Abstract 14862: Development and Validation of a Phenomapping Tool to Identify Patients With Diuretic Resistance in Acute Decompensated Heart Failure: A Multi-Cohort Analysis

Abstract

Introduction: Individuals presenting with acute decompensated heart failure (ADHF) have varying response to diuretic therapy and short- and long-term prognosis. Hypothesis: If machine learning can risk stratify patients with ADHF and identify subgroups at risk for diuretic resistance. Methods: Participants with ADHF from the ROSE-AHF and CARRESS-HF clinical trials were included (n=451) and clustered using multivariable finite-mixture models based on diuretic efficiency (fluid output over first 72 hours per total intravenous loop diuretic dose). Differences in diuretic efficiency, in-hospital length of stay, and in-hospital mortality were assessed using linear and logistic regression models. Phenogroups were externally validated in trial (DOSE/ESCAPE, ATHENA-HF) and real-world (GWTG-HF) cohorts. Results: Clustering identified 3 phenogroups. Participants in phenogroup 1 (n=271, 60%) had worse diuretic efficiency [median(IQR) = 11.6(6.6-17.9) mL/mg) compared with phenogroups 2 (n=145, 32%) and 3 (n=35, 8%) [median(IQR) = 16.3(11.2-23.9) and 20.2(12.3-49.9) mL/mg, respectively; p<0.001]. An integer-based risk score to predict phenogroup 1 (lowest diuretic efficiency) was created: BAN-ADHF ( Fig. ). Net urine output was 2600 vs. 1090mL per 24 hours in patients with scores of 5 and 15, respectively ( Fig ). In the external validation cohorts, participants with scores ≥11 vs. <11 had significantly lower global well-being, higher natriuretic peptide levels on discharge, longer length of stay, and higher risk of in-hospital mortality. Conclusions: We developed and validated a phenomapping strategy and risk score for individuals with ADHF and differential response to diuretic therapy, which was associated with length of stay and mortality.