A clinical trial was commenced by a large pharmaceutical company but recruited slowly and was terminated early, with only 68% of the planned number of paediatric subjects recruited. Following their review, the Paediatric Committee (PDCO) of the EMA recommended that an extrapolation study be conducted to evaluate the efficacy of the drug in the recruited paediatric sample. To do this, a Bayesian extrapolation approach was required, where data from previous adult studies was leveraged to increase precision for the final efficacy endpoint in the paediatric population. A tight timeline was stipulated for completion of the analysis.
The recommended analysis would involve multiple stages requiring specific expertise deployed intensively to meet the overall timeline with no guarantee of success. Prior to committing to the work KerusCloud was used to rapidly evaluate the feasibility and probability of success of the recommended approach, taking into account the volume of data available and anticipated effect sizes and variability. This allowed confirmation that the approach had a good chance of providing robust evidence of efficacy of the new drug in paediatric subjects.
The analysis could then proceed with greater confidence through the following steps:
- Adult studies for the drug, which looked at similar endpoints, were identified in the literature.
- The paediatric data was first analysed using frequentist techniques which had been previously employed within the adult studies and the results were compared to ensure consistency, increasing confidence in the appropriateness of implementing an extrapolation approach.
- A primary Bayesian extrapolation analysis was conducted, initially using data from a single adult study, and then using the results of a meta-analysis from multiple adult studies as a sensitivity analysis.
- Initially, a weak and informative prior distribution for the primary endpoint were defined (Figure 1). The informative prior was specified using the mean and standard error of the adult data.
- A Bayesian ‘dynamic borrowing’ approach was then employed, where the informative prior was combined with the weak prior, resulting in a robust weighted mixture prior distribution (Figure 2), allowing the analysis to learn how much of the adult information to borrow, based upon the consistency observed between the paediatric data and the adult prior.
Figure 1. Informative and weak priors
Figure 2. Robust Mixture priors
KerusCloud showed that the Bayesian dynamic borrowing approach was feasible and had a good probability of success before any work was undertaken.
The subsequent analysis went on to show:
- The available efficacy data for the paediatric population appeared consistent with that for the larger adult population, as there were no observed differences in treatment effect across the adult and paediatric populations.
- A significant increase in the primary endpoint in paediatric subjects could be supported through the extrapolation analysis with only a relatively small (25%) prior confidence in the paediatric treatment effect being similar to the adult treatment effect.
- Extrapolation of drug efficacy in adults to paediatric subjects could be supported in this case as part of a robust evidence package for Regulators.