Causal Graph Extraction from Medical Literature using Large Language Models for Hetionet

Causal relationships play a critical role in the medical domain, enabling deeper understanding of the connections between diseases, treatments, and biological factors, which are essential for accurate clinical decision-making. However, manually constructing causal knowledge is resourceintensive and cannot keep pace with the rapid growth of medical literature. Large Language Models (LLMs) offer the potential to automate causal graph extraction from unstructured medical text, yet they face the persistent challenge of hallucination-producing factually incorrect outputs. This paper proposes a framework for causal graph extraction from medical literature, incorporating hallucination-mitigation strategies through CO-STAR-based prompt engineering, augmented with few-shot examples and ontology constraints derived from the Hetionet knowledge graph. Extracted causal graphs are then integrated into Hetionet to enrich its biomedical knowledge base. Experiments using two state-of-the-art LLMs, GPT-4.1 and Gemini-2.0, across three medical topics (COVID-19, cardiovascular disease, and breast cancer) show that expert-validated precision ranges between 0.6383 and 0.9443, indicating relatively low hallucination levels. Integration results further demonstrate entity insertion rates of 47-68% and relation insertion rates of 80-93%, confirming significant enrichment of Hetionet. These findings highlight the effectiveness of the proposed framework in generating accurate causal graphs and enhancing biomedical knowledge graphs for downstream applications.