2004
Uppsala, Sweden
Modelling and simulation of the incidence of adverse events in clinical trials.
Filip De Ridder (1), An Vermeulen (2) and Vladimir Piotrovskij (2)
(1) Biometrics & Clinical Informatics; (2) Global Clinical Pharmacokinetics & Clinical Pharmacology, Johnson & Johnson Pharmaceutical Research and Development, B-2340 Beerse, Belgium.
Objectives: Predict the incidence of a pre-defined adverse event in dose ranging trials of a new compound. The adverse event of interest is typical for the class to which the new compound belongs.
Methods: A PK/PD model was developed using data from Phase II/III trials of a marketed compound of the same class (10 studies, ± 2500 patients). A time-to-event (hazard modelling) approach was used to model the incidence of the adverse event of interest (number of patients with at least 1 episode of the event/total number of patients randomised). This approach utilizes all available information and allows a flexible incorporation of time-varying covariates, including drug exposure, varying trial duration and dropout. A parametric model was developed describing the hazard as a function of time and covariates, including drug exposure. A similar model incorporating the drop out mechanism was also developed. Patient-specific measures of drug exposure were derived from an available population PK-model (developed in NONMEM). The hazard models were fitted by maximum likelihood and implemented in the SAS-procedure NLMIXED. The model was validated against other trials with the same compound. Using these hazard models, a population PK-model based on early Phase I data of the new compound, and adopting a few basic pharmacological and pharmacokinetic assumptions, clinical trial simulations were performed to predict the incidence of the adverse event in the planned dose-ranging trials. Sensitivity analyses were performed to assess to impact of uncertainty in some aspects of the model.
Results: The hazard for developing the adverse event of interest was best described by a steep - virtually on/off - relationship, with patient-specific average steady state concentration. This translated into shallow dose-response relationship for the new compound. Clinical trial simulation allowed to assess the incidence for the planned dose-ranging trials and to balance safety and efficacy.
Conclusion: Hazard modelling is a flexible tool to model the occurrence of (adverse) events in clinical trials. In this application, it allowed to synthesize relevant data of a marketed compound. Simulation allowed studying what can be expected in planned clinical trials based on available knowledge and gaps therein.