Unique Research · Healthcare · January 9, 2025
A new study led by researchers at Emory University’s Winship Cancer Institute and the University of Pennsylvania’s Abramson Cancer Center shows that a first-of-its-kind AI platform can help clinicians and patients assess whether — and how much — an individual patient may benefit from a therapy being tested in a clinical trial. The platform supports informed treatment decisions, expectations for novel therapies, and future care planning.
Published in Nature Medicine, the work was led by Ravi B. Parikh, MD, MPP, medical director of Winship’s Data and Technology Applications Shared Resource and associate professor of hematology and medical oncology at Emory — who develops and integrates AI applications to improve cancer care. Co–senior author is Qi Long, PhD, professor of biostatistics and of computer and information science at Penn, director of Penn’s Center for Cancer Data Science, and associate director of the Quantitative Data Science Division at Abramson. First author is Xavier Orcutt, MD, a trainee in Parikh’s lab. Other authors include Kan Chen, a PhD student trained in Long’s lab, and Ronac Mamtani, associate professor of medicine at Penn.
Parikh and colleagues developed TrialTranslator, a machine-learning framework that “translates” clinical-trial results into real-world populations. By using real-world data to simulate 11 landmark cancer trials, they could reproduce actual trial outcomes and identify which patient subgroups may respond well to trial therapies — and which may not.
“We hope this AI platform gives doctors and patients a framework to judge whether trial results apply to an individual,” Parikh said. “It may also help researchers identify subgroups for whom new treatments do not work — and spur new trials for those high-risk groups.”
Long added: “Our work shows the huge potential of using AI/ML to mine rich, complex real-world data to advance precision medicine as far as possible.”
Limited generalizability of trial results
Parikh explained that trials of potential new therapies are limited because fewer than 10% of cancer patients enroll — so trials often do not represent all patients with that cancer. Even when a trial shows a new strategy beats standard care, “many patients still do not benefit from the new treatment,” he said.
“This framework and our open-source calculator will help patients and doctors judge whether Phase III results apply to an individual cancer patient,” he said, adding that the study also provides a platform to analyze generalizability of other randomized trials — including negative ones — in the real world.
Methods
Using Flatiron Health’s national electronic health record (EHR) database, the team simulated 11 landmark randomized controlled trials comparing treatments for the four most common advanced solid malignancies in the U.S.: advanced non–small cell lung cancer, metastatic breast cancer, metastatic prostate cancer, and metastatic colorectal cancer.
Findings
Patients with low- and intermediate-risk phenotypes (machine-learning assessments of underlying prognosis) had survival times and treatment-related survival benefit similar to those in the RCTs. By contrast, high-risk phenotype patients had significantly shorter survival and less treatment-related benefit than in the trials.
The results suggest machine learning can identify groups for whom RCT findings generalize poorly to real-world patients — meaning “real-world prognosis may be more heterogeneous than among RCT participants.”
Why it matters
The team concluded that “patient prognosis, rather than eligibility criteria, better predicts survival and treatment benefit.” They recommend that prospective trials “consider more sophisticated methods to assess prognosis at enrollment rather than relying only on strict eligibility criteria.”
They also cite ASCO and Friends of Cancer Research recommendations to improve representation of high-risk subgroups in RCTs, “given that treatment effects in these individuals may differ from other participants.”
On AI’s role, Parikh said: “With appropriate oversight and evidence, we will soon see more AI-based biomarkers that analyze pathology, radiology, or EHR information to help predict whether patients will respond to certain therapies, diagnose cancer earlier, or bring better outcomes for our patients.”
— END —