A Bayesian design with efficient backfilling and expansion to accelerate phase I/II trials accommodating late-onset competing risk outcomes

Novel early-phase oncology studies increasingly seek an optimal dose (OD) by jointly considering toxicity and efficacy rather than relying only on the maximum tolerated dose (MTD). In many trials of such agents, the occurrence of either dose-limiting toxicity or disease progression typically triggers treatment discontinuation. Thus, only the first event is observed, yielding a competing-risks data structure that is further complicated by late-onset outcomes and rapid patient accrual. To fill these gaps in current design considerations, we propose a Bayesian seamless Phase I/II design that integrates overdose-controlled dose-escalation with concurrent backfilling, followed by an optimization expansion stage. The design mitigates potential conflicts between dose-escalation and backfilling data by borrowing information across doses through parsimonious yet flexible parameterizations of cause-specific hazards, and it defines the OD through a transparent, clinician-elicited risk–benefit tradeoff on the cumulative incidence function scale. Practical implementation is supported by explicit per-dose enrollment caps that bound the maximum sample size while promoting efficiency through targeted, adaptive backfilling and expansion. Simulation studies demonstrate favorable operating characteristics and robust performance across a range of sensitivity analyses. A hypothetical trial example is provided in the supplementary materials to facilitate practical implementation.