16:15 – 16:45
Assist. Prof. Ahmet Kaplan
Istanbul Medipol University, Istanbul, Türkiye
Generative AI for Healthcare Data: Transforming Natural Language Queries into SQL for Medical Patient Records
Abstract:
The integration of Generative AI models with structured medical patient data presents a transformative opportunity for healthcare analytics. At Istanbul Medipol University, we are developing a system that allows academicians and clinicians to query Relational Database Management Systems (RDBMS) using natural language, eliminating the need for manual SQL scripting. This approach leverages Text-to-SQL techniques, where large language models (LLMs) with fine-tuned open-source models translate user questions into executable SQL queries.
Key challenges include preserving patient privacy (via anonymization and access controls), handling complex medical terminologies, and ensuring high accuracy in SQL generation. We explore few-shot learning, retrieval-augmented generation (RAG), and hybrid rule-based + neural approaches to improve robustness. Additionally, we evaluate fine-tuning strategies using domain-specific medical datasets and real-time clinician feedback loops to enhance model performance in clinical contexts.
Our preliminary results indicate that Text-to-SQL systems, combined with structured EHR data, can significantly streamline medical research by enabling real-time, natural language-based data retrieval. This research aligns with Medipol University’s mission to bridge AI innovation with healthcare efficiency, improving a data-driven academic ecosystem.