The drug development process and FDA-approved prescribing generally assume that patients are
sufficiently stable and similar enough to justify population-based dosing for a given group that is usually
unchanged during therapy. Unfortunately, there is a huge body of evidence that dosing according to this
“one size fits all” paradigm results in wide variation in plasma drug concentrations between individuals
and even within the same individual over time, all of which can compromise clinical outcomes. Population
pharmacokinetic (PK) and pharmacodynamic (PD) models can control for this variability by providing
clinicians with tools to adjust doses accordingly, a process that has come to be known as Model-Informed
Precision Dosing (MIPD). However, MIPD has been better able to control for inter-individual variation
rather than interoccasion variation (IOV) within an individual over time. MIPD methods exist to track IOV
in the past, but not to account for possible future IOV. In this project we will address IOV in three novel
approaches. Our first aim uses our unique Virtual Pediatric Intensive Care Unit (VPICU) dataset with >400
clinical variables obtained from ~20,000 unstable, critically ill children in our hospital since 2009. We will
build recurrent neural networks (RNNs) to predict changes in renal function within individuals, which is
relevant to the control of renally excreted drugs. While models exist to predict renal failure, this will be
the first application of RNNs to predict creatinine clearance in children. There are >100,000 serum
creatinine measurements to validate this work. Our second aim is to account for changing PK-PD in
models that cannot be linked to a specific covariate like renal function. To do this, will incorporate
stochastic differential equations (SDEs) to capture changes in model parameters over time. Unique to our
work, we will apply SDEs in the setting of our long history of non-parametric PK-PD modeling, which
makes no assumptions about underlying probability distributions for parameter values in a model and is
particularly good at describing and controlling unusual patients, perfect for a critically ill population. We
will use >40,000 vancomycin doses and >5,000 plasma concentrations in VPICU to test our algorithms. Our
third aim is two-fold. First, we will again use RNNs to predict outcomes of VPICU patients with
Staphylococcal bloodstream infections treated with vancomycin. We will compare RNNs that include
vancomycin exposure estimated with IOV and without IOV. The second part is to use our in vitro hollow
fiber infection model (HFIM) to directly assess the effect of vancomycin IOV on both methicillin-resistant
and methicillin-susceptible Staphylococcus aureus in our laboratory. The HFIM can reproduce pediatric
PK to measure antibacterial kill and emergence of less susceptible or persister organisms over days to
weeks. Our inclusion of IOV in the HFIM is completely novel. We will deliver software tools to clinicians
to control IOV and understand the magnitude relevant to outcomes of anti-Staphylococcal therapy.