Abstract 20282: Comparative Estimates of Heart Failure Hospitalization Between Simple and Multi-State Proportional Hazard Models in a Population With a High Rate of Competing Events
Background: Cox proportional hazard models using Kaplan-Meier estimates are often used in prediction models but may lead to biases in situations of competing or semi-competing risks due to informative censoring. Multistate models generate more accurate estimates than simple Cox models. However, the magnitude of these differences is understudied and is particularly relevant when the rate of the censored event (e.g., death) is substantial. We therefore compared predictions of heart failure (HF) hospitalization from a simple Cox model of HF hospitalization alone to those from a multistate model that includes the outcomes of HF hospitalization and death. We also report the additional effect of the transition event (HF hospitalization) on the risk of death.
Methods and Results: We used the Heart Failure Endpoint evaluation of Angiotensin II Antagonist Losartan (HEAAL) study cohort, a multicenter, randomized trial of 3,834 symptomatic patients with reduced left ventricular ejection fraction to compare predictions from a simple Cox model with HF hospitalization as the only outcome and a multistate model that included HF hospitalization and all-cause mortality as outcomes of interest. The two models contained the same 12 pre-specified variables: age, gender, New York Heart Association class (III vs II), left ventricular ejection fraction, serum creatinine, serum sodium, blood pressure, weight, history of diabetes, ischemic heart disease, atrial fibrillation and prior stroke. In the presence of the competing risk of death, overall estimates of HF hospitalization were different between the two models: 1-year and 5-year estimates of 8.0% and 27.1% in the simple Cox model and 7.7% and 24.3% in the multi-state model, respectively. There was a 5.9-fold higher hazard of death after a HF hospitalization (95% CI 4.4-5.7).
Conclusions: Multistate models such as the one utilized here account for transition states, such as HF hospitalization, and absorbing states, such as death, and are known to give more accurate estimates than simple Cox models. In the present analysis, we observed modest over-estimates in the prediction of HF hospitalization when using a simple Cox model. We further observed that hospitalization markedly increased the rate of subsequent death.
Author Disclosures: J.N. Upshaw: None. M.A. Konstam: None. D. van Klaveren: None. D.M. Kent: None.
- © 2016 by American Heart Association, Inc.