Cancer3.AIAI in OncologyAI Models › TrialGPT
Tool API / hosted only Research use only clinical-LLM

TrialGPT

Three-stage LLM framework (retrieval, criterion-level matching, ranking) that matches a patient summary to clinical trials from ClinicalTrials.gov; in the paper it cut clinician screening time by more than 40% in a pilot user study.

At a glance

DeveloperNIH National Library of Medicine (with National Cancer Institute collaborators)
Released2024-11-18
Licencepublic domain (NIH code) — depends on the backing LLM
AvailabilityAPI / hosted only
KindTool
Regulatory statusResearch use only

What it does

TrialGPT is a large-language-model framework developed at the US National Institutes of Health that matches patient records to clinical trials. It analyses eligibility criterion by criterion, explains its reasoning in natural language, and ranks candidate trials for each patient.

Notable uses in oncology

In a Nature Communications (2024) evaluation, its criterion-level judgments reached 87.3% accuracy — close to physician experts — and a pilot user study showed a 42.6% reduction in screening time. The system was piloted on oncology trials managed at or related to the National Cancer Institute.

Tasks, data types and cancers

Data typeClinical text
Clinical settingClinical trials
CancerPan-cancer
Inputpatient summary (free text) + trial eligibility criteria
Outputcriterion-level eligibility labels with explanations; ranked trials

Architecture

FamilyLLM pipeline (retrieval + criterion-level reasoning)
BackboneGPT-4 class models via API in the paper; open models can be substituted

Training data

None. TrialGPT performs zero-shot patient-to-trial matching by prompting a general-purpose large language model through the OpenAI or Azure API (the repository documents runs with GPT-4 / GPT-4-turbo); no model is trained or fine-tuned, so behaviour depends on the LLM that is plugged in.

Linked datasets

Evaluation

Benchmark / datasetMetricValueExternal validationSource
SIGIR / TREC clinical-trial cohorts criterion-level accuracy vs. experts 87.3% (paper) yes Source

How to run

LibraryPython (repo) + LLM API

The code is open, but results depend on the LLM you plug in; patient data sent to a commercial API needs a compliant deployment.

Regulatory status and intended use

Regulatory statusResearch use only
Intended useScreening aid for trial coordinators; eligibility is confirmed by the study team.

Source →

Regulatory status is quoted from the source linked above and can change. Research-use-only models must not be used for clinical decisions.

Limitations and bias

  • LLM errors on nuanced criteria (lab thresholds, prior therapies); human review required.
  • Privacy: patient text leaves the institution when using hosted models.

Sources

  1. Jin Q et al. Matching patients to clinical trials with large language models. Nat Commun 2024
  2. GitHub — ncbi-nlp/TrialGPT

This page is educational — it is not medical advice and does not replace consultation with an oncologist. Diagnostic and treatment decisions are made solely by specialist physicians.