Knowledge augmented causal discovery through large language models and Knowledge Graphs: Application in chronic low back pain.

OBJECTIVE

Causal discovery from observational data is limited by structural constraints of available datasets, the absence of causal logic, and the lack of external domain knowledge. We propose a Knowledge-Augmented Causal Discovery (KACD) framework that combines data-driven structural learning with LLM-based knowledge systems to overcome these limitations, with application to chronic low back pain (cLBP).

METHODS

We evaluated KACD in three configurations of increasing complexity: a vanilla LLM, an LLM augmented with retrieval-augmented generation (RAG), and an LLM augmented with a Knowledge Graph-based RAG system (KG-RAG). Each configuration was benchmarked against a clinically derived, expert-validated causal graph for cLBP. Causal querying employed a decomposed two-phase prompting strategy - separately querying plausibility, statistical association, and temporal precedence - inspired by the methodology used by domain experts during ground truth construction. Performance was evaluated using F1-score, true positive rate, false discovery rate, and structural Hamming distance.

RESULTS

KG-RAG with plausibility-based prompting achieved the highest F1 score (0.697), outperforming standard RAG (F1 0.689), vanilla LLM (F1 0.649), and data-driven structural learning alone (F1 0.396), however this was backbone-dependent with RAG outperforming KG-RAG depending on the LLM used. Decomposed causal prompting outperformed direct causal queries across all knowledge system configurations. Combining the data-driven proto-model with any knowledge system consistently raised the performance floor of either method individually across all prompting strategies and system configurations.

CONCLUSION

LLM-driven knowledge systems can effectively augment data-driven causal discovery for biomedical causal modeling, with KG-RAG providing superior performance through cross-document synthesis and entity-centered retrieval. These findings support KACD's potential as a decision-support tool for researchers constructing causal models in complex, multi-disciplinary health domains, pending further validation of its use in practice.