Abstract
Conducting phase I trials is particularly challenging when dealing with late-onset dose-limiting toxicity (DLT) or when patient accrual is rapid relative to the DLT assessment window. In such cases, a new cohort may be ready for enrollment before the DLT assessments are complete for the current cohort, complicating dose assignment decisions. The time-to-event Bayesian optimal interval (TITE-BOIN) design was developed to address this issue by enabling real-time dose assignment for new cohorts, even when some enrolled patients' DLT data are still pending. This design has been increasingly adopted in practice. Upon reviewing trial protocols utilizing TITE-BOIN, we observed considerable variation in dosing decision criteria, ranging from no requirements to requiring that at least two thirds of enrolled patients' DLT data should be available before enrolling the next cohort. This raises a critical question: How can safety comparable with the fully staggered design be ensured? This article addresses this question by evaluating how the percentage of pending patients and the duration of their follow-up affect the operating characteristics of the TITE-BOIN design. Based on these evaluations, we generally recommend suspending accrual to allow for longer safety follow-up if fewer than 51% of patients at the current dose have their DLT data ascertained or if TITE-BOIN recommends dose escalation and the shortest follow-up time for pending patients at the current dose is less than 25% of the DLT assessment window. The rules should be further calibrated to suit specific trial considerations. Equipped with these rules, TITE-BOIN achieves similar accuracy in identifying the MTD and comparable safety to the fully staggered approach while substantially reducing trial duration and increasing patient access by enabling real-time dose decisions when some patients' DLT data are still pending.