Long-Term Risk Prediction and Detection of Aortic Stenosis From AI-Quantified Multisite Calcification on Noncontrast CT: An AI-CVD Analysis of MESA

Submitted manuscript | Original research review

Aortic stenosis (AS) can remain clinically silent until advanced disease, while structural changes may begin earlier. This MESA-focused study evaluated whether AI-derived cardiovascular phenotypes measured on noncontrast cardiac CT are associated with incident AS and can improve long-term prediction, alongside a separate model for AS detection.

Study design

The prediction analysis included 5,520 MESA participants with baseline noncontrast cardiac CT. Incident AS was identified through adjudicated events and Exam 6 echocardiography. Researchers assessed associations using multivariable Cox models and internally validated a random survival forest with nested cross-validation. A separate detection model was developed in 1,857 MESA Exam 6 participants with paired CT and echocardiography, then externally evaluated in the Framingham Heart Study.

Key findings

During a median 16.7 years of follow-up, 157 participants developed AS. Aortic valve calcium showed the strongest association with incident disease (hazard ratio per standard-deviation increase, 4.34; 95% CI, 3.33-5.66). At 10 years, the AI-CVD model using age, sex, aortic valve calcium, mitral valve calcium, and thoracic aortic calcification achieved an AUROC of 0.973, compared with 0.809 for a model based on routinely measured clinical factors. The corresponding AUPRC values were 0.256 and 0.024. In external validation, the detection model achieved an AUROC of 0.946, sensitivity of 87.8%, and specificity of 94.5%.

Interpretation

These findings suggest that AI-based quantification of multisite calcification on noncontrast CT may help identify long-term AS risk and detect existing disease. The prediction model was internally validated and the detection model externally assessed; prospective evaluation is still needed to establish performance and clinical impact in routine care.

Abstract

Background: Aortic stenosis may remain silent until severe. Objective: To assess whether AI-derived cardiovascular phenotypes from noncontrast CT are associated with incident AS and can predict and detect disease. Methods: In 5,520 MESA participants, incident AS associations were analyzed and an AI-CVD prediction model was evaluated using nested cross-validation. A separate detection model was developed in 1,857 MESA participants and externally evaluated in Framingham. Results: Over median 16.7-year follow-up, 157 participants developed AS. Aortic valve calcium had the strongest association (HR 4.34; 95% CI 3.33-5.66). The 10-year AUROC was 0.973 for the AI-CVD model vs 0.809 for the clinical model; external detection AUROC was 0.946 (sensitivity 87.8%, specificity 94.5%). Conclusions: Automated CT phenotyping may support scalable prediction and detection of AS.

Status: Submitted manuscript. The full manuscript is not attached.

Download the abstract

Download Abstract (PDF)