NIH’s TrialGPT algorithm uses AI to match patients to clinical trials

Clinical trials can help inform medical treatments - but only if the right study participants are identified. Helping patients find clinical trials they qualify for can be a laborious process for physicians. Now, researchers from the National Institutes of Health's National Library of Medicine have developed an artificial intelligence-based algorithm called TrialGPT to help speed up this process.

The program uses the large language model ChatGPT to streamline such matches. Here's how it works: First, the algorithm processes a patient summary that contains the person's relevant medical and demographic information such as age, sex and major conditions or symptoms. Then, it searches the ClinicalTrials.gov website to identify clinical trials for which the patient may be eligible and outputs a list ranked in order of relevance.

In addition, TrialGPT produces an explanation about how the person meets each study's enrollment criteria to predict more precisely whether they may be a candidate so that a clinician can use the information in conversations with their patient. ClinicalTrials.gov is a federal government database maintained by the National Library of Medicine containing information on hundreds of thousands of studies.

"About 40% of cancer trials failed due to insufficient patient enrollment," the project's principal investigator, Zhiyong Lu, Ph.D., said in a video about the program posted to YouTube. "This is because identifying an eligible patient [for] a particular trial which has many inclusion and exclusion criteria is a very time-consuming and error-prone process."

Increasing representation in clinical trials

Historically, women and people of color have been underrepresented in clinical trials as studies focused on white men as a presumed model for all. Industry sponsors repeatedly conduct research at the same large sites with the same investigators, which generally do not provide care to underserved communities and are often not easily accessible to diverse communities, according to an editorial in MedCity News. Some 50% of clinical trials are conducted in only 2% of ZIP codes, the authors noted, with research being conducted among patients who are largely white, affluent and male.

This is another issue researchers hope TrialGPT can overcome.

Lu and colleagues described the work in the journal Nature Communications in November 2024, in a study co-authored by collaborators from Albert Einstein College of Medicine in the Bronx, N.Y.; the University of Pittsburgh; the University of Illinois Urbana-Champaign; and the University of Maryland, College Park.

The researchers tested TrialGPT using three cohorts of 183 "synthetic" patients (physician-created patients using real medical data) and more than 75,000 trial eligibility notes. TrialGPT could create relevant keywords for searches and readily retrieve 90% of relevant clinical trials, an NIH blog post noted.

Researchers also conducted a user study, asking human clinicians to review six anonymous patient summaries and match them to six clinical trials. For each patient-trial pair, one clinician was asked to review the summaries manually, check if the patient was eligible for the trial, and decide if the person might qualify. Another clinician used TrialGPT to assess the patient's eligibility. Researchers found that clinicians working alone and with the aid of TrialGPT had about the same level of accuracy in matching patients, but the clinicians who used TrialGPT spent about 40% less time screening patients.

Based on these results, the research team was selected for an NIH Director's Challenge Innovation Award to further expand the technology's application to real-world patient data and clinical trials across multiple institutes and centers. The approach is expected to boost the efficiency of patient recruitment while also reducing barriers to research participation among populations that traditionally have been underrepresented in clinical research, the blog post said.

The use of AI to improve patient recruitment, retention and outcomes of clinical trials began even before the advent of ChatGPT, according to an article in Healthcare IT News. During the COVID-19 pandemic, for example, oncology organizations started looking for ways to find patients nationwide who would qualify for trials using health care data. The electronic health record vendor Epic also implemented a clinical trial matchmaking data set, and last October, Microsoft announced new AI tools to enable health systems to build customized programs for administrative tasks including clinical trial matching, the article said.

Other AI initiatives were highlighted in the MedCity News article, including Trial Pathfinder, an AI program developed by researchers at Stanford to address diversity in oncology trials. The Digital Medicine Society in 2023 launched a resource platform to advance diversity, equity and inclusion in digitized clinical trials.

Resources