A deep learning model for predicting aortic stenosis based on coronary artery calcium (CAC) scans: an AI-CVD study within the Multi-Ethnic Study of Atherosclerosis (MESA)

Background Aortic stenosis (AS) remains frequently underdiagnosed, with late presentation associated with irreversible increases in morbidity and mortality despite successful intervention. Current diagnostic pathways rely on symptom onset or murmur detection, while population-wide echocardiographic screening remains cost-prohibitive, leaving a substantial burden of undetected disease in the general population. Early identification of high-risk individuals through predictive approaches is therefore critical, particularly given the emergence of disease-modifying therapies that may be most effective when initiated early in the disease course. The AI-CVD initiative seeks to extract actionable information from coronary artery calcium (CAC) scans beyond the CAC score. Purpose We aimed to develop an AS prediction model, AI-CVD-AS, based on AI-derived features from non-contrast CAC scans. Method AI-derived features from CAC scans of 5,556 asymptomatic participants in the Multi-Ethnic Study of Atherosclerosis (MESA) were analyzed (mean age 62.2 ± 10.3 years; 47.6% male; median follow-up 17.7 years; 209 AS events). AS was defined based on aortic valve area, peak aortic jet velocity, and mean transvalvular gradient assessed at MESA Exam 6 echocardiography, supplemented by ICD diagnostic codes. Features were selected using clinical relevance and random forest modeled with a random survival forest. Model performance was assessed via 5-fold cross-validation using the area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), net benefit, and calibration metrics. Results The AI-CVD-AS model incorporated the following features: aortic valve calcification score based on the Agatston method, number of discrete calcified lesions within the aortic valve segmentation, defined as clusters of contiguous voxels with CT attenuation ≥130 Hounsfield units and a minimum area of ≥1 mm², thoracic aortic calcification based on the Agatston method, aortic artery volume, left ventricular outflow tract volume, relative left ventricular wall thickness (left ventricular mass divided by left ventricular volume), right atrial to left ventricular volume ratio, bone mineral density, AI-derived CAC score, age, and sex. The model achieved a 10-year AUC of 0.88 (95% CI: 0.83–0.91) and an AUPRC of 0.32 (95% CI: 0.18–0.54; 10-year incidence rate: 1.5%), with consistent performance across age, sex, and racial/ethnic subgroups. AI-CVD-AS demonstrated favorable calibration, with a Brier score of 0.012 ± 0.001 and a calibration slope of 1.04. Conclusions AI-CVD-AS leverages AI-derived features from CAC scans to predict aortic AS, extending the utility of routine CAC scanning beyond coronary artery disease risk assessment. However, external validation, clinical trials, and cost-effectiveness analyses are needed before clinical implementation.