ZebraMap: A Multimodal Rare Disease Knowledge Map with Automated Data Aggregation & LLM-Enriched Information...

Area of Science:

• Medical Informatics
• Genomics and Rare Diseases
• Computational Biology

Background:

• Clinical knowledge for rare diseases is fragmented across various sources, hindering diagnosis.
• Existing data formats are often unstructured and heterogeneous, complicating information retrieval.
• Delayed diagnosis is a common challenge for patients with rare diseases.

Purpose of the Study:

• To present ZebraMap, a multimodal knowledge map for consolidating rare disease information.
• To convert narrative case evidence into structured, machine-readable data for rare diseases.
• To enhance rare disease research through accessible, computable knowledge.

Main Methods:

• Utilized Orphanet for rare disease identification (1727 diseases) and linked to PubMed case reports.
• Developed an automated large language model (LLM) pipeline to parse free-text case reports into standardized fields.
• Implemented a retrieval-augmented generation (RAG) component to enrich disease-level descriptions with peer-reviewed research.

Main Results:

• Created a dataset with 69,146 structured patient-level case texts and 98,038 clinical images.
• Achieved 94.5% cosine similarity between curated and generated text, indicating high data quality.
• Demonstrated satisfactory performance in information extraction and structured data generation.

Conclusions:

• ZebraMap is the largest openly accessible multimodal resource for rare diseases.
• The platform transforms narrative evidence into computable knowledge, enabling data-driven research.
• Facilitates improved understanding and diagnosis of rare diseases through consolidated information.