Modeling Blood Urea Nitrogen (urea) Dynamics to Estimate Dialysis Adequacy

Objectives: Critically ill patients experiencing acute kidney injury (AKI) may require kidney replacement therapies for clearance of uremic toxins. Dialysis dosing and adequacy, typically expressed as clearance, are not routinely documented in the intensive care unit (ICU). Blood urea nitrogen (BUN) concentration trajectories can be used to estimate clearance, the quantity of blood “cleared” of toxins every minute. However, sparse clinical BUN data necessitate a model that can predict BUN concentration during and between intermittent hemodialysis (IHD) sessions. Previous studies using routinely collected clinical data have focused on fitting models for isolated dialysis treatments and included physiologically inaccurate assumptions regarding patient volume status, which limits the ability to estimate clearance in a dynamic and personalized fashion [ 1, 2]. Additionally, most existing work models cellular BUN dynamics [ 1– 3]. A more clinically measurable approach is to model BUN dynamics between vascular and extravascular spaces, facilitating the incorporation of interpatient variability across treatments and between patients. Thus, we derive clearance estimates using a compartmental solute and volume-based mathematical model for individual patients in a large ICU patient dataset using only routinely collected data. Methods: A vascular, 2-compartment ordinary differential equation (ODE) model was developed to predict both patient volume and BUN concentration changes between and during IHD treatment(s). Data was extracted from the electronic health record of patients with AKI undergoing IHD in the University of Pittsburgh Medical Center ICU. Parameters for BUN generation and dialyzer clearance were fit using a differential evolution algorithm to minimize the normalized root mean square error (RMSE) between the actual and predicted BUN trajectories within clinically plausible parameter bounds. BUN rebound time and percentage were used as additional metrics to determine whether model predictions matched clinical intuition. The model was parametrized on 12 patients with no apparent renal function undergoing 44 IHD treatments. Results: Patients had a median of 16.5 (IQR = 7.2) BUN measurements and underwent 2-8 IHD treatments. The model accurately predicted BUN concentration trajectories in IHD patients, with a median BUN concentration RMSE of 0.11 mg/dL per data point. Additionally, fitted parameter values for BUN generation rate (472 ± 215 mg/hour) and dialysis clearance rate (218 ± 112 dL/hour) were within physiologically expected ranges and reflected expected variability across patients and treatments. The median BUN rebound time was 31 minutes, and the median BUN rebound percentage was 38%, which were within clinically expected ranges [ 4] and indicated that the model predictions were physiologically reasonable. Conclusions: The 2-compartment vascular model fit BUN concentration profiles for all tested patients and captured variance in BUN clearance and generation between patients and IHD treatments. The model-generated trajectories can be used to characterize dialysis clearance and adequacy, as well as the potential for renal recovery after AKI in critically ill patients.