Building Evidence-Centered Digital Libraries for Rare Disease Trajectories

Abstract

Rare-disease knowledge presents a digital-library problem: clinically useful evidence is fragmented across case reports and clinical narratives, while conventional access remains largely document-centered. My prior work on primary ciliary dyskinesia (PCD) evaluated five open-source large language models (LLMs) on 58 real first-visit pediatric pulmonology notes. Models often identified patients who warranted further PCD evaluation, but clinician review showed that explanations and testing recommendations were frequently only partly sound. My doctoral research therefore asks how digital libraries can transform dispersed case-report collections into trustworthy, navigable representations of disease progression. I will curate PCD case reports, extract patient-level temporal trajectories with source origin, aggregate recurring and atypical patterns, and generate clinician-facing and patient-facing summaries from the same structured evidence. Evaluation will cover extraction accuracy, temporal reconstruction, factual support, omission, origin, audience appropriateness, and clinician usefulness, using automatic claim verification together with manual review. The intended contribution is not diagnostic automation, but an origin-preserving framework for organizing, synthesizing, and communicating sparse rare-disease evidence in digital libraries.