BAYESIAN ADAPTIVE DOSE-FINDING BASED ON EFFICACY AND TOXICITY

Abstract

A description of the conventional paradigm for early phase clinical evaluation of a new
agent is given, followed by a list of this paradigm’s logical and practical flaws. This is
provided initially to motivate the use of phase I-II clinical trial designs. The main body
of the paper consists of a review of several practical Bayesian phase I-II designs for sequentially
adaptive dose-finding based on efficacy and toxicity. These include designs taking
two general approaches. The first approach uses elicited efficacy–toxicity probability
pair trade-offs as decision criteria. Designs accommodating bivariate binary and trinary
outcomes are discussed, as well as an elaboration that uses patient covariates to choose
individualized doses. The second approach uses elicited joint utilities of ordinal (efficacy,
toxicity) outcomes as a decision criterion, also including adaptive randomization to improve
performance. Several illustrative applications of the methods are provided.