<scp>BOIN</scp> ‐ <scp>MEM</scp> : A Two‐Stage Design for Dose Optimization With Information Borrowing Across Dose Levels and Stages in Oncology Phase I/ <scp>II</scp> Trials

The launch of Project Optimus has shifted a paradigm in oncology early-phase trials toward identifying the optimal biological dose (OBD), considering both the drug's toxicity and efficacy. Conventional two-stage designs for dose optimization in phase I/II trials typically identify tolerable dose levels in Stage 1, followed by selecting the OBD based on efficacy at expansion cohorts in Stage 2. However, these designs could be inefficient because toxicity and efficacy data from patients in Stage 1 treated at dose levels excluded from Stage 2 are not explicitly used in the final OBD determination. To address this issue, we developed a novel two-stage design for dose optimization in oncology phase I/II trials. This design, termed BOIN-MEM, leverages Bayesian optimal interval design (BOIN) for Stage 1 and extends multi-source exchangeability models (MEM) to facilitate information borrowing across dose levels and stages when selecting the OBD at the end of Stage 2. This information borrowing strategy is based on utility for each dose level, which is a measure representing a drug's clinical value, calculated based on the toxicity and efficacy outcomes, and their pre-specified scores. BOIN-MEM allows for the estimation of mean utility, considering heterogeneity not only between dose levels but also between stages due to variations in the target population. Through comprehensive simulation studies with practical scenarios of dose-toxicity and dose-efficacy relationships, BOIN-MEM demonstrated higher accuracy of OBD selection and reduced patient exposure to overdoses compared with existing designs under various scenarios.