Understanding how diseases develop and progress is one of the central goals of biomedical research. Recent advances in Large Language Models (LLMs) and knowledge graphs make it possible to automatically extract and structure disease-related concepts (e.g., symptoms, risk factors, biomarkers, mechanisms) from scientific literature. We already have a prototype system that uses LLMs to extract elements of disease theory from research papers and populate a knowledge graph database. To make this system more powerful and useful, there is a need to enhance its analytical capabilities, improve interpretability, and extend it to capture temporal and multimodal aspects of disease progression.
This project will extend an existing LLM + knowledge graph disease theory prototype by implementing advanced features that improve its utility for researchers. Possible directions include:
- Multi-paper disease theory analysis: support cross-checking and comparison of extracted knowledge across multiple sources to detect consensus, conflicts, or gaps.
- Interpretability and transparency: develop mechanisms to make outputs more explainable, e.g., provenance tracking, evidence highlighting, or metadata about extraction steps.
- Causal inference for disease theory: experiment with methods to strengthen causal claims (e.g., linking risk factors → mechanisms → outcomes).
- Temporal modeling of disease progression: integrate temporal information from papers (timestamps, disease stages, longitudinal studies) to model disease dynamics.
- Joint multimodal temporal generation: extend the system to embed diverse modalities (e.g., temporal text data, timestamps, imaging sequences) to capture spatio-temporal dependencies.
The enhanced system will be demonstrated in a selected disease case study to illustrate its potential for supporting biomedical research and theory-building. As a demonstration, the system can be applied to Primary Sclerosing Cholangitis (PSC), a rare, chronic liver disease characterized by progressive inflammation and scarring of the bile ducts. PSC is strongly associated with inflammatory bowel disease but its underlying causes remain unclear, involving possible autoimmune, genetic, and environmental factors. Research data spans clinical measures (liver function tests, biomarkers), imaging (MRI/MRCP sequences), and textual reports (case studies, cohort analyses, guidelines). By extracting and linking concepts from multiple papers, the system can build a disease theory graph that highlights comorbidities, biomarkers, and progression pathways. Temporal modeling can further represent disease evolution from asymptomatic stages to advanced outcomes such as cirrhosis or cholangiocarcinoma. This case study illustrates how LLM + knowledge graph approaches can support cross-paper analysis, causal inference, and transparent provenance tracking in rare disease research.