Personalizing hemodialysis (HD) treatment in pediatric patients with end-stage renal disease (ESRD) – application and integration of quantitative pharmacology with machine learning
Verena Gotta (1), Georgi Tancev (2), Andrew Atkinson(1,3), Julia Vogt(4), Olivera Marsenic(5), and Marc Pfister (1)
(1) University Children’s Hospital Basel, University of Basel, Basel, Switzerland, (2) University of Basel, Basel, Switzerland (3) University Hospital Bern, Bern, Switzerland (4) ETH Zürich, Zürich, Switzerland, (5) Stanford University School of Medicine, Lucile Packard Children's Hospital, Stanford, CA
Objectives: End-stage renal disease (ESRD) is associated with high morbidity and mortality. A large proportion of patients receive intermittent hemodialysis (HD) while waiting for a suitable kidney transplant. Intra-dialytic small molecule diffusive clearance, assessed by urea kinetics, is currently used to quantify HD dose. The optimal individual HD strategy and intensity (HD “dose” prescription) in pediatrics is, however, not well understood. Optimal fluid removal (ultra-filtration rate, UFR) is also not known. We aimed to develop scientific evidence for personalized HD treatment in children.
Methods: A retrospective cohort analysis was performed including 1852 patients <30 years on chronic HD since childhood (≤18 years), having received outpatient HD 3x/week 2004-2016 (53903 HD sessions with information on urea removal). (I) Pharmacometric (PMX) modeling was applied to identify age-dependent factors that influence urea dialyzer clearance (KD) based on paired pre- and post-HD urea samples. An adult two-compartment PMX model was scaled to paediatric patients based on physiologic knowledge, incorporating a mechanistic KD prediction term. (II) Traditional and novel urea-clearance based HD dose metrics, i.e. volume (V) and body-surface-area (BSA) normalized spKt/V and Kt/BSA, respectively, and prescribed UFR were evaluated as predictors for survival on chronic HD in a parametric time-to-event model (Weibull regression). (III) A machine learning (ML) analysis (random forest) was applied to further explore the association of HD dose and routine laboratory measurements (105 variables in total) with 5-year survival.
Results: A total of 1691 (923+768 for model development+evaluation), 1493 and 363 patients could be included in respective analyses[2,3,4]. (I) High dialysate-to-blood-flow ratio (QD/QB) ratio was the most important covariate in the urea kinetic PMX model, largely increasing KD and reducing prediction error of post-HD concentrations in children 1-6 years (from 58% [95% CI: 35-81%] to 6% [-9-22%]). High UFR further increased total clearance, well described by an additional convective clearance term. (II) The Weibull unadjusted model predicted 5-, 10- and 15-year survival of 91%, 81% and 71%, matching non-parametric Kaplan-Meier estimates well. Kt/BSA was a better predictor of survival than spKt/V (P<0.001 vs P=0.002, log-linear relationship). Optimal Kt/BSA was suggested at 30-34 L/m2, corresponding to spKt/V >1.4 in adolescent boys, and >1.6 in adolescent girls and children <12 years. UFR was also associated with survival (P<0.001, quadratic relationship) showing markedly increased mortality <10 and >18 mL/kg/h. (III) Low albumin was the most important predictor of 5-year survival in the ML model, out of 12 variables retained. Those included variables related to nutrition, inflammation, anemia, spKt/V (increased mortality <1.4) and UFR (increased mortality <12 mL/kg/h) (total ML model accuracy: 80%).
Conclusions: Children on chronic HD are prescribed more intense HD compared to adults with higher QD/QB ratios due to technical restrictions and higher UFR likely due to high nutritional needs. This significantly increases urea dialyzer clearance, and needs to be taken into account when scaling any small molecule HD-clearance from adult to pediatric HD patients. Both the parametric Weibull and exploratory ML survival analyses support the need for intense HD prescription in children for best long-term survival. In contrast to adults, high UFR>12 mL/kg/h was not associated with greater mortality, but mortality increased with spKt/V<1.4 (adult spKt/V target: minimum 1.2, target: 1.4). The Weibull model indicated age-dependent dose-effect associations when considering V-normalized HD dose (spKt/V, disappearing when considering BSA-normalized Kt/BSA), and a U-shaped relationship with UFR. Both PMX and ML approaches can help to personalize HD treatment in children in whom randomized prospective studies are limited by small patient numbers.
 Gotta V, et al. Nephrol Dial Transplant. 2018 Sep 1;33(9):1649-1660
 Gotta V, et al. ASAIO J. 2020 Jan;66(1):115-123
 Gotta V, et al. Pediatr Nephrol. 2021 (online ahead of print)
 Gotta V and Tancev G, et al. Nephrol Dial Transplant. 2021 Feb 20;36(3):519-528