerns among healthcare providers persist. Confounding factors contribute to false-positive and false-negative AF diagnoses, leading to potential omissions. Artificial intelligence advancements show promise in electrocardiogram (ECG) interpretation. We sought to examine the diagnostic accuracy of ChatGPT-4omni (GPT-4o), equipped with image evaluation capabilities, in interpreting ECGs with confounding factors and compare its performance to that of physicians.

Methods: Twenty ECG cases, divided into Group A (10 cases of AF or atrial flutter) and Group B (10 cases of sinus or another atrial rhythm), were crafted into multiple-choice questions. Total of 100 practitioners (25 from each: emergency medicine, internal medicine, primary care, and cardiology) were tasked to identify the underlying rhythm. Next, GPT-4o was prompted in five separate sessions.

Results: GPT-4o performed inadequately, averaging 3 (±2) in Group A questions and 5.40 (±1.34) in Group B questions. Upon examining the accuracy of the total ECG questions, no significant difference was found between GPT-4o, internists, and primary care physicians (p = 0.952 and = 0.852, respectively). Cardiologists outperformed other medical disciplines and GPT-4o (p < 0.001), while emergency physicians followed in accuracy, though comparison to GPT-4o only indicated a trend (p = 0.068).

Conclusion: GPT-4o demonstrated suboptimal accuracy with significant under- and over-recognition of AF in ECGs with confounding factors. Despite its potential as a supportive tool for ECG interpretation, its performance did not surpass that of medical practitioners, underscoring the continued importance of human expertise in complex diagnostics.