AI-CVD-HF: A heart failure risk prediction model based on coronary artery calcium scans compared with PREVENT-HF

AbstractBackgroundThe AI-CVD initiative seeks to extract actionable information from coronary artery calcium (CAC) scans beyond the CAC score. We aimed to develop a heart failure (HF) prediction model, AI-CVD-HF, based on AI-derived features from non-contrast CAC scans, and compare it with PREVENT-HF.MethodAI features from CAC scans of 6743 asymptomatic participants in the Multi-Ethnic Study of Atherosclerosis (MESA) and Framingham Heart Study–Offspring (FHS-O) (mean age 62.3 ± 10.1; 47.2% male; median follow-up 17.1 years; 429 HF events) were analyzed. Features were selected using random forest and modeled with random survival forest. Performance was assessed against the base PREVENT-HF using 5-fold cross-validation for area under the receiver-operating-characteristic curve (AUC), area under the precision-recall curve (AUPRC), net-benefit, and calibration. External validation experiment was conducted to assess generalizability.ResultsSelected features for AI-CVD-HF model included AI-CAC score, left-to-right ventricular volume ratio, left atrial volume, left ventricular mass, visceral fat volume, skeletal muscle mean density, thoracic aortic calcification, age and sex. The 10-year AUC of AI-CVD-HF was 0.83 (95% CI:0.81–0.85), compared with PREVENT-HF (0.79, 95% CI:0.75–0.83; p = 0.01) and also demonstrated 32% higher AUPRC (0.25[95% CI:0.11–0.39] vs 0.19[95% CI:0.07–0.31]; p = 0.049), and consistent performance across age, sex, and race/ethnicity. Calibration metrics favored AI-CVD-HF over PREVENT-HF (Brier score [0.0356 vs 0.0383], slope [1.02 vs 1.20]). External validation showed performance and calibration consistent with internal validation.ConclusionsAI-CVD-HF, using AI-derived features from CAC scans, demonstrated more favorable discrimination and calibration than PREVENT-HF for HF prediction, extending the utility of CAC scans beyond coronary artery disease risk assessment.