Assessment of various continual reassessment method models for dose-escalation phase 1 oncology clinical trials:...

The continual reassessment method (CRM) identifies the maximum tolerated dose (MTD) more efficiently and identifies the true MTD more frequently compared to standard methods such as the 3 + 3 method. An initial estimate of the dose-toxicity relationship (prior skeleton) is required, and there is limited guidance on how to select this. Previously, we compared the CRM with six different skeletons to the 3 + 3 method by conducting post-hoc analysis on a phase 1 oncology study (AZD3514), each CRM model reduced the number of patients allocated to suboptimal and toxic doses. This manuscript extends this work by assessing the ability of the 3 + 3 method and the CRM with different skeletons in determining the true MTD of various "true" dose-toxicity relationships.

One thousand studies were simulated for each "true" dose toxicity relationship considered, four were based on clinical trial data (AZD3514, AZD1208, AZD1480, AZD4877), and four were theoretical. The 3 + 3 method and 2-stage extended CRM with six skeletons were applied to identify the MTD, where the true MTD was considered as the largest dose where the probability of experiencing a dose limiting toxicity (DLT) is ≤33%.

For every true dose-toxicity relationship, the CRM selected the MTD that matched the true MTD in a higher proportion of studies compared to the 3 + 3 method. The CRM overestimated the MTD in a higher proportion of simulations compared to the 3 + 3 method. The proportion of studies where the correct MTD was selected varied considerably between skeletons. For some true dose-toxicity relationships, some skeletons identified the true MTD in a higher proportion of scenarios compared to the skeleton that matched the true dose-toxicity relationship.

Through simulation, the CRM generally outperformed the 3 + 3 method for the clinical and theoretical true dose-toxicity relationships. It was observed that accurate estimates of the true skeleton do not always outperform a generic skeleton, therefore the application of wide confidence intervals may enable a generic skeleton to be used. Further work is needed to determine the optimum skeleton.