Use of item-based non-linear mixed effects model to improve confidence in Phase III clinical trial decision-making
Carolina Llanos-Paez (1), Claire Ambery (2), Shuying Yang (2), Misba Beerahee (2), Elodie L. Plan (1), Mats O. Karlsson (1)
(1) Department of Pharmacy, Uppsala University, Uppsala, Sweden; (2) Clinical Pharmacology Modelling and Simulation, GlaxoSmithKline plc, London, UK
Introduction: The proportion of failed trials due to lack of efficacy is strikingly greater in Phase III clinical drug development than in Phase II (55% vs. 48%) . This failure could be attributed not only to insufficient drug efficacy, but also to underpowered trials (as a consequence of a sample size calculated from a commonly overestimated true treatment effect obtained in Phase II)  or insufficiently powerful analysis methods to detect efficacy given that sample size. A non-linear mixed effect (NLME) model-based drug development decision-making framework is being proposed to increase Phase III trials probability of success through more precise estimates of the drug effect . This type of analysis has also shown to increase confidence around the same clinical endpoint when compared with a standard analysis in a Phase II trial .
Objectives: This study aims to illustrate how a new methodology to analyse patient reported outcome (PRO) data based on a longitudinal item response theory-based model (IRM) improves confidence in a Phase III clinical trial endpoint compared with a mixed model repeated measures (MMRM) analysis.
Methods: PRO data from the FULFIL trial , a Phase III trial that compared 24 weeks of once daily inhaled triple therapy (fluticasone furoate/umeclidinium/vilanterol – FF/UMEC/VI) with twice daily inhaled dual therapy (budesonide/formoterol – BUD/FOR) in patients with chronic obstructive pulmonary disease (COPD), was analysed using an IRM. PRO data were obtained from the Evaluating Respiratory Symptoms in Chronic Obstructive Pulmonary Disease (E-RS:COPD) tool , consisting of 11 items to assess respiratory symptoms . The clinical endpoint in this analysis was the change from baseline (CFB) in E-RS:COPD total score (RS-total), which is computed by taking the sum of the 11 items, over 4-week intervals for each treatment arm (FF/UMEC/VI and BUD/FOR) [5,7]. The IRM was developed, using NONMEM 7.4.4, in two steps: i) item characteristic functions (ICFs) were characterized ii) a longitudinal model was developed to describe symptoms-time course in COPD patients receiving either FF/UMEC/VI or BUD/FOR. For the longitudinal model, the empirical Bayes estimates (EBEs) of the latent variable were considered as the dependent variable taking into account their uncertainty . The point estimate and precision in the estimated clinical endpoint, CFB during the last 4-week period on treatment, were obtained through model simulations (Nsim=2,000 in 15,000 virtual subjects per treatment arm), with inclusion of the uncertainty in the estimated longitudinal IRM parameters using the $PRIOR functionality in NONMEM. The precision around the clinical endpoint obtained with IRM was compared with those of the published MMRM analysis [5,7].
Results: Data from 1801 patients (mean (SD) age of 63.9 years (8.65), 43.8 % smokers at study initiation) who received up to 24 weeks of either FF/UMEC/VI (n=907) or BUD/FOR (n=894) were included in this analysis. ICF parameters were estimated with good precision (relative standard error < 10%), and a Weibull function combined with an offset function best described the COPD symptoms-time course in both trial arms. The IRM considerably improved precision of the drug effect estimate compared to the MMRM at every time point. For example, at the end of treatment (week 21 – 24) the mean (95% CI) CFB in RS-total was -2.47 (-2.61, -2.30) with IRM compared to -2.31 (-2.62, -2.00) with MMRM in the FF/UMEC/VI arm, whereas it was -0.97 (-1.10, -0.81) with IRM compared to -0.96 (-1.27, -0.64) with MMRM in the BUD/FOR arm. Furthermore, a relative sample size of 4.00 (FF/UMEC/VI) and 4.72 (BUD/FOR) times larger would be required in the MMRM analysis to achieve the precision obtained with the IRM analysis at the end of the study (week 21 – 24).
Conclusion: This analysis shows the advantage of using a NLME model-based analysis of Phase III clinical item-response level data, over a common current approach used in drug development (MMRM) for the same clinical endpoint (CFB at end of treatment in RS-total) in both methods (IRM and MMRM). The IRM approach increased the precision in the efficacy estimate, resulting in higher probability of making the correct decision compared to MMRM analysis, which appears important in light of the many failures in Phase III attributed to underpowered trials.
 Harrison RK Nat Rev Drug Discov 2016; 15(12):817-8
 Erdmann S et al. BMC Med Res Methodol. 2020; 20:253
 Milligan PA et al. Clin Pharmacol Ther. 2013; 93(6):502–14.
 Llanos-Paez et al. ACoP 11. 2020, Abstract TUE-012
 Lipson DA et al. Am J Respir Crit Care Med. 2017; 196(4):438–46.
 Evidera. EXACT Program https://www.exactproinitiative.com/content/
 Tabberer et al. Adv Ther. 2018; 35(1):56–71.
 Schindler E et al. Pharm Res. 2018; 35(6):122.
Funding Information: GSK funded this research in the form of a Research payment to Uppsala University.
Conflict of Interest: CL-P, ELP and MOK declare that they have no conflict of interest. CA, SY, and MB are GSK employees and hold GSK shares.