PREDICTIVE MODELING OF FLUID VOLUMES TO SUPPORT FLUID MANAGEMENT IN HEMODIALYSIS PATIENTS: A RETROSPECTIVE ANALYSIS OF THE MONITORING DIALYSIS OUTCOMES (MONDO) DATABASE

Optimized fluid management is key to improving outcomes in dialysis patients, as extracellular water (ECW) overload is associated with adverse outcomes. Comorbidities and subclinical chronic inflammation, both highly prevalent in this population, can result in loss of muscle mass, reflected by a decrease in intracellular water (ICW). Estimates of total body water (TBW) are essential for quantifying and scaling the dialysis dose and thus for assessing dialysis adequacy in individual patients. The Body Composition Monitor (BCM, Fresenius Medical Care) is a validated, non-invasive tool that estimates fluid volume compartments as TBW, ECW and ICW. This study aimed to develop predictive models enabling instantaneous estimation of fluid volumes from demographic, historical biological data and treatment parameters, calibrated to BCM measurements. These models are intended to support fluid management in clinics without access to the BCM device or in situations where BCM measurements are not performed at every treatment. 

Adult patients from the Monitoring Dialysis Outcomes (MONDO) 2012 dataset were included. The variables consisted of patient characteristics, lab values, treatment parameters, and BCM measurements. Features were transformed into average values over the period beginning one day before the current estimation date and extending to the last BCM measurement, with a maximum look-back period of 90 days. The final dataset included 18,600 patients with 162,479 treatments. Data were split by patients into training (n=16740, 90%) and testing (n=1,860, 10%) subsets, stratified by region. eXtreme Gradient Boosting (XGBoost) machine learning algorithm was used to develop models. Two models were built for each of fluid volume metrics: one including prior BCM assessments (w/ BCM) and one without (w/o BCM). Model performance was evaluated using R-squared (R2), mean squared error (MSE), mean absolute error (MAE) and mean absolute percentage error (MAPE).

In the testing subset, 60% were male, 86% from Europe, 67% were aged> 60 years.  The w/BCM models achieved R2 values greater than 0.85 for all metrics on the testing subset. The w/o BCM models had lower R2 values, ranging from 0.73 to 0.81. Although accuracy is lower without BCM data, the w/o BCM models still provided acceptable predictive performance (Table 1).

For the w/BCM models, the top predictor of each fluid volume metric was the difference between two recent corresponding BCM-measured values. Among the top five features in the w/o BCM models, the shared predictors were average urea distribution volume calculated during urea kinetic modeling, age, and height (Figure 1). 

We developed predictive models for estimating TBW, ECW and ICW, both with and without BCM measurements. Although model accuracy decreased in the absence of BCM data, predictive performance remained clinically acceptable. These models may support day-to-day decisions on fluid management and body composition assessment in dialysis patients, either in clinics without BCM access or to enable instantaneous estimations between routine BCM assessments in centers where the device is only intermittently available.