An Escalation for Bivariate Binary Endpoints Controlling the Risk of Overtoxicity (EBE-CRO): Managing Efficacy and Toxicity in Early Oncology Clinical Trials

ABSTRACT

For about a decade, early clinical development in oncology is facing new challenges. This is due to two main reasons. The first one is linked to the developed molecular targeted agents (MTA) themselves for which the maximum tolerated dose (MTD) is no longer the only dose of interest. The second reason is related to the fact that costs and attrition rates are huge. When selecting a dose, evidence of early activity signal becomes required for future engagements. This implies the need to handle both toxicity and activity endpoints in the analysis and also in the dose escalation design of early-phase trials. We propose a model-based design taking into account both safety and activity for dose escalation. The proposed model involves a bivariate Gaussian latent variable depending on a monotonic toxicity curve and a quadratic activity curve. This model is fitted under the Bayesian framework that allows the incorporation of prior information. The predictive distributions of dose-response curves are used to lead the dose recommendation. Uncertainty in the dose–response relationship is taken into account to calculate the probability of being an over-toxic or a target dose. The proposed design is compared to two other widely used methods.

KEYWORDS:

• Adaptive design
• Bayesian
• dose finding
• safety
• efficacy

Acknowledgments

The authors thank Eric Parent, Pierre Mancini, and Xavier Paoletti for their advice and comments. The authors also thank the referees for helpful comments that improved the manuscript.

Notes

¹ https://biostatistics.mdanderson.org/SoftwareDownload/SingleSoftware.aspx?Software_d=2

² One may replace by to normalize the drug dose.

³ The truncated prior distribution is very useful if the posterior joint distribution is obtained by Gibbs sampling, since the truncation is simply transfered from the prior distribution to the full conditional distribution (and thus on the posterior distribution).

⁴ Correlated scenarios weren’t implemented in the EffTox and bCRM software. Hence, this simulation study wasn’t performed for both designs.

⁵ Correlated scenarios weren’t implemented in the EffTox and bCRM software. Hence, this simulation study wasn’t performed for both designs.