How is model building reported for population PK-PD? An exhaustive survey of the literature between 2002 and 2004
Dartois, C (1)(4), K Brendel (2)(4), E Comets (2), C Laveille (3), C Laffont (4), B Tranchand (1)(5), F Mentré (2), A Lemenuel-Diot (4), and P Girard (1)(6).
(1) EA3738, Faculté de Médecine de Lyon Sud, Oullins, France; (2) INSERM U738, Hôpital Bichat, Paris, France; (3) Exprimo NV, Lumnen, Belgium ; (4) Institut de Recherches Internationales Servier, Courbevoie, France; (5) Centre Léon Bérard, Lyon, France; (6) INSERM, Lyon, France.
Purpose: An exhaustive description of data, model, and methodology ensuring confidence in clinical and statistical interpretations, is an important issue in population PK and/or PD analyses. Objective of the present study was thus to perform a systematic review of all population analyses published between 2002 and 2004 in order to have an overview of the level of details provided in model description as well as software used, algorithms and the way that model was built. Model evaluation is being assessed in a second study.
Methods: Firstly, we selected articles in Pubmed (n=324) using defined keywords. Secondly, we built and validated a data abstraction form (DAF) constituted by 5 sections, describing article, authors and drug treatment generalities, clinical study from which the data arose, and focusing on modeling process and qualification. After DAF qualification, two persons read the papers and directly collected the information in a mysql database. Descriptive statistical analysis, including agreement test between the readers as well as trends over time tests for certain items of the DAF, were carried out using SAS software (version 8).
Results: We pointed out that three therapeutic classes (oncology, microbiology and analgesia) with low therapeutic index and/or high PK/PD variability were the most modeled and published (58 % of the cases). Data used in models (for a majority descriptive) were often homogenous (from one clinical trial, one arm, including one compound), and rich rather than sparse. NONMEM was the main software used (69.5 % of the cases), followed by NPEM and ADAPT. PK models were rarely complex (number of compartment <4) as well as PD models (majority of Emax). Covariate testing was often performed (58 % of the cases). Essentially based on likelihood ratio test, it mainly revealed high influence of demographics (48 %), elimination organs (30 %), or drug interactions (13 %). Overall, description of model building steps gave enough details to appreciate quality of the analyses in only 40 % of the cases.
Conclusion: As this methodology is complex and concerns data from patients with severe pathologies or drugs with low therapeutic index, description of model building steps should be improved. Our DAF can be a useful tool to summarize this required information and provides a quality survey of published population PK-PD over years.