Decoding post-resuscitation ECGs: OMI and culprit vessel prediction by experts and AI

Background/Introduction Accurate electrocardiogram (ECG) interpretation after cardiac arrest is critical for recognizing occlusive myocardial infarction (OMI), yet post-resuscitation artifacts complicate diagnosis. Artificial intelligence (AI) tools are emerging as potential aids, but their value in this high-stakes setting remains insufficiently explored. Purpose This study aimed to compare the diagnostic performance of human experts, a domain-specific neural network, and two general-purpose large language model (LLM)–based AI chatbots in detecting OMI on post-return of spontaneous circulation (ROSC) ECGs, and to assess their ability to predict the culprit coronary vessel, using coronary angiography (CAG) as the reference standard. Methods We included patients successfully resuscitated from cardiac arrest with both post-ROSC ECG and CAG available. ECGs were independently evaluated by expert cardiologists (HE), a validated neural network [Queen of Hearts (QoH)], and two LLM-based chatbots (AI-CB: ChatGPT, EKG Analyst). Each method assessed the likelihood of OMI and, where applicable, predicted the suspected culprit vessel (LAD, RCA, RCX; compared to CAG-reference with Fisher’s exact test). Results Among 105 included patients, OMI was angiographically confirmed in 60 cases (57.1%). Both AI-CB achieved perfect sensitivity (100%) but had very low specificity (0.04% and 0.02%), resulting in the lowest AUROC values. HE interpretation reached 83.3% sensitivity, 53.5% specificity, and AUROC 0.723 [95% CI 0.670–0.776]. QoH achieved highest specificity (76.2%) and PPV (78.1%) but lower sensitivity (63.6%) with AUROC 0.709 [0.659–0.759]. F1 scores were highest for HE (0.76), followed by AI-CB (0.73 each) and QoH (0.70). Inter-rater agreement between HE was moderate (Cohen’s kappa 0.431). Regarding culprit vessel prediction, HE showed moderate accuracy for LAD (51.6%; p=0.001), RCA (37.5%; p=0.007) and RCX (23.5%; p=0.259). QoH did not provide vessel classification. Both AI-CB almost exclusively predicted LAD (≈98–99 cases), resulting in low overall accuracy (≈28%; p=0.185 and 0.682) and failed to correctly identify RCA or RCX (both p=NS). Conclusion Post-ROSC ECG interpretation remains a complex task. Human experts provide the most balanced and clinically reliable performance in OMI detection and vessel prediction. Domain-specific AI tools such as QoH may offer supportive triage value, whereas general-purpose LLMs currently lack diagnostic accuracy and should not be relied upon for clinical decision-making.