AI Trial Matching Just Solved Oncology’s Worst Bottleneck

The patient matching problem nobody could solve

Cancer clinical trials have a recruitment crisis that the field has talked about for two decades and never fixed. Only 5 to 7 percent of adult cancer patients in the United States ever enroll in a trial, even though the number of active oncology trials has nearly doubled since 2015. The reason is not patient unwillingness. It is matching. Trials carry 30 to 80 eligibility criteria each, written in dense oncology jargon, and a staff coordinator has to read every line of a candidate’s chart against every line of every protocol. At a large cancer center running 400 active trials, that is mathematically impossible to do well.

By April 2026, three institutions have shown that AI can finally do this work at scale, and the operational results are not incremental. Mount Sinai, City of Hope, and the National Institutes of Health have each rolled out production systems that read a patient’s electronic record, screen it against active trial criteria, and surface ranked matches to the oncologist while the patient is still in the room. The accuracy numbers and time savings are large enough that the FDA, sponsors, and major academic centers are starting to treat this as the new baseline.

What Mount Sinai’s PRISM platform actually does

Mount Sinai launched PRISM in January 2026 inside the Tisch Cancer Center. The platform sits on top of an oncology-trained large language model called OncoLLM, which the team built specifically to handle the unstructured pathology reports, imaging notes, and molecular profiling results that make up most of a cancer patient’s chart. Structured EHR fields capture maybe 30 percent of what an oncologist needs to know to assess trial eligibility. The other 70 percent is in free text, and that is where the older rule-based matching tools failed.

PRISM ingests both. It reads the latest biopsy report, the most recent imaging, the molecular panel from the tumor, prior lines of therapy, comorbidities, and lab values, then runs that profile against the structured eligibility criteria of every open trial in the Mount Sinai system. The output is a ranked list with a citation back to the specific chart line that triggered the match. Oncologists at Mount Sinai have reported that what used to be a four-hour weekly trial review meeting is now a fifteen-minute confirmation step.

The platform was built to expand access, not just speed. Mount Sinai’s catchment area includes neighborhoods where trial enrollment has historically been a fraction of the rate at suburban academic centers. PRISM was deliberately deployed across the full Mount Sinai Health System, not just the flagship Tisch site, so that patients seen at community-affiliated clinics get the same trial screening as patients seen on the main campus.

City of Hope’s HopeLLM is matching against 200+ trials in real time

City of Hope built a different system around the same idea. HopeLLM is an internally trained foundation model fine-tuned on the institution’s own oncology corpus, which includes decades of pathology, treatment, and outcomes data. The trial matching module went live in 2025 and by early 2026 was supporting feasibility checks for more than 200 active trials across all City of Hope locations.

The operational metric that matters here is screening speed. Before HopeLLM, matching a single patient against the active trial portfolio could take a coordinator several hours of chart review per protocol of interest. After deployment, the same query returns a ranked match list in under a minute, with the relevant eligibility lines pre-flagged. City of Hope reports that the system is now embedded in the standard oncology consult workflow, meaning trial options are presented at the same visit where treatment decisions are made rather than weeks later.

That timing change matters because patients with aggressive cancers often cannot wait. A delay of two weeks to identify a trial is a delay long enough that the patient may have already started a different therapy, which then disqualifies them from trials that exclude prior treatment. Real-time matching closes that window.

TrialGPT and the NIH baseline

The federal version of this work is TrialGPT, developed jointly by the NIH National Library of Medicine and the National Cancer Institute. TrialGPT is not aimed at a single hospital. It is a public-facing tool that takes a patient summary or eligibility query and returns relevant trials from ClinicalTrials.gov along with a plain-language explanation of why each trial matches.

The published results from the TrialGPT validation study are the cleanest benchmark in the field. In a head-to-head test against expert clinicians doing manual matching, TrialGPT-assisted screening took 40 percent less time while maintaining equivalent accuracy. That is the single most-cited figure in clinical trial AI conversations right now, because it pins down the productivity gain in a clean apples-to-apples comparison rather than a marketing pitch.

A separate prospective study at a metastatic colorectal cancer center used a Common Data Model-based AI matching system against the same trial portfolio and reported 94 percent retrospective accuracy and 88 percent prospective accuracy. The chart review workload dropped tenfold, and screening time dropped 41 percent. That study is one of the first to publish prospective deployment data with a real patient cohort, which is the bar that regulatory and reimbursement reviewers care about.

The economics finally work

Patient recruitment is the single largest cost line in most clinical trials. Industry estimates put it at 30 to 40 percent of total trial budget, and the average cost to recruit one oncology trial participant ranges from $6,000 to $15,000 depending on the indication. Trials that miss enrollment targets are the ones that blow timelines and burn through sponsor budgets, and the failure rate of trials that fail to enroll on schedule is high enough to scare off investment in some rare cancer indications entirely.

Across reported deployments in 2025 and early 2026, AI-driven matching tools improved enrollment rates by roughly 65 percent and cut recruitment costs by roughly 40 percent. Those numbers are aggregated across multiple platforms — Deep 6 AI, IQVIA, Phesi, Lindus Health, BEKhealth — but they are consistent enough across vendors that the ROI argument is no longer speculative. Sponsors paying $20 million for a Phase 2 oncology trial are looking at $5 to $8 million in recruitment savings if the matching tool performs at the published median.

That math is what is moving the buying decision. Mount Sinai, City of Hope, MD Anderson, Memorial Sloan Kettering, and the Mayo Clinic have all either built or contracted AI trial matching platforms by Q1 2026. The vendor field has consolidated around a handful of players that can integrate with Epic and Cerner at the EHR layer and with Veeva or Medidata at the trial management layer.

Where this connects to the rest of oncology AI

Trial matching does not exist in isolation. It is the front door to the precision oncology stack that has been quietly assembling itself across the last three years. A patient identified by a matching system is then often enrolled in a trial that uses an AI-designed therapeutic, monitored with AI-read pathology, and tracked with digital twin simulation of expected response.

We covered the upstream side of this in our piece on AI-designed drugs clearing Phase 2, and the readout side in pathology AI catching cancers humans miss. The trial matching layer is what connects them. Without it, a patient eligible for an AI-discovered combination therapy never learns the trial exists, and the science sits stranded behind a recruitment failure.

The same dynamic applies to the rare disease diagnostic odyssey, where the bottleneck is not just diagnosis but matching the diagnosed patient to whatever sliver of a trial actually accepts that genotype. Generic LLMs cannot do this. Domain-trained models like OncoLLM and HopeLLM can, because they have learned the genotype-to-eligibility mapping from real institutional data.

What the regulators are doing

The FDA has not issued a formal guidance on AI-assisted trial matching, but it has signaled tolerance. The 2025 update to the agency’s enrichment guidance referenced the use of computational tools for cohort identification, and FDA reviewers have accepted enrollment data sourced from AI-matched populations in several recent oncology submissions. The European Medicines Agency has been more cautious but is running pilots with sponsors that want to use AI matching to support EU enrollment in trials that originated in the US.

What the regulators care about most is bias. A matching system that systematically misses certain patient populations is worse than no matching system at all, because it bakes the disparity into the official recruitment record. The published validations of TrialGPT, OncoLLM, and HopeLLM all include subgroup performance breakdowns by race, age, and insurance status, and that disclosure is becoming the de facto standard for what regulators expect.

Where the technology breaks

It is not all working. The biggest current limitation is data standardization across health systems. A patient referred from a community oncology practice to a major academic center arrives with a chart in a different EHR with different note structures, and the matching model has to reconcile that without losing the eligibility-relevant signal. Some platforms handle this well, others fail silently and produce false negatives that no one notices because the patient is simply never told about the trial.

Eligibility criteria themselves are also still written by humans for humans. Trial sponsors writing protocols in 2026 are starting to coordinate with matching platform vendors during protocol design so the criteria are machine-readable from day one, but the back catalog of older trials remains uneven. Some criteria contain subjective language — “adequate organ function,” “appropriate prior therapy” — that requires human judgment regardless of how good the AI is.

What changes from here

Two things to watch through 2026. First, predictive enrollment forecasting is moving from match-the-patient to forecast-which-patients-will-arrive. The same models that score eligibility can be turned around to project, given a hospital’s catchment population, how many eligible patients are likely to walk in over the next 90 days. That changes how sponsors choose sites and how sites choose which trials to accept.

Second, real-time alerting is showing up in production. A patient whose latest pathology report contains a specific mutation can trigger an immediate notification to their oncologist that a newly opened trial accepts that mutation, before the next clinic visit. This is the use case that turns trial matching from a periodic batch process into a continuous monitoring layer over the entire patient population. Mount Sinai, City of Hope, and a handful of other centers are already running it.

The bottleneck that defined oncology recruitment for thirty years is in the process of moving. The trials that get the AI matching layer are seeing enrollment rates that the field had previously written off as unattainable, and the centers without it are starting to look like the laggards rather than the norm.