The challenges of predicting right heart failure (RHF) post-Left Ventricular Assist Device (LVAD) may reflect heterogenous underlying pathophysiology. We hypothesized that 1) machine learning (ML) algorithms applied to multidimensional phenotypic data from patients with confirmed post-LVAD RHF will allow identification of distinct RHF phenotypes, 2) identified phenotypes will have unique clinical trajectories.
Patients with acute post-LVAD RHF (RVAD and/or ≥ 14 days inotropes post-implant, n=2,550) were identified from the ISHLT Mechanically Assisted Circulatory Support database (n=15,428); and divided into a derivation (DC, n=1,531) and validation cohort (VC, n=1,019). First, unsupervised ML (blinded to clinical outcomes) was applied to 41 pre-implant variables to identify distinct phenotypes. Then, resultant phenotypes were clinically validated by comparing outcomes of 1) RVAD/ death during index hospitalization 2) ICU Length of Stay. Results were validated in the VC. Risk discrimination of existing RHF risk scores was compared between phenotypes.
Four distinct RHF phenotypes were identified. (Figure 1) Phenotype I had the worst, and Phenotype III had the best outcomes. Results were validated in the VC. RHF risk scores were modestly accurate at predicting RHF in those with severe shock (Phenotype I) pre-implant; but performed poorly for phenotypes without prominent shock. (Table 1)
ML identifies novel pathophysiological phenotypes of RHF, among which current risk scores were useful to predict RHF only in patients in severe shock prior to implant.