Prediction of major adverse cardiovascular events (MACE) in Agatston low-risk patients using comprehensive AI analysis of low-cost screening CT calcium score exams

Coronary artery calcification Agatston score, calculated from CT calcium score exam, is widely used in cardiovascular risk prediction, particularly for determining the necessity of statin therapies in patients with Agatston score >100, as recommended by the American Heart Association (AHA). However, as high Agatston score is indicative of MACE, patients with low or zero Agatston are more likely not been detected. There is growing recognition that CT-based assessments of epicardial adipose tissue (EAT) are associated with cardiovascular events. This study aimed to enhance cardiovascular disease prediction using pathophysiologically-driven EAT assessments obtained from CTCS images in individuals with Agatston scores below 100. We analyzed a cohort of 16,042 patients (1.7% MACE) with low Agatston score (Ag<100) from the HU-CLARIFY registry (NCT04075162). EAT was segmented using our DeepFat segmentation model from non-contrast CTCS exams, generating 216 EAT hand-crafted features (fat-omics) based on morphology, intensity, and spatial location. The cohort was divided into training/testing sets (80%/20%). We employed Cox proportional hazards regression model with elastic net regularization, incorporating EAT, Agatston score and/or clinical features to predict MACE. EAT fat-omics (c-index of train/test = 0.68/0.63) outperformed Agatston score (0.61/0.55). Our proposed model include adding Agatston score and clinical features which improve results significantly (c-index=0.77/0.74). In conclusion, advanced, pathophysiological EAT features may address a valuable risk assessment tool (that Agatston missed) for patients with low Agatston scores, enhancing early detection and management strategies.