Abstract 4366932: AI-CVD vs. PREVENT for Predicting Incident Heart Failure: The Multi-Ethnic Study of Atherosclerosis (MESA)

Background: The AI-CVD initiative aims to extract opportunistic screening information from coronary artery calcium (CAC) scans to maximize cardiovascular disease prediction beyond the traditional risk factors and the Agatston CAC score. Hypothesis: In 2024, the American Heart Association introduced the PREVENT heart failure (HF) risk score based on age, sex, systolic blood pressure, body mass index, glomerular filtration rate (GFR), diabetes, smoking, and anti-hypertensive medication consumption. We sought to compare PREVENT HF vs. AI-CVD risk scores for predicting HF in the Multi-Ethnic Study of Atherosclerosis (MESA). Method: AI-CVD platform is a collection of deep learning models targeting various componenets of a CAC scan (see figure 1). We applied AI-CVD to 4,554 CAC scans of asymptomatic MESA participants aged 45–84 years (46.9% male). We used selected AI-CVD outputs included cardiac chamber volumes, thoracic skeletal muscle volume and density, epicardial fat volume, percentage of lung emphysema (<950 HU), and percentage of liver fat (<40 HU). Clinical data comprised demographic and anthropometric characteristics, laboratory results, lifestyle factors, and electrocardiogram parameters. Embedded feature selection methods were applied to identify the most important predictors of HF. The AI-CVD risk score for incident HF was developed using FasterRisk, an interpretable machine learning technique. We then compared the performance of PREVENT HF vs. AI-CVD using the area under the receiver operating curve (AUC) and DeLong’s test for predicting HF. Results: After a median follow-up of 17.7 (IQR: 13.0-18.5) years, 265 (5.8%) cases were diagnosed with HF. Age, GFR, hypertension, anti-hypertensive medication consumption, smoking, microalbuminuria, diabetes, left atrial volume, ratio of left ventricle to right ventricle volume, left ventricular mass, CAC score, epicardial fat volume, and emphysema were selected features for predicting HF. The AUC for AI-CVD (AUC: 0.84 [95% CI:0.82-0.87]) was significantly (P < 0.001) higher than for PREVENT HF (AUC: 0.77, 95% CI: 0.74-0.81) for 10-year HF prediction. Conclusion: By integrating AI-generated opportunistic screening biomarkers from CAC scans with clinical data, the AI-CVD risk score significantly outperformed the PREVENT risk score for HF prediction in MESA participants over 10 years.