Time series clustering of left atrial strain curves for risk stratification in the general population

OBJECTIVE: Currently, only the peak reservoir left atrial (LA) strain has been in use while a huge amount of useful information during different phases of cardiac cycle has been ignored. In this study, therefore, we tested the hypothesis that an unsupervised time series analysis utilizing the whole LA deformation curves will identify distinct clusters associated to risk factors and improve cardiovascular (CV) risk prediction over the traditionally used covariables in the general population. DESIGN AND METHOD: We prospectively studied 929 community-dwelling individuals (mean age, 51.6 years; 52.9% women), in whom we acquired clinical and echocardiographic data including LA strain traces at baseline and collected cardiac events on average 7.2 years later. We employed two unsupervised time series techniques such as Self-Organizing map and Variational Autoencoder with Convolutional Neural Network as building block to cluster spatiotemporal patterns of LA strain curves. Clinical characteristics and cardiac outcome were used to evaluate the validity of the clusters (k). RESULTS: According to the quantization error value for every k, the optimal number of clusters was 5 for applied methods. Overall, both methods agreed with respect to cluster assignment. Figure (left panel) presents centroid LA strain curves which represent each cluster. The first three clusters had differences in sex distribution and heart rate, but had similar low CV risk profiles. On the other hand, cluster 5 had the worst CV risk factors combination, and higher prevalence of left ventricular remodelling and diastolic dysfunction (ie, lowest e' velocity and highest E/e') compared to other clusters. We also observed an increase in the risk for incidence of adverse events between cluster 5 and other clusters (Figure, right panel). After adjustment for traditional risk factors, cluster 5 had the highest risk of cardiac events as compared to cluster 1, 2 and 3 (HR: 1.34; 95%CI: 1.05-1.73); P = 0.020). CONCLUSIONS: Unsupervised Machine/Deep Learning algorithms employed on time series LA strain curves identifies clinically meaningful clusters of LA deformation, and provides incremental prognostic information over traditional risk stratification.

Jaar van publicatie:2022

Toegankelijkheid:Closed