Mobile Menu

Data-driven eligibility criteria may make clinical trials more inclusive

Researchers have shown that relaxing the eligibility criteria of clinical trials in a data-driven manner can significantly increase the number of recruited patients with only marginal effects on trial hazard ratios. This approach may aid in designing inclusive clinical trials, without compromising on safety. 

Eligibility criteria for clinical trials

To be enrolled in clinical trials, patients must first fulfill a set of eligibility criteria. Some of these are established to reduce the risk of severe toxicity adverse events. However, the criteria are often too stringent and sometimes poorly justified. Therefore, any clinical trials suffer from low enrolment and most do not complete patient recruitment within the targeted time.

Furthermore, restrictive trials do not accurately reflect drug efficacy and safety in the populations who will use the drug after approval. Overly strict eligibility criteria may thus limit the access of patients to potentially beneficial treatments.

Broadening eligibility in a data-driven manner would thus accelerate the recruitment process and improve the generalisability of clinical trials. Importantly, this would also make clinical trials more inclusive and representative of real-world populations.

In a recent study, researchers from Stanford University and Genentech in California developed an in silico framework, Trial Pathfinder, to assess the impact of loosening eligibility criteria on treatment efficacy and cohort size in a real-world population. Their work has been published in Nature.

Trial Pathfinder

Trial Pathfinder emulates clinical trials in silico using real-world data. It first encodes the eligibility criteria from an existing clinical trial protocol into programmatic logic statements. Next, it selects individuals from a real-world dataset who meets the original requirements. Trial Pathfinder then performs survival analysis on the emulated treatment group, yielding the number of eligible patients and the resulting hazard ratio. A hazard ratio is the measure of the effect of a treatment on an outcome of interest over time. Eligibility criteria may also be altered systematically in silico to quantify how different criterion combinations affect hazard ratios. 

Developing Trial Pathfinder with the Flatiron Health database

The researchers used data from the Flatrion Health electronic health record database, which includes anonymised data from ~280 cancer clinics across the USA. Specifically, they focused on analysing 61,914 patients with advanced non-small-cell lung cancer (aNSCLC).

A total of 10 completed aNSCLC clinical trials were filtered out to be analysed with Trial Pathfinder. The trials had available trial protocols and at least 250 patients in the Flatiron Health dataset who matched the description of patients recruited for the trials. Each trial had an average of 5,167 patients.

With the Flatiron Health data, the researchers encoded common eligibility criteria, including patient characteristics, laboratory values, biomarkers and previous treatments. Although the trialled drugs are all checkpoint inhibitors, each trial had unique eligibility criteria. 

Each aNSCLC trial was then emulated with the original encoded eligibility criteria. The results were consistent with those obtained from the original randomised trials. On average, only 30% of patients in the Flatiron database who have taken the drug met the trials’ requirements. Furthermore, the hazard ratio of the full patient population is comparable or smaller to that of the patients who met the eligibility criteria.

Effects of eligibility criteria

The researchers then simulated thousands of synthetic cohorts with the Flatiron database under different criteria and estimated the hazard ratio of overall survival for each. To summarise the influence of each criterion on the hazard ratio per cohort, they used the Shapley value, which is the weighted average of the effect of adding each criterion to different sets of inclusion or exclusion rules. A Shapley value less than zero indicates that a criterion is beneficial as it decreases the trial’s hazard ratio.

The results suggested that many commonly used criteria do not significantly affect the hazard ratio or reduce trial efficacy. Moreover, it was found that in many cases the eligibility criteria were too restrictive. Patients excluded by such criteria benefitted from the treatments similarly to the patients who satisfied them. These criteria included conditions analysed by laboratory tests (e.g. blood pressure, lymphocyte or neutrophil count) and previous treatments.

For each trial, the researchers then relaxed the eligibility criteria, retaining only those that had a Shapley value of less than zero. Though the data-driven criteria reduced the hazard ratio of overall survival by 0.05 compared with using the full eligibility criteria, the number of eligible patients increased by 107%.

Analysis of three other cancer types yielded similar results, where the original eligibility criteria were overly restrictive. Data-driven criteria also significantly increased the patient population with only a marginal decrease in hazard ratio of overall survival.


Overall, this study demonstrated that many common eligibility criteria had minimal effects on a clinical trial’s hazard ratio on survival. Furthermore, relaxing the criteria more than doubled the number of eligible patients with marginal decreases of trial hazard ratios. Employing a data-driven approach to evaluate eligibility criteria may thus make clinical trials more inclusive for diverse, real-world populations. As more high-quality data becomes available, this work may be expanded beyond oncology to make clinical trials universally more representative of the patients for which they are intended to benefit.  

Image credit: Racool_studio – Freepik

More on these topics

AI / Clinical Trials / Real World Data

Share this article