A BERT-Based Generation Model to Transform Medical Texts to SQL Queries for Electronic Medical Records: Model...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Database Management

Background:

• Electronic medical records (EMRs) are stored in relational databases, necessitating SQL queries for data retrieval.
• Medical experts face challenges in formulating SQL queries due to specialized knowledge requirements.
• Current text-to-SQL generation methods have limited adoption within the medical domain.

Purpose of the Study:

• To develop a neural generation model for automatic transformation of medical text into SQL queries for EMRs.
• To create a model that integrates medical text characteristics with SQL structure for improved query generation.

Main Methods:

• Proposed MedTS (Medical Text-to-SQL) model using a pretrained Bidirectional Encoder Representations From Transformers (BERT) encoder.
• Employed a grammar-based long short-term memory network decoder to predict an intermediate syntax tree representation.
• Utilized syntax trees as intermediate representations to align with SQL's structure and reduce generation search space.

Main Results:

• MedTS achieved 0.784 accuracy in logic form and 0.899 in execution on the MIMICSQL dataset.
• The model significantly outperformed existing state-of-the-art text-to-SQL methods.
• Performance across generated SQL components was balanced and showed substantial improvements.

Conclusions:

• The MedTS model demonstrates effectiveness and robustness in medical text-to-SQL generation.
• The system shows strong potential for practical application in real-world medical scenarios.
• This advancement can enhance data accessibility and utilization within EMR systems.