Effect of Serum Albumin Changes on Mortality in Patients with Peritoneal Dialysis: A Joint Modeling Approach and Personalized Dynamic Risk Predictions

Abstract

Peritoneal dialysis (PD) is a frequently used and growing therapy for end-stage renal diseases (ESRD). Survival analysis of PD patients is an ongoing research topic in the field of nephrology. Several biochemical parameters (e.g., serum albumin, creatinine, and blood urea nitrogen) are measured repeatedly in the follow-up period; however, baseline or averaged values are primarily associated with mortality. Although this strategy is not incorrect, it leads to information loss, resulting in erroneous conclusions and biased estimates. This retrospective study used the trajectory of common renal function indexes in PD patients and mainly investigated the association between serum albumin change and mortality. Furthermore, we considered patient-specific variability in serum albumin change and obtained personalized dynamic risk predictions for selected patients at different follow-up thresholds to investigate the effect of serum albumin trajectories on patient-specific mortality. We included 417 patients from the Erciyes University Nephrology Department whose data were retrospectively collected using medical records. A joint modeling approach for longitudinal and survival data was used to investigate the relationship between serum albumin trajectory and mortality of PD patients. Results showed that averaged serum albumin levels were not associated with mortality. However, serum albumin change was significantly and inversely associated with mortality (HR: 2.43, 95% CI: 1.48 to 4.16). Risk of death was positively associated with peritonitis rate, hemodialysis history, and the total number of comorbid and renal diseases with hazard ratios 1.74, 3.21, and 1.41. There was also significant variability between patients. The personalized risk predictions showed that overall survival estimates were not representative for all patients. Using the patient-specific trajectories provided better survival predictions within the follow-up period as more data become available in serum albumin levels. In conclusion, using the trajectory of risk predictors via an appropriate statistical method provided better predictive accuracy and prevented biased findings. We also showed that personalized risk predictions were much informative than overall estimations in the presence of significant patient variability. Furthermore, personalized estimations may play an essential role in monitoring and managing patients during the follow-up period.

The joint modeling approach with a simulation study for evaluating the association between the trajectory of serum albumin levels and mortality in peritoneal dialysis patients

“…We determined the explanatory variables associated with the longitudinal trajectory of serum albumin level using a univariate mixed-effect model. Finally, we adjusted the serum albumin trajectories for significant confounders, i.e., baseline age (p<0.001), the transport property of peritoneal membrane (TPPM) (p<0.001), and peritonitis rate (PR) (p=0.012) (Table (2).…”

“…In the survival sub-model, the explanatory variables were specified using a univariate Cox proportional hazard regression model. The hazard of death was adjusted for age onset (p<0.001), history of PD (p=0.031), number of illness/comorbid diseases (NI) (p<0.001), and peritonitis rate (PR) (p<0.001) (Table (2). Finally, the fitted trajectory of serum albumin was included in the survival sub-model considering different parameterization as Equation (3.2).…”

“…The predictive ability of built models was evaluated using the time-dependent ROC area under the curve (AUC) and the Brier score (BS). The current study is, to the authors best knowledge, the first study that examined the joint modeling approach on PD patients in Turkey, and among few studies published worldwide [2,9,13,31].…”

Joint modeling of survival and longitudinal data: Carrico index data example

“…The vast majority of studies ignored repeated measurements and used either the Kaplan-Meier or the Cox proportional hazard regression model based on a single measurement (i.e. baseline or average of multiple records) of related risk factors [2,3,4,5].…”

“…It is now possible to predict the hazard of treatment of AHRF patients individually using patient-level data and changes in biological markers over time, thanks to recent advances in statistical modeling. In addition, patient-specific risk predictions for future time points can be updated dynamically as new information becomes available known [5]. It is possible to predict the results of AHRF patient survival using the Carrico Index levels at the start of the study or averaged values while the follow-up period.…”