MedT5SQL: a transformers-based large language model for text-to-SQL conversion in the healthcare domain.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Database Management

Background:

• Electronic medical records (EMRs) are increasingly prevalent, posing retrieval challenges for healthcare staff with limited database expertise.
• Accessible EMR retrieval is critical for effective medical care delivery.
• Natural Language Processing (NLP) offers a solution through Text-to-SQL conversion.

Purpose of the Study:

• To assess existing Text-to-SQL methods for EMR retrieval.
• To propose and evaluate the MedT5SQL model for converting natural language questions into SQL queries for EMR access.

Main Methods:

• The MedT5SQL model, based on the Text-to-Text Transfer Transformer (T5) Large Language Model (LLM), was fine-tuned on the MIMICSQL dataset.
• Performance was evaluated using two optimizers, varying training epochs, and the MIMICSQL and WikiSQL datasets.

Main Results:

• MedT5SQL achieved high accuracy on MIMICSQL: 80.63% exact match, 98.937% approximate string-matching, and 90% manual evaluation.
• On the WikiSQL dataset, MedT5SQL demonstrated 44.2% accuracy and 94.26% approximate string-matching.

Conclusions:

• Fine-tuned T5 models show significant potential for converting natural language medical queries to SQL.
• The MedT5SQL model offers an effective solution for improving EMR retrieval by non-technical healthcare staff.
• This research provides a foundation for advancing NLP applications in healthcare data access.