Approaches to handling missing or “problematic” pharmacology data: Pharmacokinetics

Abstract

Missing or erroneous information is a common problem in the analysis of pharmacokinetic (PK) data. This may present as missing or inaccurate dose level or dose time, drug concentrations below the analytical limit of quantification, missing sample times, or missing or incorrect covariate information. Several methods to handle problematic data have been evaluated, though no single, broad set of recommendations for commonly occurring errors has been published. In this tutorial, we review the existing literature and present the results of our simulation studies that evaluated common methods to handle known data errors to bridge the remaining gaps and expand upon the existing knowledge. This tutorial is intended for any scientist analyzing a PK dataset with missing or apparently erroneous data. The approaches described herein may also be useful for the analysis of nonclinical PK data. Overview Data from clinical trials is frequently incomplete, particularly datasets collected during large, late phase trials, during routine clinical patient care or follow-up visits. Portions of data may be missing or inaccurate due to factors such as study site noncompliance, patient noncompliance, inappropriate sample handling, data entry errors, and analytical problems. How "problematic" data are handled can impact its interpretation, especially when data used for population pharmacokinetic (PPK) modeling contains missing or erroneous data. Prior to beginning an analysis, pharmacometricians often spend a large portion of time dealing with problematic data. During data cleaning (data quality assurance), the first step is to identify missing or problematic data. Concentration-time data and dosing records are often the primary concern, but other issues, such as missing or questionable covariate data, must also be considered. Once issues/discrepancies are identified, the next challenge is to evaluate frequency of occurrence of each type of problem and the associated reason to establish appropriate methods for handling these erroneous data. Prior studies have addressed handling of specific types of problematic data, though no set of broad recommendations spanning the various types of problematic data have been previously presented. Accepted Article This article is protected by copyright. All rights reserved Through review of published methods, simulation of data sets with known errors, and evaluation using different methods for handling these errors, this tutorial aims to provide guidance for dealing with problematic clinical (and some non-clinical) concentration vs. time, dosing, and covariate data. This tutorial is intended to be utilized by scientists analyzing pharmacokinetic data with either missing data or where apparently questionable or erroneous data is present. Although data quality assurance (QA) and control (QC) are essential to successful modeling, this tutorial assumes the dataset has already undergone appropriate QC or was assembled from locked, clean data. Basic assessments include exploratory data analysis by plotting and...

The PHARMACOM‐EPI Framework for Integrating Pharmacometric Modelling Into Pharmacoepidemiological Research Using Real‐World Data: Application to Assess Death Associated With Valproate

“…Other methods for handling missing data, such as imputation of the reference value, will also impact predictions from pharmacometric models. Single imputation is easy to use and performs well when a relatively small fraction of individuals have missing covariate data 33 . However, the population from which values are imputed, either part of real‐world data or value from other studies, should be similar to the target population to which the pharmacometric model is applied 69 .…”

“…To simplify selection of a suitable approach for handling missing data, the first step is to build a missing data matrix to visualize the extent of missing data during the observational window for the study variables ( Table 4 ). Description of common methods, complete case analysis, single value imputation, and multiple imputation to handle missing data in pharmacometrics as well as clinical epidemiology can be found in recent reviews, 33,34 whereas also the importance of sensitivity analysis regarding dealing with missing data is highlighted 34 …”

Overestimation of the effect of (fos)aprepitant on intravenous dexamethasone pharmacokinetics requires adaptation of the guidelines for children with chemotherapy-induced nausea and vomiting

“…Plasma concentrations of dexamethasone and aprepitant were measured using a validated liquid chromatography mass spectrometry method, with a lower limit of quantification (LLOQ) of 1 µg/L and 0.1 µg/L respectively, as described previously [20]. The first post-dose samples below LLOQ were included as a plasma concentration of 0.5 µg/L (1/2 LLOQ for dexamethasone); other samples below LLOQ were omitted [21].…”

Visually guided preprocessing of bioanalytical laboratory data using an interactive R notebook (pguIMP)

“…However, the often nonlinear and stochastic natures of these data pose a challenge for data processing, which is why data science methods have recently made their way into systems pharmacology. 1 Although differential equation systems–based pharmacokinetic and pharmacodynamic models describing the temporal evolution of a system, such as plasma concentrations or drug effects, are well established in the preprocessing of data sets from drug research and development, 2 , 3 , 4 , 5 , 6 these additional data pose new challenges to the preprocessing of drug discovery and development data sets. This is where the strength of data science comes into play, being able to extract knowledge from high‐dimensional data, often by using machine‐learning methods (for an overview, see Badillo et al 7 ).…”

“…The analysis of biomedical data by machine learning requires data that have been cleaned of analytical laboratory errors 8 , 9 and are adequately transformed and preferably free of missing values, anomalies, 10 or values below the limit of quantification (LOQ). 2 , 5 Although likelihood‐based models have been shown to be particularly suitable for handling values below LOQ in pharmacokinetics mixed‐effects models, 2 , 3 , 4 , 5 , 6 many proposed solutions to this problem in the area of pharmacological data science are data set specific 10 , 11 and must be tailored to analyses that use machine‐learning algorithms. For example, for the treatment of missing values in gas chromatography–mass spectrometry metabolomics, predictive k‐nearest neighbors (kNN), and random forest have proven to be particularly suitable.…”