AI detection of calcification in the aortic valve and coronary arteries independently predicts mortality beyond COMPERA 2.0 and REVEAL 2.0 scores in PAH and CTEPH
N Karthikeyan, A Hokmabadi, M Sharkey, S Alabed, A Hameed, M Salehi, A Charalampopoulos, D Kiely, A J SwiftAbstract
Background
Current risk stratification in pulmonary arterial hypertension (PAH) and chronic thromboembolic pulmonary hypertension (CTEPH) relies heavily on functional capacity and markers of right-heart strain (e.g., COMPERA 2.0, REVEAL 2.0), overlooking both aortic valve calcification (AVC) and coronary artery calcification (CAC). While both are frequently under-reported on non-gated scans, AVC may additionally be masked on echocardiography by the reduced stroke volume of the right ventricle.
Purpose
We aimed to determine if opportunistic, AI-derived quantification of AVC and CAC from routine, non-ECG-gated chest CTs (contrast and non-contrast) could independently predict mortality in PAH and CTEPH.
Methods
An automated deep learning pipeline (utilising 3D ResNet-18 to exclude prosthetic valves and nnU-Net for calcification segmentation) was used to extract AVC and CAC volumes from CTs. We evaluated 548 patients from the ASPIRE registry undergoing contemporaneous right heart catheterisation (RHC), echocardiography, and CT. As the prognostic signal in the full cohort was driven by the PAH and CTEPH groups, subsequent analysis targeted this sub-cohort (n=239). To investigate subclinical disease, the multivariable Cox model was restricted to patients with a non-stenotic echocardiographic profile (peak aortic velocity < 2.5 m/s), yielding 232 patients.
Results
Within the final cohort (n=232), AI identified subclinical AVC in 13.8% and CAC in 53.9% of patients. Baseline invasive haemodynamics were comparable between the AVC-absent and AVC-present groups, including pulmonary vascular resistance (PVR) (651.6 vs. 592.8 dyn·s·cm⁻⁵, p=0.465) and pulmonary arterial wedge pressure (PAWP) (11.2 vs. 11.1 mmHg, p=0.878). Over a median follow-up of 57.7 months, the unadjusted mortality rate was significantly higher in the AVC-present cohort (Log-rank p < 0.0001; Figure 1). In multivariable analysis adjusted for COMPERA 2.0 scores, the presence of AVC (HR: 1.97, 95% CI: 1.22 - 3.17, p = 0.006) and CAC (HR: 1.89, 95% CI: 1.27 - 2.82, p = 0.002) were independent predictors of all-cause mortality. Notably, these findings were conserved when adjusting for the REVEAL 2.0 score (AVC HR: 1.63, 95% CI: 1.00 - 2.65, p=0.048; CAC HR: 1.55, 95% CI: 1.04 - 2.32, p=0.033), confirming additive prognostic value across two distinct clinical risk tools (Figure 2). Integrating these calcific markers into the COMPERA 2.0 model numerically improved Harrell’s C-index from 0.607 to 0.674 (1,000-iteration bootstrapped 95% CI: -0.048 to 0.063).
Conclusion
Automated opportunistic extraction of cardiac calcification from routine CTs identifies high-risk phenotypes in PAH and CTEPH. This provides independent, additive prognostic value beyond the established COMPERA 2.0 and REVEAL 2.0 scores, demonstrating the clinical utility of evaluating cardiovascular comorbidities to improve risk stratification in this population without additional radiation or cost.Figure 1:Kaplan-Meier Survival CurvesFor image description, please refer to the figure legend and surrounding text.Figure 2:Cox PH Forest PlotsFor image description, please refer to the figure legend and surrounding text.