부산대학교 도서관

Background:Peak global longitudinal strain (Peak-GLS) is recognized as a powerful predictor of heart failure (HF). However, the entire strain curve remain undiscovered and therefore important prognostic information regarding HF might be lost.Hypothesis:We hypothesized that analysis of the strain curve using unsupervised machine learning (uML) would reveal novel ventricular deformation patterns capable of predicting incident HF independently of Peak-GLS.Methods:Longitudinal strain curves from 3,767 subjects from the general population without prevalent HF were analyzed using uML.Results:Mean age was 56 years and 43% were male. During a median follow-up of 5.3 years, 94 subjects (2.5%) developed HF. The uML algorithm generated a hierarchical clustering tree (HCT)(Figure 1a) resulting in 9 different clusters with mean strain curves (figure 1b and 1c). Figure 1a summarizes the mean age, IRHF and number of subjects in each cluster. In multivariable Cox regression cluster 3 was significantly associated with reduced incidence of HF when compared to cluster 2 [HR 0.13, 95%CI: 0.04;0.41, P<0.001]. Cluster 4 was also significantly associated with reduced incidence of HF compared to cluster 2 [HR 0.21, 95%CI: 0.08;0.57, P=0.002]. Cluster 2 was associated with increased incidence of HF compared to cluster 3 and 4 even though cluster 2 displayed healthier clinical baseline characteristics as well as higher EF and Peak-GLS. The mean strain curve of cluster 2 had an increased absolute peak value, more rapid decline during systole, faster increase during diastole and a larger diastolic curvature compared to the mean strain curves of cluster 3 and cluster 4.Conclusion:uML was capable of identifying a cluster (cluster 2) exhibiting a deformation pattern associated with increased risk of HF despite having higher EF and peak-GLS. The deformation patterns of cluster 2 was characterized by a faster contraction, faster relaxation and a higher deacceleration after early filling.