Truace
Subscribe
44/100

Emerging evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

AI matched 24 test cancer patients to 746 trials in under a second, agreeing with doctors 95.8% of the time

Study title: Quantifying cancer clinical trial eligibility using artificial Intelligence-Based matching

Evidence type
Benchmark or lab evaluation
Sample
n = 24
Sources
1 sourcePeer-reviewed
Published
2 min read
Truvace record
TRV-2026-2470cite or check it

Study title: Quantifying cancer clinical trial eligibility using artificial Intelligence-Based matching

Researchers built a platform in which a large language model turns the free-text eligibility criteria of cancer trials into SQL code, so matching can run as explicit rules without further AI inference. Using 24 synthetic patient vignettes written by medical oncologists, matching across 746 trials took a median of 0.459 seconds per patient.

Read the study abstract

Purpose: To develop and validate an artificial intelligence-enabled platform that combines efficient screening with explicit eligibility logic by converting unstructured cancer trial criteria into reusable structured queries, and illustrates potential eligibility using synthetic patients. Methods: We downloaded actively recruiting US interventional treatment trials for advanced/metastatic breast cancer, colon cancer, and non-small cell lung cancer from ClinicalTrials.gov. Medical oncologists created 24 synthetic…

53Gain side

How the two sides scoreGain scores higher

46Problem side

The gain

In a validation using 24 synthetic patients, an AI tool matched each patient against 746 cancer trials in a median 0.459 seconds and agreed with physician consensus on 230 of 240 patient-trial pairs (95.8%).

The problem

The automated matching missed some eligible pairs, with sensitivity of 83.3% against physician consensus, and was tested only on synthetic vignettes that left out laboratory results and comorbidities.

A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.

What the scores mean

G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.

P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.

In brief

Against physician consensus on ten trials, the tool agreed on 230 of 240 patient-trial pairs (95.8%), with 100% specificity and 83.3% sensitivity. That means it rarely flagged ineligible patients as eligible, but it did miss some true matches.

Potentially eligible trials ranged from 10 to 258 per synthetic patient without geographic limits. Counts fell with worse performance status and with geographic restriction, and were relatively stable across ages 30 to 90. The work used synthetic patients, so performance on real patient records is not yet known.

Main points

  1. The team downloaded actively recruiting US interventional treatment trials for advanced or metastatic breast, colon and non-small cell lung cancer from ClinicalTrials.gov; 746 of 833 candidates met inclusion criteria.
  2. Medical oncologists wrote 24 synthetic patient vignettes; the language model converted these into structured records.
  3. Matching all 746 trials took a median of 0.459 seconds per patient, excluding data preparation.
  4. Against physician consensus, automated matching agreed on 230 of 240 pairs (95.8%), with sensitivity of 83.3% and specificity of 100%.
  5. Potentially eligible trials per synthetic patient ranged from 10 to 258 with no geographic restriction.
  6. Later-phase, randomized and molecularly selective trials generally matched fewer vignettes.
  7. Physicians edited the generated eligibility code for 30 trials as part of validation.

The full story

The platform separates the language model's work from the matching itself. The model translates trial criteria into SQL and patient information into structured records once, and matching then runs as rule-based queries, which keeps the logic explicit and reusable across patients.

Validation had two parts: physicians edited generated eligibility code for 30 trials, and two physicians independently assessed eligibility across ten trials, followed by consensus adjudication. Against that consensus, the tool's 100% specificity and 83.3% sensitivity point to missed matches rather than false ones.

In the descriptive analysis, the number of potentially eligible trials fell as performance status worsened and as geographic restrictions were applied, but barely changed across ages 30 to 90. Later-phase, randomized and molecularly selective trials matched fewer of the vignettes.

What this doesn’t fix

Matching was evaluated only on 24 synthetic patient vignettes that did not include laboratory results or comorbidities, and the authors say missed matches mean further validation is needed.

Talk it overFour questions for a class, a team or the dinner table

  1. This record found a gain and a problem in the same place. Which matters more to you, and who might answer differently?

    The gain In a validation using 24 synthetic patients, an AI tool matched each patient against 746 cancer trials in a median 0.459 seconds and agreed with physician consensus on 230 of 240 patient-trial pairs (95.8%).

    The problem The automated matching missed some eligible pairs, with sensitivity of 83.3% against physician consensus, and was tested only on synthetic vignettes that left out laboratory results and comorbidities.

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a benchmark or lab test with a sample of 24, from JNCI Cancer Spectrum.

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat Matching was evaluated only on 24 synthetic patient vignettes that did not include laboratory results or comorbidities, and the authors say missed matches mean further validation is needed.

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

The questions are the same on every record. The details under them are taken from this record as written.

Sources

1 on file
  1. JNCI Cancer Spectrumdoi:10.1093/jncics/pkag120
    Peer-reviewed2026-11-10

The debate

More from Health