Identifying patients at risk of death or hospitalisation due to worsening heart failure using decision tree analysis: Evidence from the Trans-European Network-Home-Care Management System (TEN-HMS) Study

Abstract

ObjectivesTo stratify patients recently discharged from hospital with heart failure (HF) according to their risk of death and/or hospitalisation for worsening HF (WHF), to enable timely and appropriate monitoring and intervention. MethodsData from the TEN-HMS study were used in this analysis. Chi-square automatic interaction detector (CHAID) decision trees were constructed using a 10-fold cross-validation to predict events at 1-year and compared with logistic regression (LR) models using ROC curve analysis. Results284 patients were used for training and 160 patients available at 4-month for validation. Amino-terminal pro-brain natriuretic peptide (NT-proBNP) was the strongest predictor of mortality identifying groups with high (>13,492pg/ml), medium (3127–13,492pg/ml) and low (≤3127pg/ml) risk, followed by MI, systolic blood pressure, age, heart rhythm, study randomisation group and serum sodium. NT-proBNP was also the strongest predictor for death or hospitalization for WHF identifying groups with high (>13,492pg/ml), medium (584–13,492pg/ml), and low (≤584pg/ml), followed by MI, creatinine, heart rhythm, potassium and urea. CHAID trees tended to perform better than LR-models (prediction of the composite outcome: ROC area with 95% CI, 0.797 (0.745–0.849) for CHAID and 0.738 (0.680–0.796) for LR-model; p=0.041; prediction of mortality: 0.892 (0.853–0.931) for CHAID and 0.858 (0.813–0.904) for LR; p=0.15). ConclusionsDecision trees are an alternative classification method used to differentiate risk in patients with HF. The resultant models are concise, free of subjective variables and understood easily by clinicians. Further exploration of their potential and validation in other data-sets is justified.