Phenomapping a Novel Classification System for Patients with Destination LVADs: An Unsupervised Machine Learning Cluster Analysis

Abstract

Introduction Patients with continuous flow destination therapy (DT) left ventricular assist devices (LVAD) are a heterogenous population that may have distinct phenotypic clusters. Hypothesis We hypothesize there will be distinct phenotypic clustering of individuals with DT LVADs by their clinical characteristics at implantation that are associated with different long-term risk profiles. Methods We analyzed 8245 patients with a continuous flow DT LVADs in INTERMACS and selected 25 baseline clinical characteristics including demographic, echocardiographic, hemodynamic, and laboratory data. Variables with >10% missingness were excluded except for left ventricular dimensions which were imputated. Unsupervised K-means clustering derived phenogroups and the optimal number of clusters was assessed by minimization of the Bayesian Information Criterion. The final dataset included 5999 patients with 18 selected variables for 4 distinct phenogroups. Survival analyses for events considered competing risk for cumulative incidence of transplant or the composite endpoint of death or heart transplant when appropriate. Results Of the 4 phenogroups, phenogroup 1 (n=1163, 19%) was older (median age 71 years), primarily white (81%), and most frail (9%). Phenogroup 4 (n=517, 9%) was younger (41 years), heavier (108 kg), and more diverse (white race 46%) with the least frailty (3%), largest left ventricles (LVEDD 7.5 cm), and highest rate of non-adherence (8%) amongst the clusters. Phenogroups 2 (n=648, 11%) and 3 (n=3671, 61%) were of intermediate age (69 & 61 years), weight (87 & 90 kg), and ventricular size (LVEDD 6.5 & 6.9 cm). The cumulative incidence of death, heart transplant, bleeding, LVAD malfunction and LVAD thrombosis differed among phenogroups (Figure A-D). The highest cumulative incidence of death and the lowest rate of heart transplant was seen in phenogroup 1 (P Conclusions Our exploratory unsupervised machine learning analysis highlighted 4 phenogroups that differed in fatal and non-fatal adverse events. These observations underscore the influence of phenotypic heterogeneity on post-LVAD implantation outcomes.