Case studies / Healthcare · Oncology research
How AI classified hereditary-cancer variants with 98% expert agreement
A research project with the familial cancer unit of a public university hospital. Agents gather the evidence for each gene variant, apply the clinical criteria and explain every call.

- 98%agreement with the final expert classification
- 525variants evaluated
- 27 of 37disagreements with the commercial tool settled for the AI
- 1,152anonymised patients
The situation
The unit studies families with a history of cancer. Sequencing a patient returns gene variants, and a geneticist has to classify each one into one of five categories, from benign to pathogenic. That classification guides the follow-up of the patient and their relatives.
The problem
Classifying one variant takes a geneticist 30 to 60 minutes: searching population and clinical databases, reading the evidence and applying the ACMG/AMP criteria. Sequencing keeps raising the number of variants, and expert time does not grow with it.
What we built
- Evidence gathering
Agents read the sequencing output and collect population, clinical and computational evidence for each variant from reference databases such as gnomAD and ClinVar.
- Clinical criteria
The system applies the ACMG/AMP criteria and proposes one of the five categories.
- Reasoning a geneticist can audit
Every call comes with its justification and its sources, so the expert reviews the reasoning instead of repeating the search.
The results
- 98% agreement with the final expert classification, on 525 variants.
- The AI and the hospital’s commercial tool disagreed on 37 variants. Physicians reviewed them and sided with the AI in 27.
- Evaluated on data from 1,152 anonymised patients.
Evaluated against the final classification made by the hospital’s geneticists. This is a research project: the geneticist remains responsible for every classification.
Client anonymised. Full case file and reference available under NDA.


