Abstract Body (Do not enter title and authors here): Background: Aortic valve calcification (AVC) is an early marker of aortic stenosis but is not routinely scored and reported in coronary artery calium (CAC) scans. Traditional manual quantification methods are time-consuming and subject to inter-reader variability, limiting their scalability in clinical practice. The AI-CVD initiative (see figure 1) aims to extract all useful opportunistic screening information from CAC scans to maximize prediction of cardiovascular diseases (CVD).
Hypothesis: We hypothesized that the AI-CVD measurements of AVC in CAC scans performs comparably to manual measurements by human experts.
Methods: We analyzed 742 CAC scans from the Framingham Heart Study Offspring cohort, acquired between 2002 and 2005 who also had an echocardiogram with aortic valve analysis. The AI model was trained to segment each aortic valve leaflet and quantify calcification using the Agatston criteria (>130HU and ≥1mm2). Manual AVC scores were previously computed by trained huamn experts. The primary outcome was clinically diagnosed aortic stenosis, identified cross-sectionally from health records. We compared the performance of AI-CVD and manual AVC scores using the area under the receiver operating characteristic (AUC) analysis and confusion matrix metrics, including sensitivity, specificity, positive predictive value, negative predictive value, and F1 score.
Results: Among 742 participants, 22 had a clinical diagnosis of aortic stenosis. 18 cases of aortic stenosis were correctly identified by AI-measured AVC compared to 15 cases by manual measurements. The AUC for aortic stenosis diagnosis for AI-CVD AVC was 0.934 vs. 0.883 for huamn experts. Confusion matrices and ROC curves support the improved accuracy and lower false negative rate of AI-based measurement versus human experts.
Conclusion: AI-CVD AVC score detects aortic stenosis comparably to manual measurements by huamn experts in the Framingham Heart Study Offspring Cohort. These findings support the integration of AI-enabled opportunistic screening into CAC scans to maximize CVD prediction beyond the CAC score.