Revolutionizing Drug Development: Harnessing Real-World Data in Hybrid Trial Designs

We consider two classes of nonparametric Bayesian models utilizing external data for the design or analysis of an ongoing clinical trial. The first class is build on a sequence of dependent random distributions called Shared Atoms Model (SAM) that induces shared and nested clusters for covariates and outcomes. Applying to clinical trials borrowing external information (e.g., RWE) to augment the control arm, SAM attempts to match subpopulations based on measured and unmeasured confounders. The second class of Bayesian models is based on integration of Bayesian additive regression trees (BART) and meta-analytic-predictive (MAP) prior. Using the two models, the new MAP-BART model is able to accommodate potential bias caused by both measured and unmeasured confounders when external data are borrowed to form a hybrid control and infer the treatment effect of a clinical trial. We will demonstrate the methodology using numerical examples.  

Randomized controlled trials (RCTs) are considered the gold standard for evaluating treatment efficacy, but they come with several challenges. These include high costs, lengthy timelines, ethical concerns for participants in placebo or control arms, and issues such as patient attrition and non-compliance. Recruiting patients for the control arm can be particularly challenging, especially in therapeutic areas with high unmet medical needs. To address these issues, hybrid trial designs that integrate external data sources, such as historical controls and real-world data, have emerged as a promising alternative. This paper introduces the Bayesian hybrid design with adaptive sample size through multisource exchangeability modeling (BEAM). The BEAM design leverages a modified multisource exchangeability model to dynamically borrow relevant information from multiple historical data sources, while adaptively adjusting the sample size throughout the trial. This approach ensures that the trial maintains statistical rigor and efficiency, even when heterogeneity exists between current and historical data, and mitigates the challenges associated with control arm accrual and compliance. Through extensive simulations, BEAM demonstrated robust performance in controlling type I error, reducing bias, and maintaining power compared to traditional methods and other adaptive designs. Additionally, the BEAM design offers a versatile and efficient computational framework for optimizing clinical trials, helping to reduce both the cost and time involved in drug development.  

High-quality Real-World Data (RWD) is essential for reliable analysis, yet challenges like missing data, ambiguity, and chronological misalignments frequently arise. In asthma and COPD research using Optum EHR claims data, RWD supports eligibility criteria refinement, power validation, and identification of key populations. However, reliance on complete cases for missing data can introduce selection bias. Traditional imputation methods, like mean and median imputation, are limited in addressing RWD's complexity. Advanced AI methods, such as autoencoders (AEs), variational autoencoders (VAEs), and GANs, offer robust solutions by capturing intricate data relationships. AEs and VAEs use latent spaces for data reconstruction, with VAEs enabling flexible learning of distributions. GANs further improve imputation by generating synthetic data to fill gaps. Beyond imputation, these generative AI models detect anomalies by comparing reconstructed and real data, while Bayesian networks identify low-likelihood records as errors, modeling conditional dependencies. With enhanced RWD, advanced analyses become feasible. Virtual Twins use machine learning and causal inference to pinpoint subgroups, Bayesian networks map data dependencies with transparency, and deep learning integrates unstructured data, refining clinical trial screening and design.