In a retrospective study of diagnostically challenging ECGs, an AI application demonstrated better diagnostic performance of occlusion myocardial infarction as compared to interventional cardiologists, achieving 100% sensitivity and negative predictive value versus 62-85% and 69-83% respectively.
The AI application (PMCardio) achieved 100% sensitivity and 100% negative predictive value in detecting occlusion myocardial infarction from challenging ECGs.
The three interventional cardiologists showed substantially lower sensitivity and negative predictive values compared to the AI when independently assessing the same challenging ECGs.
The AI application had a lower overall misclassification rate than all three cardiologists.
Of the 33 patients with diagnostically challenging ECGs, 13 (57% of acute coronary syndrome cases) had a confirmed occlusion myocardial infarction.
The study specifically targeted ECGs that were diagnostically challenging and generated debates about diagnosis among clinicians at the time of patient admission.
This research suggests that an AI-based smartphone application (PMCardio) outperformed experienced heart specialists in detecting dangerous coronary artery blockages from difficult-to-read electrocardiograms (ECGs). The study looked at 33 ECGs that were genuinely puzzling to clinicians at the time patients arrived in the emergency department — cases where even experts disagreed about the diagnosis. When three interventional cardiologists (specialists who perform procedures to open blocked arteries) independently reviewed these ECGs without knowing the final diagnosis, they correctly identified blockages between 62% and 85% of the time. The AI, by contrast, correctly identified every confirmed blockage — achieving 100% sensitivity — meaning it missed none of the true cases. The practical implication highlighted by this research is about missed diagnoses. When a patient has a blocked coronary artery (called an occlusion myocardial infarction or OMI), rapid treatment is critical. Missing the diagnosis on an ECG can delay life-saving treatment. The AI's 100% negative predictive value means that in this small study, a negative AI result reliably indicated no occlusion was present, which could potentially help clinicians rule out dangerous blockages more confidently. The AI's main limitation was lower specificity (60%), meaning it sometimes flagged cases as blockages when they were not — a pattern that could lead to unnecessary procedures. This research suggests that AI tools could serve as a safety net or second opinion for clinicians reviewing difficult ECGs, potentially catching dangerous heart blockages that might otherwise be missed. However, the study was small (33 ECGs), retrospective, and focused specifically on pre-selected challenging cases, so the findings should be interpreted with caution and would need confirmation in larger, prospective studies before drawing broad conclusions about clinical use.
Zimmerli A, Dupre M, Liabot Q, Sellah C, Meier D, Salihu A, et al.. (2026). Human vs Artificial Intelligence in Detecting Coronary Occlusions in ST-Elevation Myocardial Infarction: Retrospective Study.. JMIR cardio. https://doi.org/10.2196/88399