LLM-Powered Knowledge Graph of Causal Relations in Drug Reviews

This paper presents the employment of JSL-MedLlama, a decoder-only Large Language Model (LLM) trained within the medical domain, to create a knowledge graph of causal relationships from drug reviews. We leverage a dataset of causal narratives from clinical notes, MIMICause, to benchmark JSL-MedLlama for classifying causal narratives using instruction fine-tuning. The results show that it obtains satisfying performance, outperforming other encoder-only baselines. Furthermore, we validate our algorithm robustness and cross-domain generalization by testing it on the Drug Reviews dataset, a collection of patient reviews on specific drugs along with related conditions. We then deploy the model on a subset of around 19,000 Drug Reviews, generating a knowledge graph of 3,050 unique triples connecting 1,149 Drugs and 322 Conditions through the considered causal relations. The results highlight the role of decoder-only LLMs, fine-tuned within the biomedical domain, in advancing causal reasoning and generating valuable resources for real-world biomedical use cases. We make publicly available the drug-condition causal relation knowledge graph to support future research efforts in the field.