One Year Heart Failure Mortality Prediction Using SHFM versus PRISM Score

Abstract

Introduction With congestive heart failure (CHF) causing significant morbidity and mortality, a model predicting 1-year mortality can help with efficient targeting of palliative care consults. Our goal was to compare PRISM score, a model developed to predict 30-day mortality in general inpatients, to Seattle Heart Failure Model (SHFM), a CHF disease specific mortality prediction tool in predicting 1-year mortality in CHF patients. A CART model that combines PRISM and SHFM was developed to identify the appropriate set of patients who would benefit from a palliative care consult. Methods A community hospital based retrospective study used SHFM and PRISM to predict 1-year mortality in 689 patients with CHF admitted from 2012-2014. The predicted scores from the SHFM and PRISM were validated using area under the receiver operation characteristic curve (AUC). The net reclassification index (NRI), integrated discrimination index (IDI) and the likelihood ratio test (LRT) were used to determine if the addition of the PRISM score increased predictive accuracy of SFHM. Classification and Regression Tree (CART) model was used to create a decision tree to assess risk of 1-year mortality in patients with CHF. Cross-validation of the CART model was carried out by fitting the model to 2/3 of the data selected at random and assessing how well it fit to the remaining 1/3 of the data. Results The discriminatory ability of PRISM (categorical) (AUC = 0.701) was not significantly different (DeLong's test p = 0.56) than that of SHFM model (AUC = 0.686). Discriminatory ability was improved with combination of SHFM with either PRISM (categorical) model (AUC = 0.740) (p = 0.002) or PRISM (continuous) (AUC = 0.749) (p = 0.002). CART model combining PRISM as the main trunk with SHFM score > = 0.23 as the next branch increased 1-year mortality risk from 28% to 57%. Conclusion This study demonstrates that PRISM score performs similarly to SHFM as a risk prediction tool for one year mortality in hospitalized heart failure patients. This suggests that a disease specific stratification tool may not be necessary to accurately predict one year mortality in this patient population. Moreover, the addition of PRISM score to SFHM increases the accuracy of predicting 1-year mortality as compared to either of the two models alone.