Imagine a doctor needing to quickly find all patients who received a specific treatment in the last year. Instead of wrestling with complex database queries, they simply ask a question in plain English. This is the promise of AI-powered text-to-SQL systems, and new research from Korea University is pushing the boundaries of what's possible in Electronic Health Records (EHRs). Researchers at KU's DMIS lab tackled a major challenge in this field: how to handle questions an EHR system can't answer. Their innovative framework, presented at the EHRSQL 2024 shared task, uses a clever combination of question "templatization" and a fine-tuned large language model (LLM) similar to ChatGPT. First, the system rephrases incoming questions into a standardized format. This helps the LLM, trained on the EHR's database structure, to generate the correct SQL query. But what if a question asks for information the database doesn't contain? The KU-DMIS system has a built-in verification step. It actually runs the generated SQL query against the database. If it hits an error, the system recognizes the question as unanswerable and politely declines to respond, preventing inaccurate or misleading results. This research achieved impressive results in the EHRSQL 2024 competition, demonstrating the power of this approach. While the system currently relies on a resource-intensive LLM, the team is looking ahead to more accessible open-source models. This work opens exciting possibilities for making EHR data more readily available to medical professionals, ultimately leading to faster insights and better patient care.