Written by Catherine Burger
Comparing the accuracy of ECG interpretation between Emergency Medicine Providers (EMP) and 3 separate AI models, EMPs narrowly took the lead.
Who does it better… humans or AI?
This study involved 30 practitioners certified by the Turkish Minister of Health, each with 5–10 years of experience vs. 3 separate AI models: GPT-4o, Canva-GPT, and ECG Reader-GPT. GPT-4o is a general-purpose vision-language model (VLM); Canva-GPT is optimized for visual input processing; and ECG Reader-GPT is configured via OpenAI’s custom platform for ECG interpretation. Each group was given 50 multiple-choice questions from a reference textbook, including 12 rhythm disorders, 12 ischemic disorders, 12 conduction disturbances, 5 channelopathies and genetic syndromes, 5 normal ECGs, and 4 ECGs with other pathology. EMPs completed the questions once, while the AI models completed the questions once daily for 30 consecutive days (with question order randomized each time). Median number of correct answers were compared between the groups.
EMPs had the highest overall accuracy, with a median of 41.5 questions correct out of 50 (IQR 37.0-43.0), followed closely by ECG Reader-GPT, with median 39.5 questions correct (IQR 39.0-41.0); then Canva-GPT, with 29 questions correct (IQR 27-30), and finally GPT-4o, with 26 questions correct (IQR 23-27). ECG Reader-GPT narrowly outperformed EMPs in the following sub-categories: rhythm disturbance, ischemic syndromes, other pathologies, and normal ECGs. However, EMPs outperformed ECG Reader-GPT on conduction disturbances and channelopathies. Data on how the AI models performed over time was not published in this article. Authors point out that generalizability of these results is limited, since real-life clinical scenarios are not multiple-choice questions. In addition, details on how ECG Reader-GPT was trained were not available to the authors of this study, so it’s possible that it was trained using the same reference text as the questions in the study.
How does this change my practice?
While I am not ready to rely on a computer for ECG interpretation, this article shows the promise of future AI-assisted ECG interpretation, especially when computers are trained specifically for this task.
Source
Task-specific versus general-purpose AI models in ECG analysis: A comparative study with emergency medicine specialists. Am J Emerg Med. 2025 Sep;95:220-226. doi: 10.1016/j.ajem.2025.06.068. Epub 2025 Jun 27. PMID: 40609173
