BOIN12: Bayesian Optimal Interval Phase I/II Trial Design for Utility-Based Dose Finding in Immunotherapy and Targeted Therapies

Abstract

PURPOSE For immunotherapy, such as checkpoint inhibitors and chimeric antigen receptor T-cell therapy, where the efficacy does not necessarily increase with the dose, the maximum tolerated dose may not be the optimal dose for treating patients. For these novel therapies, the objective of dose-finding trials is to identify the optimal biologic dose (OBD) that optimizes patients’ risk-benefit trade-off. METHODS We propose a simple and flexible Bayesian optimal interval phase I/II (BOIN12) trial design to find the OBD that optimizes the risk-benefit trade-off. The BOIN12 design makes the decision of dose escalation and de-escalation by simultaneously taking account of efficacy and toxicity and adaptively allocates patients to the dose that optimizes the toxicity-efficacy trade-off. We performed simulation studies to evaluate the performance of the BOIN12 design. RESULTS Compared with existing phase I/II dose-finding designs, the BOIN12 design is simpler to implement, has higher accuracy to identify the OBD, and allocates more patients to the OBD. One of the most appealing features of the BOIN12 design is that its adaptation rule can be pretabulated and included in the protocol. During the trial conduct, clinicians can simply look up the decision table to allocate patients to a dose without complicated computation. CONCLUSION The BOIN12 design is simple to implement and yields desirable operating characteristics. It overcomes the computational and implementation complexity that plagues existing Bayesian phase I/II dose-finding designs and provides a useful design to optimize the dose of immunotherapy and targeted therapy. User-friendly software is freely available to facilitate the application of the BOIN12 design.

TITE‐BOIN12: A Bayesian phase I/II trial design to find the optimal biological dose with late‐onset toxicity and efficacy

“…Thanks to the use of the quasi‐beta‐binomial model, the values of can be precalculated for all possible outcomes in and theirs ranks can be pretabulated and included as a decision table in the protocol of a trial 17 . As a result, when conducting the trial, simply look up the decision table to determine the most desirable dose in Steps 2b and 2c.…”

“…Another innovation of BOIN12 is that it uses posterior probability Pr(u > u b | D), rather than the posterior mean of u (ie, the most commonly used approach), to determine the most desirable dose for treating patients at each interim, where u b is a utility benchmark. Because the tail posterior probability Pr(u > u b | D) accounts for both the mean location of u and its uncertainty, it leads to more appropriate decisions of dose assignment than the common approach based on the mean of u. Denoting 𝜙 T and 𝜙 E as the upper limit of DLT probability and lower limit of efficacy probability, respectively, the recommended default value for u b is u + (100 − u)∕2, where u 17 The justification of this default value, and several specific examples showing the advantageous property of using tail probability to guide dose assignment, are provided in Section S.1. Consider a phase I/II trial with J doses under investigation.…”

“…Zhou et al proposed the two‐stage U‐BOIN design, in which the trade‐off between efficacy and toxicity is measured by utility elicited from clinicians 16 . Recently, Lin et al proposed the utility‐based Bayesian optimal interval phase I/II (BOIN12) design to find OBD for immunotherapies and targeted therapies 17 . One attractive feature of BOIN12 is its simplicity.…”

Early Completion of Model-Assisted Designs for Dose-Finding Trials

“…The model-assisted designs include the modified toxicity probability interval (mTPI) design, 11 the Keyboard design, 12 the BOIN design, 13,14 and variations of these designs. [15][16][17][18][19] The modelassisted designs are reviewed by Zhou et al 1 and Yuan et al 20…”

Application of Multi-Armed Bandits to Model-assisted designs for Dose-Finding Clinical Trials

“…Recently, model-assisted designs which are no complex assumptions and have superior performance have been proposed. The model-assisted designs include Bayesian optimal interval (BOIN) design [3], keyboard design [4], and related designs [5,6,7,8,9,10,11,12,13]. The BOIN design conducts a dose-assignment by using an interval for the toxicity rate that is optimized by minimizing errors in dose-assignment decisions.…”