Clinically useful and reliable mortality scoring based on serum creatinine, ejection fraction, and age in Heart Failure: A logistic regression cost–benefit analysis

Abstract

Background: Accurate prediction of mortality in heart failure (HF) is essential for timely clinical decision-making. However, many risk models rely on multiple inputs that may be impractical in routine settings. This study evaluated whether a simplified logistic regression model using serum creatinine, ejection fraction, and age could provide predictive performance and clinical utility comparable to more complex models.

Results: Model 2, which included serum creatinine, ejection fraction, and age, demonstrated strong performance (AUC = 0.786; Brier score = 0.166) with excellent calibration (Hosmer–Lemeshow p = 0.765; Emax = 0.064) and the highest net clinical benefit across threshold probabilities ranging from 0.10 to 0.60. Compared to Model 1, it significantly improved patient classification (NRI = 0.111; 95% CI: 0.009–0.210), while the addition of serum sodium (Model 3) or other predictors (Model 4) yielded no further reclassification gains (NRI = 0.000). Model 2 stratified patients into three predicted risk groups: low (≤33%), moderate (34–66%), and high (≥67%). This score-based classifier achieved 92.3% overall accuracy with excellent diagnostic performance across categories (e.g., High-risk: sensitivity = 1.00; specificity = 0.93).

Methods: We performed a secondary analysis of a publicly available dataset of 299 HF patients from two centers. Four logistic regression models were developed: Model 1 (serum creatinine and ejection fraction), Model 2 (Model 1 plus age), Model 3 (Model 2 plus serum sodium), and Model 4 (a full model including all available predictors: age, anaemia, creatinine phosphokinase, diabetes, ejection fraction, high blood pressure, platelets, serum creatinine, serum sodium, sex, and smoking). Model performance was assessed using AUC for discrimination, Brier score, and Hosmer–Lemeshow test for calibration, and decision curve analysis (DCA) for clinical utility. Net Reclassification Improvement (NRI) was used to quantify reclassification gains. A point-based scoring system was derived from the best model using age as the reference to weight predictors and estimate individual mortality risk. Patients were grouped into low, moderate, or high-risk criteria, where a confusion matrix was used to evaluate sensitivity, specificity, and predictive values for each category.

Conclusion: A simplified logistic regression model including only serum creatinine, ejection fraction, and age provides comparable predictive accuracy and greater clinical utility than a full model using 11 predictors. These findings support its use as a practical and low-cost tool for early risk stratification in heart failure care.