Approaching Higher Dimension Imaging Data Using Cluster-Based Hierarchical Modeling in Patients with Heart Failure Preserved Ejection Fraction

Abstract

Abstract Heart failure with preserved ejection fraction (HFpEF) is a major cause of morbidity and mortality, accounting for the majority of heart failure (HF) hospitalization. To identify the most complementary predictors of mortality among clinical, laboratory and echocardiographic data, we used cluster based hierarchical modeling. Using Stanford Translational Research Database, we identified patients hospitalized with HFpEF between 2005 and 2016 in whom echocardiogram and NT-proBNP were both available at the time of admission. Comprehensive echocardiographic assessment including left ventricular longitudinal strain (LVLS), right ventricular function and right ventricular systolic pressure (RVSP) was performed. The outcome was defined as all-cause mortality. Among patients identified, 186 patients with complete echocardiographic assessment were included in the analysis. The cohort included 58% female, with a mean age of 78.7 ± 13.5 years, LVLS of −13.3 ± 2.5%, an estimated RVSP of 38 ± 13 mmHg. Unsupervised cluster analyses identified six clusters including ventricular systolic-function cluster, diastolic-hemodynamic cluster, end-organ function cluster, vital-sign cluster, complete blood count and sodium clusters. Using a stepwise hierarchical selection from each cluster, we identified NT-proBNP (standard hazard ratio [95%CI] = 1.56 [1.17–2.08]) and RVSP (1.37 [1.09–1.78]) as independent correlates of outcome. When adding these parameters to the well validated Get with the Guideline Heart Failure risk score, the Chi-square was significantly improved (p = 0.01). In conclusion, NT-proBNP and RVSP were independently predictive in HFpEF among clinical, imaging, and biomarker parameters. Cluster-based hierarchical modeling may help identify the complementally predictive parameters in small cohorts with higher dimensional clinical data.