Bayesian Approach to Personalized Benefit-Risk Assessment

ABSTRACT

Benefit-risk assessment is critical in evaluating the effectiveness of a new drug before and after the approval. Some benefit-risk measures depend on the probabilities of benefit-risk categories in which the subject-level benefit and risk outcomes are characterized. The existing benefit-risk methods for analyzing the categorical data depend only on the frequencies of mutually exclusive and collectively exhaustive categories that the subjects fall in, and thus ignore the subject-level differences. We propose a Bayesian method for analyzing the subject-level data with multiple visits. A generalized linear model is used to model the subject-level response probability, with respect to a “reference” category, assuming a logit model with subject-level category effects and multiple visit effects. The random longitudinal visit effects are modeled by a multivariate normal distribution with zero means and first-order autoregressive structured variance-covariance matrices. In the proposed Bayesian setup, a Dirichlet process is used as a prior for the subject-level category effect to catch the similarity among the subject responses. We develop an efficient Markov chain Monte Carlo algorithm for implementing the proposed method, and illustrate the estimation of individual benefit-risk profiles through simulation. The performance of the proposed model fit is evaluated using two model selection approaches, namely, the deviance information criterion (DIC) and the log-pseudo marginal likelihood (LPML). We analyze a clinical trial data using the proposed method to assess the subject-level or personalized benefit-risk in each arm, and to evaluate the aggregated benefit-risk difference between the treatments at different visits.

Keywords:

• Benefit-risk measures
• Dirichlet process
• Generalized linear model
• Gibbs sampling
• Metropolis–Hastings algorithm
• Model selection

Acknowledgments

The authors thank two anonymous referees whose comments have greatly helped improve the contents of this article. This article reflects the views of the authors and should not be construed to represent FDA's views or policies.