Abstract
<title>Abstract</title> <p>Occluded myocardial infarction (OMI) remains under-recognized in patients without ST-elevation, delaying revascularization despite a high risk of adverse outcomes. We fine-tuned and validated a foundation model for detecting OMI from standard 12-lead electrocardiograms (ECGs), recorded between 2016 and 2022. The study included 17 165 ECGs from 11 338 patients. Among these, 4 038 encounters proceeded to acute invasive coronary angiography, identified from the Norwegian Registry of Invasive Cardiology (NORIC), and 1 338 of these were diagnosed with OMI. OMI was defined as an acute culprit lesion with Thrombolysis In Myocardial Infarction (TIMI) flow grade 0–2 treated with percutaneous coronary intervention. A pretrained ECG foundation model was fine-tuned and evaluated on a held-out test set. The model achieved an area under the receiver operating characteristic (AUROC) curve of 0.93 and an area under the precision-recall curve (AUPRC) of 0.72 for detecting OMI cases. Performance was high for OMI with ST-elevation (AUROC 0.98) and moderate for OMI without ST-elevation (AUROC 0.84, AUPRC 0.20), reflecting the inherent difficulty of detecting non-ST-elevation occlusions. Among OMI cases with available door-to-procedure timestamps, the model identified 59 (53%) of those with delayed door-to-procedure time (>30 minutes), of which 32 (54%) were NSTEMIs. These findings highlight the potential of AI-enabled ECG interpretation to improve early recognition and timely management of high-risk OMIs.</p>