Bayesian Nonparametric Common Atoms Regression for Generating Synthetic Controls in Clinical Trials

Area of Science:

• Biostatistics
• Health Informatics
• Clinical Trial Design

Background:

• Randomized controlled trials (RCTs) are expensive and challenging.
• Electronic health records (EHR) offer a valuable source of real-world data.
• Supplementing traditional trials with real-world evidence is increasingly important.

Purpose of the Study:

• To develop a method for constructing synthetic control arms using EHR data for single-arm trials.
• To propose a novel nonparametric Bayesian common atoms mixture model for this purpose.
• To enable robust inference of treatment effects using real-world data.

Main Methods:

• Utilized EHR data to identify equivalent patient strata compared to the treatment arm.
• Employed a nonparametric Bayesian common atoms mixture model.
• Implemented a density-free importance sampling scheme for data resampling.
• Constructed synthetic control arms for single-arm trials.

Main Results:

• The proposed method demonstrated higher statistical power in detecting treatment effects compared to alternatives.
• Effectiveness was particularly notable for nonlinear response functions.
• The method was successfully applied to glioblastoma studies using historical trial data.

Conclusions:

• The novel Bayesian approach effectively generates synthetic control arms from EHR data.
• This method enhances the power of treatment effect inference in single-arm trials.
• It offers a cost-effective and efficient alternative to traditional RCTs, especially in oncology.