Decoding calcific mitral valve disease: a novel deep learning model uncovers the role of calcium burden

Abstract

Abstract Background Mitral annular calcification (MAC) is a common degenerative disease that affects older adults and causes calcium deposition in the mitral annulus, increasing the risk of mitral valve (MV) dysfunction, cardiovascular events, conduction abnormalities, and mortality. While the primary imaging modality recommended for the clinical evaluation of MAC is echocardiography, it has been reported to be found incidentally in approximately 8% of routine thoracic computed tomography (CT) scans. CT scan is a sensitive imaging modality for assessing MAC, providing the resolution necessary to understand the anatomical involvement of calcification. Research has shown that the MAC burden is associated with disease activity and progression. Still, the exact correlation between the quantitative calcium burden and the degree of severity of mitral valve disease is unclear. Purpose Our study aimed to determine the association between MAC, as assessed on CT, and significant mitral stenosis (MS) or mitral regurgitation (MR), as visualised on echocardiography. In addition, we sought to determine the quantitative MAC cutoffs that would optimally predict significant MS or MR on echocardiography. Methods A retrospective cohort study was conducted at our academic centre. Inclusion criteria were: age≥60 years, resting transthoracic echocardiogram performed between 2013-2022, non-gated chest CT performed within one year before echocardiography, and any degree of MAC documented on the echocardiography report. CT scans were requested for clinical indications unrelated to mitral valve disease. Exclusion criteria were: prior mitral valve intervention, endocarditis, and congenital mitral valve disease. Significant calcific mitral valve disease was defined as ≥moderate MR and ≥mild MS on the echocardiography report, ascertained by expert readers based on multiparametric ASE criteria. MAC volume was quantified on the multi-slice CT DICOM images using a 3D U-Net deep neural network previously trained by our group on an independent cohort. Results The cohort consisted of 1,560 unique patients with a mean age of 79 years and 61% females. The echocardiographic prevalence of significant MR and MS was 10% and 4%, respectively, for 211 affected patients. The CT-based mean MAC volume was 949 mm³ in patients with significant MR or MS, as opposed to 334 mm³ in those without (P<0.001). MAC volume >1000 mm³ was the optimal cutoff to predict significant MS (specificity 90%, sensitivity 51%, area under ROC curve 0.79) and significant MR (specificity 89%, sensitivity 21%, area under ROC curve 0.60). Adjusting for age, sex, and comorbid conditions did not affect the observed association between MAC volume and mitral valvulopathy. Conclusion A novel deep learning model for quantifying MAC volume from non-cardiac clinical CT scans was efficient in screening for calcific mitral valve disease, particularly MS, and identifying patients who may benefit from further evaluation.