Precision information extraction for rare disease epidemiology at scale

Area of Science:

• Biomedical Informatics
• Computational Biology
• Public Health

Background:

• Millions worldwide live with rare diseases, necessitating better data for research and policy.
• Current methods for collecting rare disease epidemiologic information (EI) are slow and error-prone.
• Lack of comprehensive EI hinders understanding of rare disease variations and outcomes.

Purpose of the Study:

• To develop an automated pipeline for extracting epidemiologic information (EI) on rare diseases.
• To create a deep learning framework for Named Entity Recognition (NER) tailored to rare disease literature.
• To improve the efficiency and accuracy of curating EI for rare disease research and public health.

Main Methods:

• Developed a curated epidemiology corpus using weakly-supervised machine learning and manual validation.
• Adapted and fine-tuned the BioBERT model for NER on rare disease abstracts.
• Created the EpiPipeline4RD, a system with a web interface and API for EI extraction.

Main Results:

• Achieved high performance in EI extraction with F1 scores of 0.817 (entity-level) and 0.878 (token-level).
• Demonstrated comparable qualitative results to existing curated databases like Orphanet.
• Case studies confirmed efficient and precise extraction of EI for specific rare diseases.

Conclusions:

• EpiPipeline4RD significantly enhances manual curation of rare disease literature.
• This automated approach supports initiatives like the NIH Genetic and Rare Diseases Information Center (GARD).
• The pipeline advances public health efforts for the rare disease community through improved data accessibility.