Inside the shift: How oncologists are using AI for trial matching
Industry Buzz
They [AI tools] crumble on the 80% of EHR data that is in the clinical notes, imaging reports, and physician shorthand… If a physician writes 'arrhythmia, currently controlled,' that means something very different to me as a trained clinician compared to what it means to a keyword-parsing algorithm.
—Eleonora Fedonenko, MD
The biggest gain that AI brings in is not speed. It's the fact that the eligible patients get flagged before the physician even thinks to look. And when that matching is run alongside structured imaging data, accuracy is improved in ways manual review simply cannot replicate.
—Jason Schroder, DO
Clinical trial enrollment in adult oncology remains low, with only 7.1% of patients enrolling nationally.[] Fewer than 17% percent of eligible patients enroll in studies across most U.S. centers.[] Screening still depends on manual chart review, protocol cross-checking, and fragmented data.[]
However, now large health systems and cancer centers are starting to deploy AI tools to automate trial matching directly within the EHR. Early data and clinician feedback show measurable gains in efficiency and identification of eligible patients.[]
Mount Sinai: Systemwide deployment of oncology-specific AI
In January 2026, Mount Sinai launched PRISM, an AI-powered clinical trial matching platform integrated across its health system.[]
The system uses an oncology-trained large language model (LLM), OncoLLM, to analyze structured and unstructured EHR data, including pathology reports and clinical notes. The platform addresses a known bottleneck, low patient participation in oncology trials, in part due to a lack of awareness among both physicians and patients.
Karyn Goodman, MD, vice chair of clinical research at Mount Sinai, described the impact: “By deploying an AI platform trained specifically for oncology, we can identify trial opportunities earlier, more consistently, and more equitably, allowing clinicians to focus on meaningful conversations with patients rather than manual chart review.”[]
Mount Sinai reports that PRISM enables real-time screening across multiple hospitals, extending access beyond academic flagship sites.
MSK-MATCH: High-accuracy eligibility screening
At Memorial Sloan Kettering, a separate AI system, MSK-MATCH, focuses on eligibility prescreening. In a retrospective study across breast cancer trials, the system achieved 98.7% accuracy for patient-level eligibility classification, and for cases requiring manual review, it reduced screening time from 20 minutes to 43 seconds per patient.[]
The model processes free-text clinical documents and generates explainable eligibility decisions. About 62% of cases required no manual review.
This type of triage model can reduce case recruitment burden since AI handles high-confidence matches, while complex cases remain under physician oversight.
TrialGPT and LLM approaches
Academic groups are testing generalizable LLM-based systems. TrialGPT, developed with NIH support, retrieves up to 90% of relevant trials from large databases and shows high agreement with physician assessments.[]
Sandip Patel, MD, professor in the Department of Medicine at the University of California San Diego Health, and co-author of an article on machine learning innovations,[] noted the clinical relevance: “There are many friction points in clinical trials, starting at accrual…. It’s important for us to have tools that can help screen patients for these clinical trials.”[]
These tools move beyond structured data matching. They interpret inclusion and exclusion criteria in natural language and apply them to full patient records.
Related: IMRT, AI, and fewer cases: Is radiation oncology facing oversupply?Data integration and real-world matching performance
AI matching depends on the integration of fragmented data sources.[]
Board-certified internist Eleonora Fedonenko, MD, states, “Current AI tools are good on structured fields such as lab values, ICD codes, and discrete checkboxes.”
But she adds, “They [AI tools] crumble on the 80% of EHR data that is in the clinical notes, imaging reports, and physician shorthand… If a physician writes "arrhythmia, currently controlled," that means something very different to me as a trained clinician compared to what it means to a keyword-parsing algorithm. The difference between what I state in my notes and what a keyword-parsing algorithm interprets in my notes is not going to narrow as the screening of the notes becomes more efficient.”
To overcome such limitations, newer architectures combine structured data with the parsing of unstructured text. This improves sensitivity to complex eligibility criteria such as prior therapies, biomarker status, and timing of disease progression.
Systems like MatchMiner aggregate genomic data, clinical notes, and registry information. MatchMiner was shown to generate 7 to 10 trial matches per patient in multidisciplinary tumor boards.[]
TrialMatchAI, an end-to-end system, demonstrated over 90% accuracy in eligibility classification in validation datasets.[]
Outside academic centers, interest is growing in how AI may improve access.
Jason Schroder, DO, a board-certified anesthesiologist treating patients with cancer, lung, and cardiac disease, says, “After experiencing years of patients coming in for surgery when they were already past the point where a trial helping them would have been an option, I've developed a real frustration with the way the medical system takes so long to get people connected to options. Traditionally, trial matching involved a physician running through their head a few patients they happened to recall when reviewing a new protocol. That's a memory test, not a system, and patients pay the price for it."
He further adds, “The biggest gain that AI brings in is not speed. It's the fact that the eligible patients get flagged before the physician even thinks to look. And when that matching is run alongside structured imaging data, accuracy is improved in ways manual review simply cannot replicate.”
What changes in clinical workflow
Across these platforms, three consistent changes emerge.
Earlier identification of eligible patients. AI runs continuously in the background rather than relying on manual review at clinic visits.
Reduced administrative burden. Chart review shifts from primary screening to validation of AI-flagged candidates.
Broader access. Systemwide deployment allows patients in community settings to be screened against trials traditionally limited to academic centers.
Limitations remain
Model bias, data quality, and interpretability require ongoing oversight. Most centers maintain a human-in-the-loop model for final eligibility decisions.
As Dr. Schroder states, “Rural patients, uninsured patients, and people that never entered a major health system's EHR simply don't exist in the data pools that AI learns from. You can't beat a patient you didn't see. AI optimizes the population that already was in front of you, and it does it well. But closing the access gap requires getting more people into early detection pipelines before a diagnosis makes the issue forcing, and that's the work that no matching algorithm does on its own.”
Dr. Fedonenko puts it this way: “On diversity and enrollment, AI can provide greater opportunities but cannot resolve the upstream issues, such as lack of trust, transportation issues, and the fact that most trial sites are far from where underrepresented communities are located. Simply speeding up matching will make these disparities more apparent without closing them.”
Still, AI-driven trial matching has moved beyond pilot studies. Health systems are integrating these tools into routine oncology workflows, with measurable effects on screening efficiency and trial access.
Related: Biomarker testing: The gap between guidelines and what actually happens in the clinic