Clinical Trial Matching
Educational note: this page reflects public evidence last checked on 2026-06-04. Trial matching tools are decision-support and workflow software; they do not determine eligibility, enrollment, treatment, or trial safety by themselves.
TL;DR
Matching a cancer patient to a clinical trial is a hard data problem because trial criteria are specific, often nested, and partly written as free text, while patient state is distributed across pathology, genomics, imaging, labs, treatment history, performance status, and clinician notes. NLP, LLMs, rule engines, and knowledge graphs can help produce a candidate trial list, but the clinically safe endpoint is human review by trial staff and the treating team. A matching system should optimize recall, provenance, and explainability before it optimizes "automation." Sources: [1], [2], [3]
1. The problem space
Adult cancer trial enrollment has historically been low, often described around the 2-5% range for treatment trials depending on setting, disease, and data source. The barrier is not only patient willingness. It is also infrastructure: trial availability, eligibility complexity, clinician awareness, site capacity, insurance/logistics, molecular testing, and rapid disease progression. Sources: [1], [4]
A matcher has to connect:
- Trial facts - NCT ID, status, phase, sites, arms, interventions, inclusion/exclusion criteria, biomarkers, prior therapies, labs, age, organ function, measurable disease, and washout rules.
- Patient facts - diagnosis, stage, histology, molecular alterations, prior lines, current treatment, ECOG, labs, organ function, comorbidities, location, dates, and preferences.
- Evidence provenance - where each patient fact came from, when it was measured, and whether it is current enough for screening.
2. Parsing trial criteria
ClinicalTrials.gov API v2 exposes structured fields such as NCT ID, title, status, phase, interventions, age groups, and eligibility text. The hard part is turning eligibility prose into computable logic without losing nuance.
Examples:
- "EGFR exon 19 deletion or L858R" is a molecular inclusion criterion.
- "No untreated active brain metastases" is not the same as "no brain metastases ever."
- "Progressed after platinum" depends on dates, treatment intent, and disease context.
Extraction can use regular expressions, ontologies, named-entity recognition, LLMs, or hybrid neuro-symbolic systems. For clinical use, extracted rules need traceable links back to original text. Sources: [2], [3], [5]
3. Parsing patient data
The patient side is usually messier than the trial side:
- Genomics - VCF/MAF/PDF reports, gene panels, MSI/TMB/HRD, ctDNA, tissue source, assay limits.
- Pathology - histology, grade, receptor status, PD-L1 scoring, IHC/FISH results.
- Treatment history - line of therapy, exposure, progression, toxicity, dates.
- Clinical state - ECOG, organ function, measurable disease, active infection, brain metastases, pregnancy, comorbidities.
- Logistics - country, site distance, travel capacity, insurance, language, and caregiver constraints.
No model should silently infer missing facts. A safe matcher should say "unknown, needs review" rather than converting absent data into eligibility.
4. Matching engine behavior
| Output | Meaning | Safe use |
|---|---|---|
| Potential match | Key disease/biomarker constraints appear compatible | Send to human prescreening |
| Rule-out | A clear exclusion is documented | Keep evidence trail and allow override if data are stale |
| Unknown | Required criterion not found or not current | Request missing data or manual review |
| Ineligible | Human-reviewed final determination | Belongs to trial staff/clinician workflow, not the algorithm alone |
The implementation should favor explainable criterion-level results:
NCT12345678
include: ER+/HER2- breast cancer -> matched from pathology report, 2026-03-02
include: ESR1 mutation -> unknown, no ctDNA/tissue report found
exclude: active untreated CNS metastases -> unknown, last brain MRI not available5. Failure modes
- Treating an LLM-generated match as enrollment eligibility.
- Using stale tumor genomics after clonal evolution or new therapy.
- Missing exclusion criteria hidden in free text.
- Ignoring trial status, site activation, cohort closure, or slot availability.
- Overfitting a matcher to one hospital's EHR vocabulary.
- Failing to distinguish "no evidence found" from "evidence of absence."
- Not logging why a trial was hidden from the user.
6. What technologists can build
- ClinicalTrials.gov API harvesters with status and last-update timestamps.
- Eligibility parsers that preserve source text and criterion IDs.
- FHIR/mCODE/OMOP adapters for patient facts.
- Molecular-rule engines that normalize genes, variants, fusions, and biomarkers.
- Human-in-the-loop review queues for coordinators and molecular tumor boards.
- Audit logs for every match, rule-out, unknown, and manual override.
See also
References
- Unger JM, Vaidya R, Hershman DL, Minasian LM, Fleury ME. Systematic review and meta-analysis of the magnitude of structural, clinical, and physician and patient barriers to cancer clinical trial participation. J Natl Cancer Inst 2019;111:245-255. PMID 30856372. https://doi.org/10.1093/jnci/djy221
- National Library of Medicine. ClinicalTrials.gov API v2 documentation. https://clinicaltrials.gov/data-about-studies/learn-about-api
- Ni Y, Wright J, Perentesis J, et al. Increasing the efficiency of trial-patient matching: automated clinical trial eligibility pre-screening for pediatric oncology patients. BMC Med Inform Decis Mak 2015;15:28. PMID 25881112. https://doi.org/10.1186/s12911-015-0149-3
- Unger JM, Hershman DL, Fleury ME, Vaidya R. Association of patient comorbid conditions with cancer clinical trial participation. JAMA Oncol 2019;5:326-333. PMID 30589950. https://doi.org/10.1001/jamaoncol.2018.5953
- PubMed. Automated classification of eligibility criteria in clinical trials to facilitate patient-trial matching for specific patient populations. PMID 28339690. https://pubmed.ncbi.nlm.nih.gov/28339690/