Modeling mortality risk in hemodialysis patients using laboratory values as time-dependent covariates

Proportional hazards analyses assume that the magnitude of mortality risk for a predictor variable remains proportional over time. In a time-dependent model, the explanatory variable violates this assumption, and repeat observations are required to accommodate the change in risk that occurs over time. Using a retrospective cohort design, we tested the following laboratory values for a time-by-covariate interaction: hematocrit (HCT), serum albumin (ALB), and serum creatinine (CR). A random sample of 4,083 hemodialysis patients whose specimens were analyzed in a central laboratory over a 3-year period served as the study group. Using the baseline observation, we discovered significant probability values for the interaction terms TIME*CR (P = 0.013) and TIME*ALB (P < 0.01). The beta coefficient for TIME*HCT was not significant, indicating that the proportional hazard assumption was not violated by this covariate. Based on these results, we fitted a multivariate regression model containing two time-dependent covariates (CR and ALB) using a data structure that incorporated repeat observations of these laboratory values. Patients with high ALB levels experienced the lowest mortality risk. Similarly, serum CR levels were higher in long-term survivors. This analysis verifies the clinical importance of monitoring ALB and CR levels over time and the hazard of using a single laboratory observation to make long-term projections. Additionally, the predictive values of the time-dependent covariates ALB and CR affirm the significance of nutritional approaches directed toward maintaining visceral and somatic protein content throughout renal replacement therapy.