Trial predictions vs. trial simulations in Model-based Drug Development: integrating uncertainties to evaluate the predictive probability of success.

In a Model-Based Drug Development strategy, the first objective is to design studies such that the most reliable model estimates are obtained, in order to optimize the design of future studies and to take decisions based on predictions. The objectives of the work is to present from a theoretical and practical point of view how to perform trial predictions, as opposed to trial simulations, by integrating the uncertainty of the parameters. The difference between prediction and simulation is important in early development when limited data or prior information are available. Indeed ignoring the uncertainty of parameter estimates can lead to wrong decisions. First, will be provided methodology, derived from Bayesian statistics, to perform trial predictions from the parameter estimates and their uncertainty, when obtained with conventional frequentist population methods. Second, a practical implementation in R will be shown. This generalized prediction shell can cope with any kind of structural population models: Ordinary Differential Equation, single & multiple doses, infusion, etc... The proposed shell is also flexible to allow the testing of various scenarios and study designs, and therefore evaluate the predictive probability of success of different protocols. When joint models for efficacy and safety are established, the Prediction-based Clinical Utility Index (p-CUI) and its distribution can directly be obtained for more riskless decision making. Examples will be shown to highlight in early phases the differences existing between trial prediction and trial simulation. This approach is required to permit Model-Based Drug Development strategy, and impact successfully decision in early clinical phases