AI Trial Matching: screening cancer patients for clinical trials

An AI tool for oncology: it identifies clinical trials that a patient is eligible to join, and finds patients who can be enrolled in a trial.

Problem:

• Trial coordinators and oncologists manually compare long lists of criteria against the free-text content of Czech medical reports.
• As a result, they either identify a suitable patient too late, or the patient is excluded by a criterion that was present in the documentation all along.

How it works:

• We download oncology trials in the Czech Republic and their criteria from ClinicalTrials.gov. We store them in IRIS in FHIR format: the trial as a ‘ResearchStudy’, and the criteria as a ‘Group’ within it.
• The model converts the criteria into separate Czech rules, each with verbatim citations from the register.
• Numerical limits and time windows are structured fields and are calculated by the code.
• Rules that cannot be verified from the documentation are sent directly for manual assessment.
• From the patient’s record, the model generates a profile: diagnosis, stage, biomarkers, treatment, performance status, comorbidities, medications and laboratory results, all with a date and a reference to the source. The ‘Classification System One’ model then reads only this profile.
• The core of the tool: the Jev model decides on each rule individually. It receives the profile and the rule as a query and returns only the probability of two answers: ‘yes’ or ‘no’. If the profile provides no indication, it assigns equal weight to both. The code from the issue determines the result: ‘yes’, ‘no’, or otherwise ‘unknown’ (i.e. data missing).
• Every definite answer must be supported by evidence; otherwise, it is ‘unknown’. The code verifies that the citation appears verbatim in the document.
• The coordinator can view the matrix of patients and criteria, the timeline and the details of each pair.
• Searches can be conducted in both directions: from candidates to the study, and from the study to the patient.
• There is an AI assistant on every page. The user selects a study, criterion, patient or verdict with the pointer, and the assistant explains what it sees and how the data are related.

Security:

• Missing patient data is marked as ‘unknown’ and represents a task for the coordinator.
• An excluded patient remains visible, along with the criterion that led to their exclusion.
• We display the numbers of criteria that have been met, not met and are unknown.

Deployment and data:

• The Next.js application is deployed on Vercel.
• The evaluation backend runs as a function on Vercel, with one call per patient-study pair.
• A Supabase auxiliary database stores the screening status, patient profiles and change history.
• The database from the InterSystems challenge, IRIS for Health with FHIR R4, is deployed on Azure. We have loaded sample data into it, which we received from the organisers.
• We call the models via OpenRouter with zero-data-retention.
• Only identifiers and results are logged.

What we’ve achieved:

• Deployed the entire workflow from registration to review by the coordinator, with an audit trail.

Access to the demo

• Email: demo@rakathon.cz
• Password: Winners