Electrocardiogram-Based Deep Learning Prediction of Coronary Artery Calcium and Adverse Cardiovascular Outcomes in South Asians. - PROJECT SUMMARY/ABSTRACT South Asians (SA) make up over a quarter of the world’s population and represent one of the fastest-growing ethnic groups in the United States. SA have a disproportionately high prevalence of atherosclerotic cardiovascular disease (ASCVD) compared to other groups, which often occurs at younger ages and leads to worse outcomes. SA descent is a risk enhancer for ASCVD. Nonetheless, current prediction tools underestimate the risk in this population, likely due to underrepresentation in derivation studies. Coronary artery calcium (CAC) is a highly specific marker that enhances ASCVD prediction in diverse populations, including SA. However, implementation of CAC is limited in certain clinical settings such as remote, low-resource, or underserved communities. Our scientific premise is grounded in the critical need to identify high-risk individuals who could benefit from timely preventive interventions. Artificial intelligence (AI)-based deep learning methods have revolutionized the diagnostic potential of the 12-lead electrocardiogram (ECG-AI), a widely available and affordable tool. ECG-AI has accurately predicted CAC and identified individuals at increased risk of ASCVD, providing a practical, accessible, and scalable enhancement to the primary prevention of ASCVD. These advancements offer an opportunity to address gaps in care, particularly in underserved populations. The overarching goal of this project is to investigate the correlates, predictive performance, and clinical utility of an ECG-AI model that estimates CAC, specifically in SA, by leveraging the uniquely characterized Precision-Cardiometabolic Risk Reduction in South Asians (P-CARRS) Cohort. To achieve this goal, we propose the following specific aims: Aim 1. To investigate sociodemographic, behavioral, and biological mechanisms underlying ECG-AI-based CAC scores in SA and compare them to computed tomography (CT)- based CAC. Aim 2. To evaluate the agreement between ECG-AI-based CAC scores and CT- based CAC scores in SA as continuous measures and clinically actionable thresholds (e.g., CT-derived CAC = 0, 1–99, 100–299, >300), we will calculate model performance metrics (e.g., area under the curve, sensitivity, specificity, and accuracy), agreement metrics (e.g., concordance correlation coefficients and Bland-Altman plots), and calibration. Aim 3. We will assess the association between ECG-AI-based CAC scores derived from ECGs and incident ASCVD events in South Asians. ECG-AI will be evaluated both as a standalone and additive predictor alongside traditional risk factors while analyzing potential improvements in predictive accuracy and calibration. The proposed research is an important next step after an initial landmark finding and will also provide important training for the applicant as he seeks to train in a career in translational research. The proposed in- depth examination of this new ASCVD screening tool will not only have important public health implications for this high-risk SA group but also inspire similar investigations in other groups with disparate ASCVD outcomes. Lastly, it will set the stage for future career development awards evaluating interventions and other populations.
