Rare diseases: this AI could save years of diagnosis, the first tests amaze doctors
Between thick medical files, consultations that follow one another and hope that stretches, diagnostic wandering remains the daily life of many patients suffering from rare diseases. For many, doctors struggle to connect scattered symptoms and sometimes contradictory examinations, due to lack of time and reference points in an immense medical literature.
THE rare diseases affect more than 300 million people worldwide, and
diagnosis often takes more than five years, with its procession of referrals, errors and unnecessary interventions. This “odyssey” weighs heavily on families and health systems. A new AI, DeepRarepublished in the journal Naturepromises to shake up this scenario thanks to
evidence-based predictions. It remains to be seen how it could, very concretely, shorten this wandering.
DeepRare, an agentic AI facing the puzzle of diagnosing rare diseases
Thought like a co-pilot more than like an oracle,
DeepRare is based on a so-called “agentic” architecture. At the center, a host powered by a large language model coordinates around forty specialized digital tools: some transform clinical descriptions into standardized terms, others analyze DNA, query PubMed, Orphanet or OMIM, or even search for similar cases in a database of nearly 70,000 patients.
From simple free-text medical notes, lists of symptoms coded with the Human Phenotype Ontology and, if available, sequencing files, the system produces a ranked list of probable rare diseases. Tested on 6,401 cases covering 2,919 diseases, it identified from its first proposal the correct diagnosis in 57.18% of cases, well ahead of the second best model, note the authors.
Evidence-based predictions to shorten the diagnostic odyssey
Where many AI tools are satisfied with an answer,
DeepRare provides for each hypothesis a chain of reasoning linking symptoms, genetic data, scientific articles and similar cases. Ten rare disease specialists judged this path to be logical in 95.4% of cases. “DeepRare is one of the first computational models to exceed the diagnostic performance of expert physicians in the complex task of phenotyping and diagnosing rare diseases“, summarized the team in an article from Naturecited by Medical Xpress.
In a head-to-head test of 163 complex cases, five experienced doctors found the correct diagnosis first in 54.6% of cases. The system reached 64.4%. With genetic data, its rate rose to 63.6% in Hunan Hospital and 69.1% in Xinhua, above Exomiser, another tool of this type. “Our work not only advances the diagnosis of rare diseases but also shows how the latest powerful agentic systems based on large language models can reshape today’s clinical workflows.“, they added.
What DeepRare can change and what remains to be proven
A web application already allows clinicians to enter a patient’s data, add a genetic file and obtain, in a few minutes, a short list of possible diagnoses accompanied by references. The study remains retrospective and a few erroneous citations were observed, but the authors are now planning validations on large cohorts to measure, with supporting figures, how many years of wandering this co-pilot will really be able to avoid.