Journal of Translational Medicine (Oct 2023)

Circulating miR-133a-3p defines a low-risk subphenotype in patients with heart failure and central sleep apnea: a decision tree machine learning approach

  • David de Gonzalo-Calvo,
  • Pablo Martinez-Camblor,
  • Thalia Belmonte,
  • Ferran Barbé,
  • Kevin Duarte,
  • Martin R. Cowie,
  • Christiane E. Angermann,
  • Andrea Korte,
  • Isabelle Riedel,
  • Josephine Labus,
  • Wolfgang Koenig,
  • Faiez Zannad,
  • Thomas Thum,
  • Christian Bär

DOI
https://doi.org/10.1186/s12967-023-04558-w
Journal volume & issue
Vol. 21, no. 1
pp. 1 – 13

Abstract

Read online

Abstract Background Patients with heart failure with reduced ejection fraction (HFrEF) and central sleep apnea (CSA) are at a very high risk of fatal outcomes. Objective To test whether the circulating miRNome provides additional information for risk stratification on top of clinical predictors in patients with HFrEF and CSA. Methods The study included patients with HFrEF and CSA from the SERVE-HF trial. A three-step protocol was applied: microRNA (miRNA) screening (n = 20), technical validation (n = 60), and biological validation (n = 587). The primary outcome was either death from any cause, lifesaving cardiovascular intervention, or unplanned hospitalization for worsening of heart failure, whatever occurred first. MiRNA quantification was performed in plasma samples using miRNA sequencing and RT-qPCR. Results Circulating miR-133a-3p levels were inversely associated with the primary study outcome. Nonetheless, miR-133a-3p did not improve a previously established clinical prognostic model in terms of discrimination or reclassification. A customized regression tree model constructed using the Classification and Regression Tree (CART) algorithm identified eight patient subphenotypes with specific risk patterns based on clinical and molecular characteristics. MiR-133a-3p entered the regression tree defining the group at the lowest risk; patients with log(NT-proBNP) ≤ 6 pg/mL (miR-133a-3p levels above 1.5 arbitrary units). The overall predictive capacity of suffering the event was highly stable over the follow-up (from 0.735 to 0.767). Conclusions The combination of clinical information, circulating miRNAs, and decision tree learning allows the identification of specific risk subphenotypes in patients with HFrEF and CSA.

Keywords