Construction of a mortality score in French heart failure outpatients: Model comparisons

Abstract

Introduction Cardiologists need to assess the mortality risk of their patients with heart failure (HF). The meta-analysis global group in chronic heart failure (MAGGIC) validated a precise score, but its complexity prevents cardiologists to use it in clinical settings. We aimed to derive a simple mortality score on a small cohort of HF patients of the CHU of Montpellier, using a piecewise constant Poisson model. Our secondary objectives were to compare the estimates and the validity of our model with those of other models. Methods At inclusion, we assessed age, sex, body mass index, ejection fraction, serum creatinin, NT-ProBNP, NYHA status, diabetes, chronic obstructive pulmonary disease, smoking habits, and treatment by beta-blockers. The outcome measure was the time to death. The censoring date was 4.4 years after first inclusion. The assumptions of linearity and of proportional hazards were assessed in Cox models, and potential predictive variables were defined on clinical and statistical arguments. The time intervals of the Poisson Model PM1 were chosen on clinical and statistical arguments. Variables were selected with a backward procedure, targeting the minimal AIC of PM1. We converted the Poisson model predictor to an integer score, and represented discrimination (ROC curve) and calibration (observed and predicted mortality risk) graphs. We built 4 other models, with the same variables as PM1: PM2/with 1 time interval, PM3/with 3 time intervals, defined on statistical arguments, PM4/with age as a time-dependent variable, and a Cox model. In order, to compare the global results and clinical interpretation of these models, we represented graphically the relative risks and hazard ratios of these models, with their confidence intervals. In order to compare the validity of the individual predictions, we performed a cross-validation for each Poisson model. The size of our training samples was 2/3 of the size of our complete sample. We present the mean percent of the validation samples that was correctly classified. Results We included 182 patients, and performed a complete case analysis on 134 patients. Selected variables were age >75 years, Ejection Fraction 5000, NYHA 3–4, no beta-blockers, diabetes, chronic pneumopathy, current smoker. Age did not verify the assumption of proportional hazard, but the interaction with time was not included for usability reasons. The integer risk score was built in attributing 1 point to each risk factor. The c-statistic of this score was of 0.77. Patients with scores ≤2, 2–5, and ≥6, had mortality risks ≤0.2, 0.2–0.8, and ≥0.8, respectively. The relative risks of the 4 Poisson models and the hazard ratios of the Cox model were not graphically different (figure). We performed the cross-validation on 100 training samples of 89 subjects. The mean percent of correctly classified patients in the validation samples was of 63.3, 63.4, 60.1, and 63.0% for PM1, PM2, PM3, and PM4, respectively. Discussion We propose a simple mortality risk score for chronic heart failure patients. Its discriminative ability and its calibration were moderate. The statistical choices seem to be of mild importance, as more complicated or simpler models did not increase or decrease the predictive ability. The research of a user-friendly score could be pursued in larger cohorts, keeping in mind that the clinical aim and the statistical choices should be simplified.