Echocardiographic phenogrouping by machine learning for risk stratification in the general population

Abstract

Abstract Background Improved taxonomical classification of routinely measured echocardiographic parameters is needed for better phenotypic characterisation of the asymptomatic stages of cardiac maladaptation. This would create opportunities to intervene early in the course of the heart disease and prevent progression to more advanced stages and adverse events. Purpose We tested the hypothesis that an unbiased clustering analysis utilizing echocardiographic indexes reflecting left heart structure and function and hemodynamic measurements could identify phenotypically distinct groups of asymptomatic individuals in the general population. Methods We prospectively studied 1407 community-dwelling individuals (mean age, 51.0 years; 51.5% women), in whom we performed clinical and echocardiographic examination at baseline and collected cardiac events on average 8.5 years later. Cardiac phenotypes that were correlated at r>0.8 were filtered, leaving 15 less redundant echocardiographic and hemodynamic features for phenogrouping. We employed two methods of unsupervised machine learning: agglomerative hierarchical clustering and Gaussian mixture using expectation minimization algorithm. The optimal number of phenogroups was chosen based on the combination of the cohesion/separation approach (silhouette index) and stability. Cox regression was used to demonstrate the clinical validity of phenogroups. Results Overall, both methods agreed with respect to cluster assignment (RI=0.75). Unbiased clustering analyses classified study participants into 3 distinct phenogroups that differed markedly in cardiac structure/function indexes and hemodynamics used for cluster analysis (Figure, left panel). Indeed, phenogroup 3 had the worst left ventricular diastolic function (ie, lowest e' velocity and left atrial reservoir strain, but highest E/e', deceleration time, and left atrial volume index), highest left ventricular mass index, as well as highest systolic blood pressure and pulse pressure (Figure, left panel). The phenogroups were also different in other clinical characteristics and incidence of cardiac events (Figure, right panel). Even after adjustment for traditional risk factors, phenogroup 3 had the highest risk of cardiac events as compared to cluster 1 (Figure, right panel). Conclusions Unsupervised learning algorithms integrating routinely measured cardiac imaging and hemodynamic data can provide a clinically meaningful classification of cardiac health in asymptomatic individuals. This approach might facilitate early detection of cardiac maladaptation and improve risk stratification. Funding Acknowledgement Type of funding source: Public grant(s) – National budget only. Main funding source(s): Research Foundation Flanders