Applying artificial intelligence to constructing external control arms that support comparative analyses of single-arm clinical trials using historical clinical trial data or real-world data is an innovation that may enhance the effectiveness and efficiency of drug development through potentially reducing the number of patients, amount of time, and cost needed to advance the drug development cycle.

During the 3rd Annual Pharma AI Summit, with the theme “Leveraging Artificial Intelligence for Driving the Digital Future of Pharmaceutical Industry,” four panelists discussed the key elements toward external control arms. The panel included the following:

• Jim Z. Li, Health Economics and Outcomes Research Team Lead, Global Medical Analytics and Real World Evidence, Viatris
• Aaron Galaznik, Chief Scientific Officer, Carevive
• Elizabeth Lamont, Senior Medical Director for Integrated Evidence Acorn AI, a Medidata Company
• Kelly H. Zou, Head of Global Medical Analytics and Real World Evidence, Viatris

Below are a few questions they answered.

What is an external control arm?

A collection of patients from outside a clinical trial of interest (i.e., target trial) whose outcomes are compared to target trial patients. ECAs have historically been assembled from varied data sources, including disease registries, clinical care data, billing claims, and historic clinical trial data. They can provide useful insights when randomized controlled trials are not possible due to ethical or logistical reasons.

Why consider ECAs for clinical trials?

ECAs have particular value when randomized controlled clinical trials are infeasible or challenging, such as when there is difficulty recruiting patients, severe disease, and paucity of treatment options. In the case of severe diseases without treatment options, having a concurrent comparator may be challenging. In such situations, ECAs may supplement or replace such a comparator to allow investigators to make inferences regarding the efficacy of a novel therapy.

Additionally, ECAs may have value early in the clinical development process through either or both of the following:

1. Informing a go/no-go decision after phase I or phase II trials are complete

2. Providing more accurate effect size estimates, which might ensure the likelihood of well-planned and adequately powered subsequent randomized controlled trials

ECAs may be particularly vital to the timely approval of new therapies for conditions for which it is historically difficult to carry out randomized controlled trials.

Can ECAs help achieve patient diversity?

Because they are created from external data sources, ECAs have the potential to reflect a broader, more diverse patient population than can be accessed within any single randomized controlled trial. This is true whether external controls are created from pooled trial data or real-world data sources. In the case of real-world data sources, one can further extend to populations not normally found in clinical trials or enhance generalizability of the comparator population pool.

Additionally, when ECA’s are constructed via close matching of patients with trial participants, one can further compare patients selected for the ECA versus those who are not to better quantify and contextualize the generalizability of a trial population.

How do you construct ECAs?

It usually follows a three-step approach. The first step is to identify appropriate data sources, which can be from historical randomized controlled trials or real-world data sources. Data from large and well-conducted randomized controlled trials for the same disease may be suitable when conducted following the International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use’s good clinical practice. Such data usually is more accurate and complete than real-world data sources, has baseline demographic and clinical characteristics variables that are similar to the target clinical trial, and is more likely to use similar definitions for disease and measures for outcomes.

The second step is data processing, which includes data ingestion, deidentification, cleaning, and standardization to generate a robust, validated, and analysis-ready data file with the same structure and formats as the data in the target clinical trial.

The last step is data-matching via appropriate statistical and data science methodologies to match and select patients from the analysis-ready data file into a particular ECA. This step is to ensure the selected patients in the ECA will have the same, or highly similar, characteristics as the patients in the target clinical trial.

While there are multiple methods for patient matching, propensity scoring is probably the most used method. This method is also widely used in studies reporting real-world evidence studies, although regression-based adjustments can also be used in those studies.

More advanced methods such as Bayesian mixed models with commensurate and power prior distributions, random forests, neural networks, cluster analysis, and microsimulation have also been used or explored for patient matching. These advanced methods provide a fertile ground for statistics to intersect with data science through machine learning and AI technologies.

Editor’s Note: The views expressed here are the authors’ and do not represent those of their employers.