Reframing AI for Rare Disease Recognition.

Area of Science:

• Computational biology
• Medical informatics
• Genomics

Background:

• Rare diseases impact over 300 million globally, with diagnostic delays hindering timely treatment and research participation.
• Current computational rare disease recognition (RDR) methods struggle with incomplete and heterogeneous knowledge resources, requiring extensive expert curation.
• Direct application of large language models (LLMs) for diagnosis faces knowledge bottlenecks and raises concerns about cost, reproducibility, and data governance.

Purpose of the Study:

• To introduce GEN-KnowRD, a novel framework that leverages LLMs to construct a computable knowledge base for rare disease recognition.
• To evaluate GEN-KnowRD's performance in general-purpose rare disease screening and early discrimination using longitudinal electronic health records.
• To demonstrate the benefits of a knowledge-layer-first approach in RDR, decoupling knowledge generation from patient-level inference.

Main Methods:

• GEN-KnowRD utilizes LLMs to generate schema-guided rare disease profiles and assesses their quality.
• A computable knowledge base, PheMAP-RD, is constructed for local deployment.
• Lightweight inference pipelines integrate this knowledge for disease screening and early discrimination from electronic health records.

Main Results:

• GEN-KnowRD significantly improved disease ranking in general-purpose screening benchmarks (up to 345.8% improvement vs. HPO-centered framework).
• The framework outperformed advanced end-to-end LLM reasoning (up to 129.1% improvement) and expert-curated knowledge variants.
• In a real-world idiopathic pulmonary fibrosis cohort, GEN-KnowRD showed robust gains in early discrimination performance.

Conclusions:

• Repositioning LLMs to the knowledge layer enhances rare disease recognition (RDR) performance.
• GEN-KnowRD provides a scalable, updatable, and reusable infrastructure for rare disease diagnosis, screening, and research.
• This approach addresses limitations of existing RDR methods, offering improved accuracy and efficiency.