Opportunistic Screening for ASCVD using a Multimodal Deep Learning Risk Prediction Model - PROJECT SUMMARY Atherosclerotic cardiovascular disease (ASCVD), such as stroke and heart attack, is the leading cause of morbidity and mortality globally, responsible for approximately 19 million deaths annually. Risk assessment is the cornerstone for primary prevention of ASCVD. Pooled Cohort Equations (PCE) are currently used to guide risk assessment and tailor preventive therapies. However, these and other risk prediction tools remain imperfect and have significant limitations including being static and based on a small number of simple clinical variables as well as having poor performance across diverse populations. Moreover, they do not incorporate imaging that may contain known prognostic biomarkers of future risk. For example, CT scans of the chest contain coronary artery calcium (CAC), thoracic aortic calcium (TAC), intrathoracic (IF), and body composition (BC) metrics. However, these biomarkers are not routinely reported in clinical practice nor accounted for in PCE. Finally, PCE has also been shown to misestimate risk in certain ethnicities despite identical risk profiles. We will develop a comprehensive graph-based fusion model, “ADMIRE” (AscvD Multimodal rIsk pREdiction), that incorporates imaging and non-imaging data across two diverse sites – Mayo Clinic and Emory Health System. Our multidisciplinary team of radiologists, informaticists, AI scientists, and preventive cardiologists will leverage our prior experience with developing fusion models for risk prevention. We will validate our previously developed biomarker segmentation models on a diverse cohort as well as develop a comprehensive semantic segmentation model that incorporates multiple known prognostic biomarkers (AIM 1). We will then apply debiasing techniques to develop ‘fair’ models and evaluate performance on cohorts stratified by demographic (e.g., race, gender, social determinants of health surrogates) and imaging (e.g., scanner type) factors (AIM 2). Finally, we will use a novel graph-based technique to create a fusion model to show performance against PCE and on stratified cohorts based on demographics (AIM 3). We designed this study such that dependencies between experiments are reduced. We hypothesize that multimodal fusion models incorporating imaging and non-imaging biomarkers will have a greater prognostic performance than PCE. We further hypothesize that algorithmic model debiasing can allow more effective at-risk prediction for minority patients in which PCE is known not to perform well. Our proposal could potentially allow greater opportunistic screening of patients for primary prevention of ASCVD and overcome limitations of current risk assessment tools such as PCE.
