Advanced Bayesian Approaches to Borrow Strength from Real World, Historical Trial Data, and Related Subgroups for Efficient Drug Development

Using external control data can improve the efficiency of randomized controlled trials (RCTs). However, the heterogeneity of distributions of covariates between current and external control data complicates direct borrowing. Covariate adjustment methods, such as outcome regression or propensity score approaches, combined with dynamic borrowing techniques, have been proposed to mitigate this problem. However, key prognostic factors may be partially missing in real world data, forcing researchers to rely on complete case analyses. Under certain covariate-dependent missingness mechanisms, outcome regression models can still yield unbiased estimates, whereas propensity score methods may fail. In this study, we propose a Bayesian g-formula approach that dynamically borrows external control data with partially missing data. Through simulation study, we evaluate the performance of our proposed method relative to existing approaches, including propensity score matching with dynamic borrowing. Our findings highlight the potential advantages of the Bayesian g-formula framework in complete case analyses while preserving the benefits of external data utilization. 

The primary goal of dose-finding trials for novel anti-cancer agents is to identify an optimal dose (OD) that balances efficacy and tolerability, unlike traditional cytotoxic agents where the maximum tolerated dose (MTD) is sought. Novel agents may exhibit non-monotonic dose–efficacy relationships, necessitating model-based and model-assisted designs to determine ODs. The FDA's Project Optimus aims to optimize cancer drug doses for efficacy, safety, and tolerability. Advances in biomarker development and precision medicine have shifted cancer treatment paradigms towards targeting specific molecular profiles, requiring new trial designs like early-phase basket trials. These trials assess new treatments across multiple cancer types or subtypes, identifying sensitive and insensitive baskets. We extend the BOIN-ET design to phase I/II basket trials, introducing the "BOIN-ETB" design, which is model-assisted and straightforward to implement. A simulation study explores the BOIN-ETB design's properties and compares its operating characteristics with other dose-finding approaches in oncology. 

In randomized Phase III oncology trials, long-term time-to-event endpoints are crucial for assessing treatment benefits. However, Phase II trials often rely on short-term binary tumor response as a surrogate endpoint, which can lead to high failure rates in Phase III trials, as tumor response may not accurately reflect survival benefits. Additionally, many oncology trials collect biomarker data to identify participants more likely to respond to experimental treatments, highlighting the need for biomarker-based designs to enrich trial populations.
This presentation introduces a constrained hierarchical Bayesian model considering latent biomarker subgroups (CHBM-LS) for long-term time-to-event endpoints in Phase II randomized trials. CHBM-LS aggregates biomarker populations into latent subgroups and addresses treatment effect heterogeneity across biomarker levels. We compare our design with other approaches, demonstrating that CHBM-LS improves the accuracy of hazard ratio estimates and enhances the power to detect true effects while maintaining control over the Type I error rate. 

Evidence-based decision-making is crucial at every stage of clinical development. In common diseases, phase II proof-of-concept (PoC) studies play a vital role as gatekeepers to increase the efficiency of asset selection for late-stage development by guiding early decisions to terminate the development of ineffective assets and accelerate the development of promising ones. The process of developing Go/No-Go decision criteria and examining the study operating characteristics is called quantitative decision-making (QDM) assessment which triggers cross-functional discussions about optimality of the entire clinical development plan. However, in rare disease drug development, phase II PoC studies are often small or even infeasible due to the rarity of target diseases, and thus the PoC declaration is frequently addressed by insufficiently informative data, which leads to high probabilities of inconclusive results and undermines the ability to make definitive Go/No-Go decisions. A potential solution to the QDM assessment with limited sample size is to utilize a Bayesian framework for incorporating existing prior information such as data from previously completed clinical trials and real-world data. We introduce a technique of Bayesian information borrowing into the QDM assessment via power prior, allowing for the incorporation of historical data, which makes the Go/No-Go decision reliable even with the limited amount of PoC data. For instance, data from dose-escalation cohorts can be leveraged to augment preliminary efficacy data of the treatment, and furthermore historical trial data or real-world data can serve as external controls when concurrent controls are not available. Through a simulation study, we present the operating characteristics of the Bayesian QDM with information borrowing. 

Randomized Controlled Trials (RCTs) are the gold standard in clinical trials. If Historical Data (HD) on the standard of care is available, hybrid control design can provide more evidence than a standalone RCT with unequal allocation. HD for the control group can often be derived from Real World Data, which frequently includes missing covariates data. However, such missingness may introduce bias depending on missing data mechanisms and analytical methods. In this study, we propose addressing covariate missingness under the missing at random assumption by multiple imputation. In the analysis stage, we utilize a combination of propensity score matching and modified power prior. The simulation showed that complete case analysis caused bias under outcome and covariate-dependent covariate missingness, while multiple imputation provided nearly unbiased estimates and improved precision when HD was similar to the current trial data. HD was dynamically borrowed based on outcome similarity: improved estimation accuracy with reduced bias when outcomes were similar, while reasonably controlling the type 1 error when outcomes were dissimilar.