Integrating External Data in Clinical Trials: From Traditional Borrowing Methods to Machine Learning Approach

This thesis focuses on the incorporation of external data in clinical trial. Randomized clinical trials are the gold standard for investigating the relationship between a new treatment and the outcome of interest. Sometimes, it can be difficult to reach the adequate sample size for detecting a significance of the treatment effect; for example, when dealing with rare disease because of the small-sized populations, or, in case of life-threaten diseases, where it is unethical to assign patients to the control group, as little benefit is expected with respect to the experimental arm. To overcome these issues, there has been a growing interest in leveraging external data for enhanced decision making. Data from previous clinical trials or observational studies constitute a valuable resource to augment or fully replace the control arm of current trials. The integration of different data sources comes with several challenges, such as the quality and the heterogeneity between studies. For this reason, a family of statistical techniques, called borrowing methods, have been developed for integrating external and current data. A scoping review about borrowing methods was conducted to provide a comprehensive overview of the amount of available methods, their characteristics and clinical applications. Then, a simulation study was performed to compare the operating characteristics of some of them in the context of a phase II single-arm clinical trial for evaluating the safety of a new medical device. %Consecutive versions of a medical device have the same action mechanism, making them suitable for the borrowing purpose. Finally, a new machine learning-based method was developed to borrow external data in a new clinical trial, providing an innovative application of distance metric learning approaches in the clinical setting.