Derivation and Validation of Three Novel Phenotypes of Cardiogenic Shock

Abstract

Cardiogenic shock (CS) mortality remains unacceptably high. Risk stratification of CS patients may improve trial design and treatment strategies. We tested the application of the expert consensus-based Society for Cardiovascular Angiography and Intervention (SCAI) CS stages in a retrospective dataset. Furthermore, we hypothesized that unsupervised clustering of CS patients is feasible and identifies clinically applicable phenotypes of CS. The Cardiogenic Shock Working Group (CSWG) registry is a retrospective multicenter registry of patients with all-cause CS. Descriptive analyses were performed on 1,414 CSWG patients to quantify the association of clinical variables and in-hospital mortality. In myocardial infarction-caused CS (AMICS) patients, a random forest classifier was used to identify variables that predict in-hospital mortality in CSWG. Consensus k means clustering was performed on these variables in 1,000 randomly assigned 85% samples of 408 AMICS patients from the CSWG training cohort and 701 AMICS patients from the Danish National Patient Registry (DNPR). CSWG patients with heart failure (CSWG-HF, n=480) were then assigned to the clusters. Variables associated with in-hospital mortality in all-cause CS included SCAI Stages, but also pulmonary artery catheter use, and venous congestion. In AMICS the mortality driving continuous variables comprised glomerular filtration rate, lactate, HCO3, platelets, alanine aminotransferase, and white blood cell count. Independent, unsupervised consensus k means clustering on these variables identified three phenotypes of CS that were associated with in-hospital mortality in similar patterns in CSWG, DNPR, and CSWG-HF. The phenotypes revealed unique demographic, clinical, metabolic and hemodynamic profiles. This is the first report to employ an unbiased, machine learning approach in CS and identified 3 phenotypes of patients with CS due to MI. These findings may inform treatment strategies and future CS clinical trials.