ChatGPT-Like AI Tool Promises to Speed Treatment Advances and Free Doctors’ Time by Matching Patients with Clinical Trials

Monica Bertagnolli shared on LinkedIn:

“A promising new AI algorithm developed by researchers at NIH can help clinicians quickly match their patients to relevant clinical trials for which they could be eligible. This tool can help speed volunteer enrollment and ultimately aid faster progress in medical research, benefitting many.

Clinical trials are essential for advancing new treatments that improve patient care and lives. But far too many clinical trials face challenges in identifying and enrolling eligible trial participants. Now, an NIH-led team has introduced an artificial intelligence (AI) tool that promises to speed up the process of matching patients to clinical trials to help boost enrollment. They call it TrialGPT.

As reported in Nature Communications , TrialGPT takes advantage of large language models, a type of AI that can generate human-like responses to questions and explanations familiar to users of ChatGPT. The research team adapted it for matching patients to thousands of possible clinical trials in a data-efficient and transparent way. While earlier studies have shown the potential for using this type of AI for answering clinical questions, designing clinical trials, and retrieving initial lists of potential trials, TrialGPT is the first end-to-end solution, generating a list of potential trials before more precisely matching and ranking them. The team’s preliminary testing of this tool suggests TrialGPT can achieve a high degree of accuracy while cutting the time required of clinicians for screening patients.

These advances come from researchers at the NIH National Library of Medicine (NLM) and National Cancer Institute (NCI), Bethesda, MD, and collaborators, led by NLM Senior Investigator Zhiyong Lu. The study was co-authored by researchers from Albert Einstein College of Medicine, New York City; University of Pittsburgh; University of Illinois Urbana-Champaign; and University of Maryland, College Park.

Here’s how TrialGPT works: The first step is TrialGPT-Retrieval, in which the tool processes a summary of patient information containing relevant medical and demographic details such as age, sex, and major conditions or symptoms. The algorithm uses any available notes to filter out irrelevant trials from ClinicalTrials.gov and return an initial list of candidate clinical trials for which a given patient may prove eligible.

Next, TrialGPT-Matching produces an explanation about how the individual in question meets each trial’s eligibility criteria to predict more precisely whether they may indeed be a candidate. The final step is TrialGPT-Ranking, which produces an annotated list of potential clinical trials—ranked by their relevance and an individual’s likely eligibility—that clinicians can use to define a top 10 list of clinical trial opportunities for discussion with the patient.

To find out how well it might work in practice, the researchers put it to the test with three publicly available datasets representing 183 “synthetic” patients and more than 75,000 trial eligibility annotations. Synthetic patient data are physician-created vignettes that closely mimic real medical data for research purposes without containing anyone’s personal information. The team found that TrialGPT could use the data to readily retrieve 90% of relevant clinical trials. Comparisons of TrialGPT-Matching to matching by human experts on more than 1,000 patient-eligibility criterion pairs found that the tool could accurately explain the relevance of each criterion to a patient, find relevant sentences, and predict eligibility with near human accuracy. The TrialGPT-Ranking function also produced results similar to that of experts.

In a pilot user study conducted at NCI, the researchers compared patient-trial evaluations based on short summaries about six patients made by one medical expert with TrialGPT and another who made the same evaluation manually without TrialGPT. Both experts conducted evaluations with and without AI to account for any differences in their speed or skill. The study found that clinicians using TrialGPT could generate similarly accurate lists of trial options in 40% less time.

More study is needed to assess TrialGPT’s practical application in real-world settings across diverse groups of patients. But these findings already show the remarkable potential of AI technology for connecting patients to relevant trial opportunities, with tremendous potential for speeding trial recruitment and treatment advances while giving clinicians more time for other tasks only humans can do, including caring for their patients.

The research team was selected for a 2024 Director’s Challenge Innovation Award to further assess TrialGPT in real-world clinical settings while addressing disparities in clinical trial enrollment. The goal is to make clinical trial recruitment both more efficient and more effective in the future while helping to reduce barriers to participation for populations that have traditionally been underrepresented in clinical research.”

Monica Bertagnolli is the Director of the National Institute of Health (NIH), USA. She is the President of Alliance for Clinical Trials In Oncology Foundation. She previously served as the Richard E. Wilson Professor of Surgery in the field of surgical oncology at Harvard Medical School, a surgeon at Brigham and Women’s Hospital, and a member of the Gastrointestinal Cancer Treatment and Sarcoma Centers at Dana-Farber Cancer Institute.

Dr. Bertagnolli is the Founding Chair of the minimal Common Oncology Data Elements (mCODE) executive committee and has held multiple positions nationally, including being Past President and Chair of the Board of Directors of the American Society of Clinical Oncology. In 2021, she was elected to the National Academy of Medicine, having previously served on the National Academies National Cancer Policy Forum.

For more Updates, Follow OncoDaily.