DeepRare: Revolutionizing Rare Disease Diagnosis with AI-Powered Agentic Technology

The Challenge of Diagnosing Rare Diseases

Rare diseases affect around 400 million people globally, with over 7,000 distinct disorders, most of which, about 80%, are genetic in origin. Diagnosing these diseases is notoriously challenging due to clinical heterogeneity, low prevalence, and limited clinician exposure. Patients often endure lengthy diagnostic journeys averaging over five years, including misdiagnoses and invasive procedures, which negatively impact treatment efficacy and quality of life.

Limitations of Existing Diagnostic Tools

Current diagnostic tools like PhenoBrain and PubCaseFinder use structured clinical terminologies and case records to identify rare diseases. Meanwhile, large language models (LLMs) such as Baichuan-14B and Med-PaLM help manage multimodal data. However, traditional tools struggle to adapt to evolving medical knowledge, and general LLMs may miss subtle rare disease nuances, limiting diagnostic performance.

Introducing DeepRare: An AI-Driven Diagnostic Platform

Developed by researchers from Shanghai Jiao Tong University, Shanghai AI Laboratory, Xinhua Hospital, and Harvard Medical School, DeepRare is the first agentic diagnostic system powered by LLMs specialized in rare disease identification. Its three-tier architecture includes a central host server with long-term memory and a state-of-the-art LLM, multiple specialized analytical agent servers for tasks like phenotype extraction and variant prioritization, and extensive external resources such as clinical guidelines and genomic databases.

How DeepRare Works

Clinicians input patient data in various formats (free text, HPO terms, genomic VCF data). The central host coordinates agent servers to retrieve relevant clinical evidence from external sources tailored to the patient's profile. Diagnostic hypotheses are generated and iteratively refined through a self-reflective mechanism, reducing errors. The system outputs a ranked list of diagnostic candidates supported by transparent reasoning linked to authoritative clinical references.

Performance and Validation

DeepRare was evaluated across eight benchmark datasets comprising 3,604 clinical cases representing 2,306 rare diseases in 18 specialties worldwide. It achieved a top-ranked diagnosis recall accuracy of 70.6% when combining phenotypic and genomic data, surpassing the next best method, Exomiser, by 17.4 percentage points. Its accuracy improved significantly in multimodal scenarios (from 46.8% with phenotype data alone to 70.6%).

Clinical Usability and Expert Feedback

Clinician evaluations of 50 complex cases showed a 95.2% agreement rate on the system's diagnostic validity and traceability. Physicians praised its efficiency in producing accurate, clinically relevant references, reducing diagnostic uncertainty. DeepRare is accessible via a user-friendly web application that supports structured patient data input, genetic files, and imaging reports.

Key Features of DeepRare

• First comprehensive agentic AI diagnostic system for rare diseases integrating advanced LLMs and analytical modules.
• Hierarchical modular architecture ensuring systematic and traceable diagnostics.
• Superior diagnostic accuracy demonstrated on extensive international datasets.
• Robust integration of phenotypic and genomic data enhancing diagnostic recall.
• High expert agreement on clinical relevance and transparent reasoning.
• Practical clinical integration through an intuitive web interface.

DeepRare represents a groundbreaking advancement in rare disease diagnostics, significantly improving accuracy and enabling timely interventions through AI-powered technology.