Querying Healthcare Data Lake Using LLMs Without Direct Data Exposure: A Feasibility Study

The integration of diverse healthcare data into unified data lake infrastructures is a promising way to support research and clinical decision-making. Yet, querying these data lakes, such as those based on MongoDB, remains complex, often requiring specialized expertise and custom scripts. Large language models (LLMs) offer new opportunities by translating natural language queries into executable commands, but direct access raises major privacy concerns. We present a method for indirect LLM-assisted querying of a healthcare data lake that preserves data confidentiality by avoiding any direct model access to real data. The multistep workflow includes user intent clarification, schema and example guided query generation, and expert validation before execution. In a preliminary evaluation using five representative test queries, the pipeline consistently produced syntactically valid code aligned with user intent. Qualitative assessment showed its potential for reducing technical barriers while also highlighting current limitations, including hallucinated fields and missing lookups. These findings suggest that indirect LLM-assisted querying represents a promising step toward more secure and usable access to healthcare data lakes, though further large-scale validation is required.