Abstract 319: Development and Validation of a Mortality Prediction Model for Hospitalized Heart Failure Patients

Abstract

Background: Clinical predication models for hospitalized patients with heart failure have good performance characteristics but require medical record review, which is both costly and time-consuming. In contrast, administrative claims data are readily available for research but have limited clinical detail. Methods: We used highly detailed billing data to derive a multivariable in-hospital mortality prediction model for patients with heart failure. We validated the model in a separate multi-hospital dataset derived from electronic health record (EHR) data. For the derivation, we included patients aged ≥18 years admitted to one of 433 hospitals that participate in the Premier, Inc. Data Warehouse (PDW) between January 2009 and June 2011. Patients with a principal ICD-9-CM diagnosis of heart failure or principal diagnosis of respiratory failure with secondary diagnosis of heart failure were included. After dividing the dataset into derivation and validation sets, we used generalized estimating equations to estimate the parameters of a generalized linear model that adjusted for clustering within hospitals. The model included selected patient demographics, comorbidities, history of prior heart failure admissions, and initial medications and therapies (e.g., inotropes, mechanical ventilation) administered in the first 2 hospital days. We applied the models to the validation set. Finally, we applied the coefficients from the model to a novel multi-hospital dataset, HealthFacts (Cerner Corp), which contains information about hospitalizations and is derived from the electronic health record of 76 participating sites (years 2010-2012). Results: In PDW, we identified 200,832 patients ≥18 years old with a diagnosis of heart failure. Of these, 80% (160,839) were randomly assigned to the derivation cohort and the remaining 20% (39,993) were assigned to the validation cohort. Mortality was 4.0%. In the derivation cohort, the model showed a c-statistic of 0.79 Factors that were most strongly associated with mortality included age over 80 years, early inotropes, early vasopressors, and diagnosis of acute kidney injury that was present on admission. In the validation cohort, the c-statistic was 0.79. We then validated the model in the HealthFacts dataset, which had a higher percentage of African American patients (28% vs. 17%). In this cohort, the c-statistic was 0.82. Conclusions: A heart failure mortality model based on detailed administrative data available in the first 2 hospital days had excellent performance characteristics in a derivation and an external, EHR-based validation dataset. When clinical data are not available, this model may be useful for severity adjustment in comparative effectiveness studies of heart failure patients.