Use of Risk Models to Predict Death in the Next Year Among Individual Ambulatory Patients With Heart Failure.

Importance: The clinical practice guidelines for heart failure recommend the use of validated risk models to estimate prognosis. Understanding how well models identify individuals who will die in the next year informs decision making for advanced treatments and hospice.

Objective: To quantify how risk models calculated in routine practice estimate more than 50% 1-year mortality among ambulatory patients with heart failure who die in the subsequent year.

Design, Setting, and Participants: Ambulatory adults with heart failure from 3 integrated health systems were enrolled between 2005 and 2008. The probability of death was estimated using the Seattle Heart Failure Model (SHFM) and the Meta-Analysis Global Group in Chronic Heart Failure (MAGGIC) risk calculator. Baseline covariates were collected from electronic health records. Missing covariates were imputed. Estimated mortality was compared with actual mortality at both population and individual levels.

Main Outcomes and Measures: One-year mortality.

Results: Among 10 930 patients with heart failure, the median age was 77 years, and 48.0% of these patients were female. In the year after study enrollment, 1661 patients died (15.9% by life-table analysis). At the population level, 1-year predicted mortality among the cohort was 9.7% for the SHFM (C statistic of 0.66) and 17.5% for the MAGGIC risk calculator (C statistic of 0.69). At the individual level, the SHFM predicted a more than 50% probability of dying in the next year for 8 of the 1661 patients who died (sensitivity for 1-year death was 0.5%) and for 5 patients who lived at least a year (positive predictive value, 61.5%). The MAGGIC risk calculator predicted a more than 50% probability of dying in the next year for 52 of the 1661 patients who died (sensitivity, 3.1%) and for 63 patients who lived at least a year (positive predictive value, 45.2%). Conversely, the SHFM estimated that 8496 patients (77.8%) had a less than 15% probability of dying at 1 year, yet this lower-risk end of the score range captured nearly two-thirds of deaths (n = 997); similarly, the MAGGIC risk calculator estimated a probability of dying of less than 25% for the majority of patients who died at 1 year (n = 914).

Conclusions and Relevance: Although heart failure risk models perform reasonably well at the population level, they do not reliably predict which individual patients will die in the next year.