Knowledge graph-augmented large language models for reconstructing life course risk pathways: a gestational diabetes...

Area of Science:

• Epidemiology
• Artificial Intelligence
• Medical Informatics

Background:

• Synthesizing complex epidemiological evidence for life-course exposure-outcome pathways is challenging.
• Existing methods struggle to integrate fragmented literature for inferring progressive risk.
• Gestational diabetes mellitus (GDM) and dementia serve as a critical case study for understanding life-course risks.

Purpose of the Study:

• To develop and evaluate a knowledge graph-augmented large language model (LLM) framework.
• To infer life-course exposure-outcome pathways, specifically linking GDM to dementia.
• To enhance epidemiological evidence synthesis using causal knowledge graphs and LLMs.

Main Methods:

• Constructed a causal knowledge graph from empirical epidemiological associations in scientific literature.
• Integrated the knowledge graph with GPT-4 using four graph retrieval-augmented generation (GRAG) strategies.
• Evaluated GRAG strategies using semantic triples, human experts, and LLM-based reviewers for reliability, novelty, and clinical relevance.

Main Results:

• The knowledge graph-augmented LLM identified 108 maternal candidate mediators between GDM and dementia.
• GRAG strategies significantly outperformed baseline LLM approaches in inferring bridging variables.
• The structured approach improved accuracy and reduced confabulation compared to standard LLM outputs.

Conclusions:

• Augmenting LLMs with epidemiological knowledge graphs facilitates reasoning over fragmented literature to reconstruct risk pathways.
• Human-AI collaboration is crucial for interpreting and applying LLM-derived clinical relevance.
• This framework offers a promising approach for life-course epidemiology, aiding early detection and study design.