Overview
- Researchers describe DeepRare, which integrates clinical notes, patient symptoms, and genomic data to generate ranked rare-disease diagnoses with evidence-linked reasoning.
- Built around a large language model with specialized agent modules, the system extracts symptoms, matches diseases, and annotates and ranks genetic variants.
- In one evaluated cohort, DeepRare correctly identified diagnoses in 69% of cases, outperforming the widely used Exomiser tool at 56%.
- In a head-to-head review of 163 cases, the system reached correct diagnoses in about 79% of instances compared with 66% for rare-disease specialists.
- Independent reviewers judged DeepRare’s reasoning chains accurate in 95% of cases in the study, and authors note potential to support non-specialist clinicians and reduce diagnostic delays, though findings reflect research testing rather than clinical deployment.