gms | German Medical Science

As the treatment landscape in oncology is evolving, the paradigms that have been established over the past decades for assessing the safety and efficacy in clinical trials need to be adapted. Traditionally, the goal of Phase I dose escalation trials is to establish the maximum tolerated dose (MTD), based on the assumption that both, efficacy and toxicity, increase with increasing dose level. However, as new treatments like targeted agents are emerging, which are less toxic and more tumor-specific, it may no longer be necessary to treat patients at the MTD. In some cases, an MTD may not even be reached among the investigated dose levels. In those cases, other parameters, like pharmacodynamic/ pharmacokinetic markers or level of target engagement, are used to indicate potential efficacy and, thereby, can help to determine an efficacious dose (or dose range) among the safe doses.

For many years the standard design for Phase I was the rule-based 3+3 design. Nowadays, Bayesian designs like the Bayesian two-parameter logistic regression model (BLRM) proposed by Neuenschwander et al. [1] are used more often as they have shown advantages in terms of flexibility and operational characteristics over the 3+3 design. Those Bayesian designs model the dose toxicity curve while taking the associated uncertainty into account and then recommend a dose for the next cohort based on all available data. However, all these designs are focused on dose limiting toxicities (DLTs) and the MTD.

In recent years, there have been suggestions how to incorporate efficacy parameters into Bayesian dose escalation. A recent and a novel approach for incorporation of efficacy data into Bayesian modeling will be presented and compared.

The first approach, a gain function suggested by Yeung et al. [2], combines separate models: a standard BLRM to model the toxicity (dose vs. DLT), and a Bayesian conditional linear model to model the efficacy response (dose vs. efficacy marker). Based on these two models, a gain function is set up which models the safety-efficacy trade off and which then recommends the dose with the maximum gain for next cohort.

Our novel approach uses methods by Tervonen et al. [3] which were originally developed for late stage or post-marketing benefit risk analyses. Separate Bayesian models for safety and efficacy are set up, equivalent to the first approach. Using the posterior samples of the models, rank acceptability indices are calculated for each dose based on their benefit-risk score distribution. The dose with the highest rank is then the preferred dose for the next cohort.

The applicability of the two approaches will be compared for application examples. Furthermore, aspects like performance, extendibility, generalizability, and the sensitivity to parameter changes will be evaluated. Overall, the feasibility of incorporating efficacy modelling into Bayesian dose escalation designs will be examined and discussed.

The authors are employees of Merck Healthcare KGaA.

The authors declare that an ethics committee vote is not required.