Long-term mortality risk stratification in heart failure using classification and regression trees with a combination of biomarkers

Abstract

Background: Mortality remains high in heart failure (HF). Several statistical models can be used in order to evaluate the additive usefulness of the combination of biomarkers reflecting different pathophysiological pathways. Objectives: To assess the performance of SPSS Classification and Regression trees (CART) for risk of death stratification using serum biomarkers. Patients and methods: We analyzed 876 consecutive outpatients (72% men, median age 70.3 years, median LVEF 34%). A combination of biomarkers reflecting myocyte injury (hs-cTnT), myocardial stretch (NT-proBNP) and ventricular fibrosis and remodelling (ST2) was used. Results: During a median follow-up of 4 years, 370 patients died. Using semi-automatic CART (only selecting the minimum cases for parental (80) and filial (40) nodes), 12 nodes were obtained. At first step, hs-cTnT (cutoff point 16 ng/L) yielded 2 nodes with mortality rates of 18.4% (node 1) and 57.1% (node 2). At second step, ST2 (cut points 45 ng/L for low hs-cTnT and 91.6 ng/L for high hs-cTnT) yielded nodes with mortality rates of 13.9% (node 3), 35.7% (node 4), 54.2% (node 5), and 87.5% (node 6). At third step, hs-cTnT emerged again significant to further split node 3, and node 5 was divided by NTproBNP levels (cutoff point 1846.5 ng/L). The last step used again ST2 to split node 9. The mortality in the terminal branch nodes ranged from 3.3% (node 7) to 87.2% (node 6). Agreement between predicted and observed death was 70%. Conclusions: Using a simple CART decision tree with the biomarkers hs-cTnT, ST2 and NTproBNP good stratification of risk of death was easily achieved in chronic HF outpatients. (Figure presented).