Abstract 4145329: Automated Detection of ST-elevation Myocardial Infarction in High-Acuity Settings Using Artificial Intelligence Applied to Images of 12-lead Electrocardiograms

Background: The diagnosis of suspected ST-elevation myocardial infarction (STEMI) requires electrocardiograms (ECGs) done in emergent settings across the world. Despite access to ECG machines, the expertise of clinicians reading them is scarce. Automated reads are often inaccurate. We sought to address this challenge by developing an AI-enabled approach for automated cardiologist-level diagnosis from photos of 12-lead ECGs. Methods: We identified patients undergoing ECGs in a large, diverse US health system (2015-2023). We used self-supervised pretraining to develop a convolutional neural network to first detect ECG features without any diagnostic flags, designed to enhance the efficiency of training a STEMI-specific model. To ensure the model learned the signature for STEMI independent of other ECG features, each STEMI ECG (cases defined by reads by cardiologists) was matched to 10 ECGs of those without STEMI (controls) but of the same age, sex, and with the same set of conduction and rhythm abnormalities on their ECGs. To mimic variations in ECG formats in real-world clinical settings, the training ECGs were randomly plotted using any of 2880 unique formats, with variations in lead layouts, background colors, label fonts, grid, and signal line widths. The model was evaluated on a held-out set of ECG images from patients presenting to the emergency (ED) with chest pain. Results: The model development set included 49,995 ECGs from 39,697 individuals (age 65 years [IQR, 52-76], 18,926 [48%] women, 5941 [15%] Black). The test set included 7551 patients who had presented to the ED with chest pain (1 ECG/individual, median age 50 years [IQR, 34-62], 4063 [54%] women, 1661 [22%] Black). Of these, 104 (1.4%) had STEMI. The model achieved an AUROC of 0.95 (95% CI, 0.93-0.96) and AUPRC of 0.21 (0.15-0.28), with a sensitivity of 90.3% (84.0-95.8) and specificity of 87.7% (87.0-88.4). The model performance was consistent among demographic subgroups of age (AUROC 0.95-0.97), sex (AUROC 0.93-0.97), and race/ethnicity (AUROC 0.95-0.96). Conclusion: We have developed an AI algorithm that can diagnose STEMI on photos of ECGs with high accuracy, potentially enabling the triaging of patients in low-resource settings by offering accurate diagnosis at the point of care. The reliance on ECG images across formats and layouts ensures its use without changing ECG acquisition or related infrastructure.