A recent study reveals that a pioneering artificial intelligence (AI) platform can assist healthcare professionals and patients in evaluating the potential benefits of specific therapies underway in clinical trials. This innovative AI tool aims to facilitate informed treatment choices, clarify the expected advantages of new therapies, and aid in future care planning.
A recent study from the Winship Cancer Institute at Emory University and the Abramson Cancer Center at the University of Pennsylvania shows that a groundbreaking AI platform can assist both clinicians and patients in understanding how much a specific therapy being tested in a clinical trial might benefit an individual patient. This platform aims to support informed decisions regarding treatments, clarify the potential benefits of new therapies, and help plan future care.
The findings, published in Nature Medicine, were led by Dr. Ravi B. Parikh, a board-certified medical oncologist and medical director of the Data and Technology Applications Shared Resource at Winship Cancer Institute. He is also an associate professor at the Emory University School of Medicine, where he works on integrating AI applications to enhance cancer patient care. Dr. Qi Long, a professor of Biostatistics, Computer Science, and founding director of the Center for Cancer Data Science at the University of Pennsylvania, was a co-senior author. The principal author of the study was Dr. Xavier Orcutt, a researcher in Parikh’s lab, along with contributions from Kan Chen, a PhD student in Long’s lab, and Ronac Mamtani, an associate professor of medicine at the University of Pennsylvania.
Parikh and his team developed TrialTranslator, a machine learning framework that “translates” clinical trial outcomes to reflect real-world patient populations. By replicating 11 major cancer clinical trials using actual patient data, they identified specific groups of patients who might respond positively to treatments in clinical trials, as well as those who may not.
“We hope this AI platform will serve as a tool for doctors and patients to determine the applicability of clinical trial results to individual cases,” says Parikh. “Additionally, this research could help identify patient subgroups for whom new treatments might be ineffective, prompting the development of clinical trials targeted at these higher-risk groups.”
“Our findings highlight the vast potential of utilizing AI and machine learning to leverage complex real-world data and push the boundaries of precision medicine,” Long adds.
Limitations of Clinical Trial Results
Parikh notes that clinical trials for potential new therapies are often restricted, with less than 10% of cancer patients participating. This leads to a disconnect between trial results and the wider cancer patient population. Even if a trial indicates that a new treatment shows better outcomes compared to the standard care, “there are many patients for whom the new treatment does not yield benefits,” he explains.
“This framework and our open-source calculators will empower patients and physicians to evaluate whether the results from phase III trials apply to individual cancer patients,” he says, adding that “this study opens avenues to analyze the broader applicability of various randomized trials, including those with negative outcomes.”
Analysis Methodology
Parikh and his team analyzed a national electronic health records (EHR) database provided by Flatiron Health to emulate 11 pivotal randomized controlled trials focused on standard cancer treatments for the four most common advanced solid tumors in the U.S.: advanced non-small cell lung cancer, metastatic breast cancer, metastatic prostate cancer, and metastatic colorectal cancer.
Key Findings
Their analysis indicated that patients classified with low- and medium-risk phenotypes—traits defined by machine learning to evaluate patient prognosis—exhibited survival rates and treatment benefits that mirrored those observed in randomized controlled trials. Conversely, patients identified with high-risk phenotypes experienced significantly lower survival times and treatment benefits compared to trial participants.
These outcomes imply that machine learning can pinpoint real-world patient groups where the results of randomized trials may not apply. They emphasize that “real-world patients often present with more varied prognoses than those participating in randomized trials.”
Significance of the Research
The research team concludes that their findings suggest that a patient’s prognosis is a more reliable predictor of survival and treatment advantage than mere eligibility criteria. They urge that future trials should embrace more sophisticated methods of assessing patients’ prognoses upon entry, rather than relying strictly on eligibility requirements.
Furthermore, they highlight recommendations from the American Society of Clinical Oncology and Friends of Cancer Research advocating for improved diversity of high-risk groups in randomized trials, given that treatment responses might vary from other participants.
Regarding the advancement of AI in studies like this one, Parikh mentions, “With proper regulation and evidence, we are likely to see a growing use of AI-derived biomarkers that can analyze pathology, radiology, or electronic health records, helping forecast patient responses to therapies, diagnose cancers earlier, or lead to enhanced prognoses for patients.”
This research received support from various grants from the National Institute of Health: K08CA263541, P30CA016520, and U01CA274576.
