ECG-based detection of occlusion myocardial infarction: a dedicated deep neural network versus multimodal large language models and physicians — a retrospective diagnostic accuracy study

Overview

Emergency ECG interpretation for acute coronary occlusion is shifting from the STEMI paradigm toward the broader occlusion myocardial infarction (OMI) concept. This retrospective diagnostic accuracy study compared the OMI-trained PMcardio Queen of Hearts (QoH) AI-ECG model against two general-purpose multimodal LLMs (ChatGPT 5.2, Gemini 3 Pro) and ten emergency physicians (five specialists, five residents) on 36 ECGs from patients referred for emergent coronary angiography. QoH achieved the highest discrimination (AUC 0.96) and sensitivity (95.8%), significantly outperforming physicians and both LLMs, while the LLMs showed only fair-to-moderate answer consistency across repeated queries and, for Gemini 3 Pro, marked overconfidence on incorrect diagnoses.

Published in: BMC Emergency Medicine
Published on: 07 August 2026

Background

Accurate ECG interpretation for acute coronary occlusion is a critical, time-sensitive task in emergency care, and the field is shifting from the STEMI paradigm to the broader occlusion myocardial infarction (OMI) concept. This retrospective diagnostic accuracy study compared a dedicated occlusion-detection deep neural network (PMcardio Queen of Hearts, QoH) against two general-purpose multimodal large language models (ChatGPT 5.2, Gemini 3 Pro) and emergency physicians for ECG-based OMI detection.

Methods

The authors evaluated 36 twelve-lead ECGs (24 angiographically confirmed OMI, 12 non-OMI) drawn from patients referred for emergent coronary angiography (McCabe et al. dataset). Each ECG was interpreted by QoH (queried once, deterministic), by ChatGPT 5.2 and Gemini 3 Pro (each queried 5 times per ECG in separate sessions to test response consistency), and by 5 emergency medicine specialists and 5 emergency medicine residents.

Results

QoH showed the strongest discrimination (AUC 0.96, 95% CI 0.91–1.00), with 86.1% accuracy, 95.8% sensitivity, and 66.7% specificity. ChatGPT 5.2 reached an AUC of 0.81 (66.7% accuracy, 54.2% sensitivity, 91.7% specificity), while Gemini 3 Pro had the weakest performance (AUC 0.62, 55.6% accuracy, 58.3% sensitivity, 50.0% specificity). EM specialists achieved 75.0% accuracy, 66.7% sensitivity, and 91.7% specificity — the best result among physicians — while EM residents reached 58.3% accuracy. QoH's sensitivity was significantly higher than every other interpreter, including EM specialists (p=0.016), ChatGPT 5.2 (p=0.006), Gemini 3 Pro (p=0.004), and residents (p=0.008); the AUC gap versus ChatGPT (Δ0.15) and the accuracy gap versus EM specialists were numerically large but did not reach significance in this underpowered 36-ECG sample. Across the 5 repeated runs, LLM responses were only fair-to-moderately consistent (Fleiss' κ 0.24–0.49). Gemini 3 Pro was poorly calibrated, reporting a median confidence of 95% even on incorrect diagnoses (Brier score 0.40, worse than the random-guess baseline of 0.25), whereas ChatGPT 5.2 was better calibrated (Brier 0.230) with a small but real confidence drop on wrong answers.

Conclusion

The dedicated occlusion-detection model showed the strongest and most clinically favorable profile, particularly for sensitivity — the most dangerous error in this setting is a missed coronary occlusion. General-purpose multimodal LLMs are not currently safe for autonomous OMI diagnosis and, at most, could serve as an adjunct or second-reader tool pending prospective validation.