@misc{oai:ir.soken.ac.jp:00004076,
 author = {庄子, 聡 and ショウジ, サトシ and SHOJI, Satoshi},
 month = {2016-02-26},
 note = {Clinical drug development for a new drug proceeds in a stepwise manner. As the first step,
phase I clinical trials are conducted to investigate pharmacokinetics and safety of the drug
in humans, usually in healthy volunteers. After the phase I clinical trials, early efficacy in
a selected patient population is confirmed in a phase IIa clinical trial. If the results are
satisfactory, this is followed by a larger phase II or III clinical trial to learn and confirm
how to use the drug in the target patient population.
In phase I clinical trials, pharmacokinetics (PK) following the drug administration to
humans, typically to healthy volunteers, is thoroughly investigated by each individual
based on dense PK data (e.g., tens of blood samples sequentially collected from each
individual). In phase II and III clinical trials, PK data are collected to comprehend PK in
the target population. Main objectives of the PK study in phase II and III clinical trials are
generally to understand variability of PK in the target population and to identify factors
causing the variability rather than to capture individual PK profile in each patient.
Population PK analysis captures population mean PK profile, individual variability from
the mean PK profile (between-individual variability), and residual variability within the
individual (within-individual variability), which fits in objectives of PK study in phase II
and III clinical trials. Therefore, the population PK approach is often used for PK analysis
in phase II and III clinical trials.
However, because of ethical and medical concerns, relatively few samples per patient,
that is sparse PK data, are obtained from patients participating in the phase II and III
clinical trials and it is sometimes difficult to estimate all the PK parameters with the
sparse data alone. One effective approach to address this issue is to borrow strength from
dense PK data typically obtained from healthy volunteers (HVs) in the past phase I
clinical trials under the assumption of exchangeability between these two datasets.
Conventional approaches include fixing some PK parameter values to those obtained
from the dense PK data or fitting a PK model to a combined dataset of these two.
The Bayesian approach using prior distributions elicited from dense PK data from
HVs provides a general framework for the population PK analysis with this type of
information sharing. However, the Bayesian approach can still suffer from estimation bias
when the exchangeability assumption is violated or when the prior distribution is
misspecified. In order for improving estimation accuracy with reduction of estimation
bias under possible violations of the exchangeability assumption, we propose to introduce
an information weighting that allows discounting the information from dense PK data in
specifying the prior distributions.
We firstly present an application of the conventional approach, combining sparse PK
data with dense PK data, to real PK data of a new drug under development, pregabalin, in
order to demonstrate usefulness of the approach and extract the issues. Data used for this
population PK analysis were obtained from 14 clinical studies. The analysis provided an
accurate and reliable estimation of the relationship between pregabalin clearance and
renal function (creatinine clearance) in patients. Although model validation suggested
adequacy of the model to interpret PK characteristics in patients with sparse PK data, it
should be cautious of potential bias due to violation of exchangeability assumption,
specifically parameters which sparse data have less information to estimate.
Following application of the conventional approach, we present a proposed approach,
introduction of an information weighting on informative prior distributions for population
pharmacokinetic studies with sparse data. In our simulation setting, the optimal value for
the information weighting ranged from 0.2 to 1.0 when the exchangeability assumption
was held. On the other hand, when the exchangeability assumption was violated,
information weighting with 0.2 to 0.5 was optimal. In the real data analysis, we obtained
reasonable estimates. We chose the estimation results when information weighting of 0.5
for a plausible violation of the exchangeability assumption, as our simulation exercises
suggested that information weighting with 0.2 to 0.5 was optimal when the
exchangeability assumption was violated.
In this thesis, we present usefulness of the conventional approach for evaluating PK
parameters of interest in patients with sparse PK data, where the approach needs caution
of potential bias due to violation of exchangeability assumption. For the proposed
approach, through investigating the impact of various information weights on estimation
accuracy (bias and precision of parameter estimation), we identify a range of optimal
value in information weighting to obtain accurate estimates of PK parameters. In future
studies, it is worthwhile to identify optimal combinations of distinct weights for
respective PK parameters under more complex PK models (e.g., multiple-compartment
models and models with non-linear elimination) under extensive situations where the
exchangeability assumption is violated., 総研大甲第1596号},
 title = {Population Pharmacokinetic Analysis of Sparse Data; Use of Dense Data from Earlier Clinical Trials as Prior Information},
 year = {}
}