Theodoros Papathanasiou (1,2,4), Rune Viig Overgaard (2), Anders Strathe (2), Trine Meldgaard Lund (1), Andrew C. Hooker (3)
(1) Department of Drug Design and Pharmacology, Faculty of Health and Medical Sciences, University of Copenhagen, Copenhagen, Denmark (2) Novo Nordisk A/S, Pharmacometrics (PMX), Søborg, Denmark (3) Department of Pharmacy, Uppsala University, Uppsala, Sweden (4) Novartis Pharma AG, Early Development Analytics-Translational Medicine, Basel, Switzerland (Present affiliation)
Objectives: Model-based adaptive optimal design (MBAOD) is a promising methodological tool for designing efficient dose-finding studies. MBAODs are allowed to adapt at prespecified interim evaluation steps, driven by study characteristics that optimize an optimal design criterion [1]. Simulation studies have shown that MBAODs are less susceptible to misspecification during the design phase, as compared to classical fixed optimal designs [2-4]. However, these findings were based on the assumption that the model structure is known a priori, which is typically not the case in most clinical development programmes.Thus, we aimed to extend the MBAOD framework by introducing a Model Selection (MS) or a Model Averaging (MA) step at each interim step to account for model structure uncertainty.
Methods: We performed Clinical Trial Simulations for hypothetical dose-finding trials. The “ground truth” simulation model was a simplified version of a population model describing the weight loss trajectories following daily s.c. administration of liraglutide 3.0 mg [5]. Response adaptive allocation probabilities were driven by the D-optimality criterion, computed 1) for the true simulation model (SM), 2) the best model in terms of AIC (MS), or 3) by weighting models based on MA. Five, prespecified, candidate model structures were used to estimate the underlying Dose-Response (D-R) relationship (no drug effect, linear, log-linear, Emax and sigmoid Emax). Estimation and simulation of the time-course shared the same model structure (turnover model). Subjects were randomized 1:6 to placebo or one of six active treatment doses (adaptations were performed for the active groups only). The adaptive designs consisted of 1) a fixed allocation period (learning phase) and 2) an adaptive allocation period (adaptive phase). We assumed that subjects could be enrolled in the trial with a rate of five subjects per week. During the learning phase, all subjects were randomized equally across arms. When 25% of the prespecified total study population completed 25% of the total trial duration, the adaptive phase was initiated. A stopping rule was included to stop the trials early for futility (lack of drug effect).
The proposed methodology was tested under three scenarios:
- Informative: drug effect present and explored doses distributed around the hypothetical ED50
- Uninformative: drug effect present and explored doses below the hypothetical ED50
- No drug effect
All designs were compared using dose-finding predictive performance criteria. 500 trials were simulated for each scenario and each modeling approach. Model based optimizations and MBAODs were performed using the R package PopED [6] and MBAOD [7], respectively. All simulations and estimations were performed in NONMEM version 7.3 [8] using PsN [9].
Results: Our results show that MBAODs lead to more accurate dose-finding as compared to fixed designs. When accounting for model uncertainty using MS or MA, the predictive performance was comparable to the one obtained under the true simulation model. For the informative scenario, all methods led to accurate and precise prediction of the minimum effective dose. For the uninformative scenario, and no design adaptations, MA led to better predictive performance as compared to MS. When the designs were allowed to adapt, improvements in both accuracy and precision of the minimum effective dose were observed across all evaluated methods. The best performance improvement was seen for the MA approach when the true simulation model was not included in the estimation models set, followed by the MS approach. For the no drug effect scenario, an early stop for futility was observed for SM and MS, but not with the MA approach, where the majority of simulated trials failed to stop for futility.
Conclusions: Our results indicate that model structure uncertainty can be taken into account when performing MBAOD with minimal to no impact to model-based decision-making at the end of a dose-finding study as compared to the “best-case” situation where the ground truth is known a priori. The best results were seen when MBAOD methods were paired with a model-selection step at the interim evaluations. Overall, the proposed methods can help to relax the assumptions regarding the model structure, and potentially increase the applicability of MBAOD in dose-finding trials.
References:
[1] Strömberg EA, Hooker AC. The effect of using a robust optimality criterion in model based adaptive optimization. J Pharmacokinet Pharmacodyn. 2017;44(4):317-24.
[2] Foo LK, Duffull S. Adaptive optimal design for bridging studies with an application to population pharmacokinetic studies. Pharm Res. 2012;29(6):1530-43.
[3]. Maloney A, Karlsson MO, Simonsson US. Optimal adaptive design in clinical drug development: a simulation example. J Clin Pharmacol. 2007;47(10):1231-43.
[4] Pierrillas PB, Fouliard S, Chenel M, Hooker AC, Friberg LF, Karlsson MO. Model-Based Adaptive Optimal Design (MBAOD) Improves Combination Dose Finding Designs: an Example in Oncology. AAPS J. 2018;20(2):39.
[5] Papathanasiou T, Strathe A, Agersø H, Lund TM, Overgaard RV. Impact of dose-escalation schemes and drug discontinuation on weight loss outcomes with liraglutide 3.0 mg: A model-based approach. Diabetes Obes Metab. 2020.
[6] Nyberg J, Ueckert S, Strömberg EA, Hennig S, Karlsson MO, Hooker AC. PopED: an extended, parallelized, nonlinear mixed effects models optimal design tool. Comput Methods Programs Biomed. 2012;108(2):789-805.
[7] Hooker AC, JGC van Hasselt. MBAOD: Model Based Adaptive Optimal Design (2014): GitHub; [accessed 2020 May 10]. https://github.com/andrewhooker/MBAOD.
[8] Beal S, Sheiner L, Boeckmann A, Bauer R. NONMEM user’s guides. (1989–2009). Ellicott City: Icon Dev. Solut. 2009.
[9] Lindbom L, Pihlgren P, Jonsson EN, Jonsson N. PsN-Toolkit–a collection of computer intensive statistical methods for non-linear mixed effect modeling using NONMEM. Comput Methods Programs Biomed. 2005;79(3):241-57.
Reference: PAGE 30 (2022) Abstr 10225 [www.page-meeting.org/?abstract=10225]
Poster: Methodology - Study Design