Prediction of Mortality and Hospitalization Risk Using Nutritional Indicators and Their Changes Over Time in a Large Prevalent Hemodialysis Cohort

Objective(s) Malnutrition and protein–energy wasting are associated with morbidity and mortality in hemodialysis patients. Existing nutritional scores rely primarily on cross-sectional data. Using readily available nutritional indicators, we developed models to predict the risk of mortality and hospitalization in prevalent hemodialysis patients. Design and Methods In this retrospective study, we constructed prediction models of 1-year mortality and hospitalization using generalized linear models, generalized additive models (GAM), classification tree, and random forest models. The models were compared using area under the receiver-operating characteristics curve (AUC) and calibration curves. Model predictors included nutritional and inflammation indicators, demographics, comorbidities, and slopes of all continuous variables over 6 months. Patients were randomly split in the ratio 2:1:1 into training, testing, and validation cohorts, respectively. We included patients with hemodialysis vintage ≥1 year from Fresenius Medical Care North America clinics from July 2011 to December 2012 (N = 21,802 in mortality analysis; N = 13,892 in hospitalization analysis).The outcome variables were 1-year mortality and hospitalization. Results For mortality prediction, GAM was the best model (AUC = 0.85, 95% confidence interval=0.83-0.86), comprised of neutrophil-to-lymphocyte ratio slope, serum bicarbonate slope, and vintage as nonlinear predictors, and age, serum albumin, and creatinine as linear predictors. For hospitalization prediction, GAM was also the best model (AUC = 0.70, 95% confidence interval = 0.62-0.79) and included neutrophil-to-lymphocyte ratio slope, bicarbonate slope, volume of urea distribution, vintage, and phosphate slope as nonlinear predictors, in addition to albumin, congestive heart failure, age, phosphate, equilibrated normalized protein catabolic rate, and creatinine as linear predictors. Both models demonstrated good calibration, with mild overestimation of hospitalization risk at the highest risk interval. Conclusions The GAM model can accurately predict the risk of mortality and hospitalization. Application of these prediction models could inform allocation of nutritional interventions to patients at highest nutritional risk.