Unsupervised clustering of patients with severe aortic stenosis: A myocardial continuum

Abstract

BackgroundTraditional statistics, based on prediction models with a limited number of prespecified variables, are probably not adequate to provide an appropriate classification of a condition that is as heterogeneous as aortic stenosis (AS). AimsTo investigate a new classification system for severe AS using phenomapping. MethodsConsecutive patients from a referral centre (training cohort) who met the echocardiographic definition of an aortic valve area (AVA) ≤ 1 cm2 were included. Clinical, laboratory and imaging continuous variables were entered into an agglomerative hierarchical clustering model to separate patients into phenogroups. Individuals from an external validation cohort were then assigned to these original clusters using the K nearest neighbour (KNN) function and their 5-year survival was compared after adjustment for aortic valve replacement (AVR) as a time-dependent covariable. ResultsIn total, 613 patients were initially recruited, with a mean±standard deviation AVA of 0.72±0.17 cm2. Twenty-six variables were entered into the model to generate a specific heatmap. Penalized model-based clustering identified four phenogroups (A, B, C and D), of which phenogroups B and D tended to include smaller, older women and larger, older men, respectively. The application of supervised algorithms to the validation cohort (n=1303) yielded the same clusters, showing incremental cardiac remodelling from phenogroup A to phenogroup D. According to this myocardial continuum, there was a stepwise increase in overall mortality (adjusted hazard ratio for phenogroup D vs A 2.18, 95% confidence interval 1.46–3.26; P<0.001). ConclusionsArtificial intelligence re-emphasizes the significance of cardiac remodelling in the prognosis of patients with severe AS and highlights AS not only as an isolated valvular condition, but also a global disease.