Primary Submission Category: Applications in Health and Biology
CausalRAG: An Overview And Its Application to Precision Oncology
Authors: Brennan Kelley,
Presenting Author: Brennan Kelley*
LLMs, when deployed in clinical support roles, have routinely hallucinated recommendations that are unvalidated. To combat this, RAG, or Retrieval-Augmented Generation, was integrated, which has partially addressed the hallucination issue by grounding the LLM outputs in retrieved literature. However, this approach can still be contextually inappropriate, and treatments that do not have enough data to evaluate their efficacy can slip through. CausalRAG adds an additional layer of validation, where the recommendations for treatment go through a causal inference pipeline, helping to ensure that the treatments recommended by the LLM produce the intended outcomes. The pipeline involves an Empirical Bounding Box, Abstract extraction, and the DoWhy and EconML packages. The poster presentation introduces CausalRAG and display the results of some trials of CausalRAG in the context of the TCGA-BRCA breast cancer dataset containing records for 1,050 patients. Along with a breakdown of the Causal Pipeline.