AI-MI: A Deep Learning Model to Predict Actionable Acute Coronary Syndrome Using 12-Lead ECGs

Background Chest pain is among the most common chief complaints in Emergency Departments (EDs), and differentiating acute coronary syndrome from low-risk chest pain etiologies is often nontrivial. Early and accurate identification of acute coronary syndrome (ACS) is a key quality metric in Emergency Departments as diagnostic delay results in myocardial infarction, while unnecessary coronary angiography leads to adverse effects, prolonged hospital stays, and poor utilization of cardiac catheterization laboratory resources. Objectives We present AI-MI, a novel deep learning technology that predicts whether a patient presenting with chest pain will undergo same-encounter coronary revascularization. Methods We use a 1D ResNet convolutional neural network architecture pretrained on an arrythmia- detection task as a foundation model, and we train it using Electronic Health Record data from NYU Langone Health EDs from 2009 to 2024. We define three cohorts within this dataset – patients from whom a troponin has been ordered, those with a positive troponin, and those scheduled for coronary angiogram – to evaluate our model’s applicability at sequential stages in the ACS diagnostic/therapeutic algorithm. Results AI-MI achieves AUROCs of 0.91, 0.81, and 0.67, and AUPRCs of 0.35, 0.39, and 0.57, within the “troponin ordered,” “troponin positive,” and “coronary angiogram” cohorts respectively. AI- MI trained using sequential ECGs consistently and significantly outperformed that using individual ECGs. Conclusions AI-MI can be used in the point-of-care evaluation of chest pain to improve differentiation of ACS from low-risk chest pain, both promoting early cardiology consultation for revascularization and decreasing the rate of unnecessary coronary angiography.