Use of Machine Learning Techniques to Create External Control Arms: Guidance for Submissions to Health Technologic Assessment

PURPOSE:

External control arms (ECAs) are increasingly used to contextualize single-arm trial evidence. In the absence of randomization to balance the treatment groups to be compared, clinical profiles from trial patients are typically matched to external controls using propensity score matching before outcome analyses are performed. Alternative analytical approaches based on machine learning (ML) predictions, such as G-computation and Doubly Debiased Machine Learning, are being considered for inferring efficacy outcomes. The use of such advanced techniques is guided by data sparsity and high-dimensionality. For this reason, most health technology assessment (HTA) bodies do not provide specific recommendations on the methods for matching clinical trials to observational data in their guidelines.

The workshop aims to present the benefits and challenges of using matching methods based on machine learning techniques for creating ECAs in clinical trials. Our speakers will engage in real-time polling upfront to understand the audience (5 minutes). The workshop will conclude with an interactive audience discussion (10 minutes).

DESCRIPTION:

Dr. Guelfucci will introduce the topic to the audience and introduce the speakers (5 minutes).

Dr. Guelfucci will provide a comprehensive description of the methods used to compare with external controls in the most recent ECAs submitted to HTA agencies and their appraisals (10 minutes).

Dr. Mayer will delve into the context in which ML-based methods for ECAs are relevant and discuss their data requirements, inference benefits and challenges (15 minutes).

Lastly, Dr. Kreif will address recent developments and the necessary requirements to ensure their suitability for submissions to HTA bodies. She will also explore potential reporting checklists that can assist in meeting these requirements (15 minutes).

The session will close with a Q&A session with the audience on the use and acceptability of ML-based matching methods for external control arms (10 minutes).