Incorporating pre-clinical information into phase I trials: a Bayesian decision-theoretic approach
Haiyan Zheng, Lancaster University
Bayesian model-based procedures become increasingly important to design phase I dose-escalation trials, where the principal aim is to estimate the maximum tolerated dose of a novel compound. Conventionally, vague prior distributions have been used for model parameters. However, information on the dose-toxicity relationship is commonly available from pre-clinical toxicology studies by the time first-in-man trials are conducted. Should the pre-clinical experimental findings be commensurate with the toxicity in humans, incorporating them into the prior will lead to more efficient decision process and enhanced precision of the resulting recommended dose in the phase I trial. Such advantages, however, must be balanced against the risk that more patients may be treated with excessively toxic doses when prior-data conflict commences.
We propose a Bayesian decision-theoretic approach, with which the degree of commensurability can be dynamically measured during the course of a fully sequential phase I trial. Central to this approach is a utility function that formalises the benefits of incorporating pre-clinical information under various potential risks. In particular, pre-clinical information is used to predict whether the incoming patients would experience dose-limiting toxicity (DLT) or not. These predictions are optimal in the sense of maximising the prior expected utility. At each interim analysis these prior predictions are compared with the actual human outcomes. The attained predictive utility, expressed as a fraction of the maximum utility achieved when all prior predictions are correct, is then helping quantify the weight to be attributed to the pre-clinical information. Simulations demonstrate our approach is competitive and robust, and can essentially improve the design and analysis of phase I trials without undermining the safety of patients.
Symmetric loss functions in restricted parameter spaces
Pavel Mozgunov, Lancaster University
For a parameter defined on the whole real line, the squared error loss function is a proper choice as it infinitely penalizes boundary decisions. However, this approach might lead to sub-optimal solutions in problems when a parameter is defined on a restricted space. We invoke some general principle that an appropriate loss function should infinitely penalize boundary decisions. We propose several generalizations of the squared loss function for parameters defined on the positive real line and on an interval. Superior properties of the proposed loss functions over the squared loss function are demonstrated by estimation examples for widely-used distribution. We would also show how the proposed form of the loss functions can be used in dose-finding clinical trials.
Add to my calendar