Knowledge augmented causal discovery through large language models and knowledge graphs: application in chronic low...

Area of Science:

• Causal inference and artificial intelligence in biomedical research.

Background:

• Data-driven causal discovery is limited by dataset constraints and lack of external knowledge.
• Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) offer potential to augment causal discovery.

Purpose of the Study:

• To evaluate knowledge graph-based RAG (GraphRAG) for augmenting causal discovery in chronic lower back pain.
• To compare GraphRAG performance against LLM-augmented, RAG-augmented, and data-alone causal discovery.

Main Methods:

• Utilized an expert-defined causal graph for chronic lower back pain as ground truth.
• Implemented and benchmarked GraphRAG, RAG, and LLM augmentation for causal discovery.
• Explored various prompting strategies for causality assessment.

Main Results:

• GraphRAG achieved the highest F1 score (0.745) in augmenting causal discovery.
• GraphRAG outperformed RAG (0.714), LLM augmentation (0.636), and data-alone discovery (0.396).

Conclusions:

• GraphRAG represents a significant advancement in augmenting causal discovery.
• Integrating domain knowledge via graph-based RAG can accelerate causal modeling for complex conditions like chronic lower back pain.