Comparing the Predictive Attributes of the Seattle Heart Failure Model for Death Versus Heart Failure Hospitalization

Abstract

Introduction During periods of relative disease stability, risk prediction models for heart failure (HF) such as the Seattle Heart Failure Model (SHFM) may provide incremental, objective information for assisting therapeutic decisions. Given the predominance of modifiable risk factors in the SHFM, which may only transiently characterize the HF milieu, the SHFM's predictive information may only be useful for a limited time after its assessment. Accordingly, this study sought to identify the “warranty period” of the SHFM - the length of time following assessment that the SHFM provides acceptable predictive information - for two salient HF outcomes: death and heart failure hospitalization (HFH). Secondary aims included comparing the discriminatory capacity of the SHFM and the differential predictive power of individual model elements with respect to these outcomes. Methods This retrospective cohort study incorporated data from the electronic medical records (EMR) of patients from a single, large integrated health care system. HF patients were identified through the appropriate ICD9 codes and baseline dates assigned as the first completed office visit ≥2 years following the first EMR-documented encounter. The office visit environment was chosen as a point of presumed disease stability - a context where prediction models should be most valuable. The 14 SHFM elements were determined from EMR documentation on or prior to the baseline office visit. Study endpoints were all-cause death, HFH, and death/HFH, with HFH defined as inpatient admissions with HF as the primary discharge diagnosis. Warranty periods for the SHFM were considered by evaluating its predictive performance within non-overlapping 3-month intervals following baseline in a series of landmark analyses. Modifiable and non-modifiable SHFM elements were analyzed separately. Results Among 20,733 HF patients for whom quantitative risk estimates from the SHFM were calculated at an office visit, 46%, 15%, and 52% experienced death, HFH, or death/HFH over an average follow-up period of 3.9 years. The SHFM had decreasing predictive performance over time with decrements in c-statistics from baseline to 12 months post-baseline of 0.05, 0.03, and 0.03 for death, HFH, and death/HFH, respectively. The declining predictive performance was attributable solely to modifiable SHFM elements. The SHFM predicted 1-year death better than 1-year HFH (c=0.66 vs. c=0.62). Among the 4 strongest predictors of death or HFH at 1 year, diuretic dose, hemoglobin, and % lymphocytes predicted both outcomes; however, the strongest predictor of 1-year death was age, and of 1-year HFH, ejection fraction. Conclusions The SHFM predicted death better than HFH, but the model's predictive power for both endpoints declined over time, attributed solely to the model's modifiable elements. At least annual updating of the modifiable elements is recommended for proper clinical application of the SHFM.