An agentic system for rare disease diagnosis with traceable reasoning.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Genomics

Background:

• Rare diseases impact over 300 million globally, presenting significant diagnostic challenges.
• Patients face a lengthy 'diagnostic odyssey' with delays, misdiagnoses, and increased burdens.

Purpose of the Study:

• To introduce DeepRare, a multi-agent system for rare disease differential diagnosis decision support.
• To leverage large language models and specialized tools for enhanced diagnostic accuracy.

Main Methods:

• DeepRare integrates over 40 specialized tools and knowledge sources.
• It processes heterogeneous clinical data including free text, Human Phenotype Ontology (HPO) terms, and genetic results.
• A multi-agent system powered by large language models generates ranked diagnostic hypotheses with transparent reasoning.

Main Results:

• DeepRare demonstrated exceptional performance across 2,919 diseases in 14 medical specialties.
• Achieved an average Recall@1 of 57.18% in HPO-based tasks, outperforming other methods by 23.79%.
• In multi-modal tests, DeepRare reached 69.1% accuracy, surpassing Exomiser (55.9%). Expert review confirmed 95.4% agreement on reasoning validity.

Conclusions:

• DeepRare significantly advances rare disease diagnosis by providing accurate, evidence-based diagnostic hypotheses.
• The system demonstrates the potential of large-language-model-driven agentic systems to transform clinical workflows.
• This approach promises to reduce the diagnostic odyssey and improve patient outcomes for rare diseases.