Anatomically indexed aortic valve calcium score more accurately predicts transaortic peak velocities and gradients compared to radiomics features in patients with severe aortic stenosis.

Abstract

Background: This study examines the use of computed tomography (CT)-derived calcified and non-calcified aortic valve (AV) features, including radiomics-based quantitative imaging biomarkers, for predicting aortic stenosis (AS) severity and evaluating sex-specific differences.

Methods: In this retrospective, single-center study, 270 patients (50 ​% female) with severe AS and preserved left ventricular ejection fraction were assessed in the primary-cohort using both echocardiography and CT angiography. Correlation-based feature selection and Lasso regression were employed to refine the most predictive features. Logistic regression models were developed for the overall-, male-, and female-cohorts, evaluating the predictive power of calcified and non-calcified AV features for identification of peak aortic valve jet velocity (PAV) ​≥ ​4 ​m/s and mean pressure gradient (MPG) ​≥ ​40 ​mmHg.

Results: Statistical methods reduced the initial 44 CT variables to 13 in overall-cohort models, 10 in male-cohort models, and 12 in female-cohort models. The inclusion of these additional features significantly improved model performance compared to using the AV calcium score (AVCS) alone or its indexed variants. Indexing the AVCS to anatomical features resulted in modest improvements, with ROC-AUC values increasing from 0.71 (non-indexed Agatston) to 0.78 (indexed to sinus-of-Valsalva (SOV) volume) for PAV prediction in overall-cohort models. However, incorporating the full set of selected features further enhanced predictive accuracy, raising the ROC-AUC to 0.80. Similar trends were observed for MPG, with the best-performing models achieving a ROC-AUC of 0.84 compared to 0.73 using the non-indexed AVCS score alone. The male- and female-cohort models demonstrated similar improvements, with sex-specific feature sets significantly enhancing performance beyond indexed AVCSs.

Conclusions: Indexing the AVCS to SOV volume and aortic annulus area enhances the predictive power of AS severity models, though incorporating a broader set of calcified and non-calcified CT features provides the greatest improvement. These findings underscore the importance of considering anatomical and sex-specific differences in the assessment of AS.