Large language model chatbot-based text-to-SQL application for database analyses in liver diseases and hepatology research

Authors:

Abstract
Objective: Large language models (LLMs) may improve analysis of structured clinical data by enabling natural language (NL) queries. We developed an interactive, PHI-compliant “Text2SQL” chatbot as a proof-of-concept for hepatocellular carcinoma (HCC) research. Materials and methods: Using the University of California, San Francisco’s secure artificial intelligence platform, we configured the chatbot to translate NL queries into Structured Query Language (SQL), execute them, and return both code and results. Clinicians posed 30 questions, and outputs were compared with SQL written by a data scientist. We assessed prompt interpretation, SQL accuracy, output accuracy, and error types according to strict binary criteria. Results: Prompt interpretation accuracy was 73.3%, SQL accuracy ranged from 53% to 63%, and output accuracy from 53% to 63%. Frequent errors included group comparisons, ambiguous variables, and exploratory questions. Discussion: Text2SQL enables PHI-compliant interaction with clinical datasets without coding, though performance varies across query types. Conclusion: LLM-powered Text2SQL offers a promising approach for clinician-friendly clinical data access.
Funding Information
  • Financial/Grant Support (K23DK139455)
  • National Institute of Diabetes and Digestive and Kidney Diseases, Ge (P30DK026743)
  • National Institute of Diabetes and Digestive and Kidney Diseases, Ge (UCSF Liver Center Grant)
  • National Institute of Diabetes and Digestive and Kidney Diseases, Ge (R01DK133527)
  • National Institute of Diabetes and Digestive and Kidney Diseases (K24AG080021)
  • National Institute on Aging (T32DK060414)
  • National Institute of Diabetes and Digestive and Kidney Diseases