Feature-based clustering of left ventricular strain curve for risk stratification in the general population

Abstract

Abstract Objective Global left ventricular (LV) strain has been identified as an important predictor of adverse cardiovascular (CV) events. However, currently only the peak strain is used from the LV deformation curve, neglecting the temporal information hidden in all phases of cardiac cycle. Therefore, we employed unsupervised machine learning methods on time-series-derived features from LV strain to identify distinct clinical phenogroups associated with the risk of developing adverse events in the general population. Design and Method We prospectively studied 1185 community-dwelling individuals (mean age, 53.2 years; 51.3% women), in whom we acquired clinical and echocardiographic data including LV strain traces at baseline and collected CV events (n=116) on average 8.7 years later. We utilised Gaussian Mixture Model (GMM) on different features derived from LV strain curve, including slope during systole and early diastole, peak strain, slope during late diastole and the duration and height of diastasis (Figure, left panel). We evaluated the clinical significance of the clusters (k) by comparing the clinical characteristics and CV adverse outcome. Results Based on inertia and BIC scores the optimal number of clusters (k) was 4. The first two clusters had differences in heart rate, but had similar low CV risk profiles. Cluster 4 had the worst CV risk factors combination, and higher prevalence of LV remodelling and diastolic dysfunction (i.e. lowest e’ velocity and highest E/e’) compared to other clusters. Kaplan–Meier showed an increased cumulative incidence risk for all CV events across clusters (Figure, right panel). After adjustment for traditional risk factors, the risk was significantly higher in clusters 3 (HR: 1.28; 95% CI: 1.01-1.61; P=0.038) and 4 (HR: 1.20; 95% CI: 1.02-1.43; P=0.034) compared to the average population, while the risk of adverse events did not reach the significant level in subjects with an abnormal peak LV strain (<17%) (HR 1.12; P=0.66). Conclusion Unsupervised machine learning algorithm employed on features derived from time series LV strain curves identifies clinically meaningful clusters, and provides important prognostic information over the peak LV strain.