Dense Retrieval for Electronic Health Record With Knowledge Injection and Synthetic Data

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Information Retrieval

Background:

• Electronic Health Records (EHRs) are crucial for clinical practice but face retrieval challenges due to semantic gaps.
• Existing dense retrieval models lack sufficient medical knowledge or use mismatched training data, limiting their EHR retrieval capabilities.
• Previous EHR retrieval systems often lack generalizability and are trained on limited query sets.

Purpose of the Study:

• To introduce DR.EHR, a novel series of dense retrieval models specifically designed for effective EHR retrieval.
• To address the limitations of current models by developing a two-stage training pipeline that incorporates extensive medical knowledge and large-scale data.

Main Methods:

• A two-stage training pipeline was utilized, leveraging MIMIC-IV discharge summaries.
• Stage one involved medical entity extraction and knowledge injection from a biomedical knowledge graph.
• Stage two employed large language models for diverse training data generation, training DR.EHR variants (110M and 7B parameters).

Main Results:

• DR.EHR models significantly outperformed existing dense retrievers on the CliniQ benchmark, achieving state-of-the-art results.
• Models demonstrated superiority in various match and query types, excelling in challenging semantic matches like implication and abbreviation.
• Ablation studies confirmed the effectiveness of individual pipeline components, and generalization was shown on EHR QA datasets.

Conclusions:

• The proposed DR.EHR models represent a significant advancement in EHR retrieval, offering a robust solution for clinical applications.
• The two-stage training pipeline effectively addresses the need for medical knowledge and large-scale data in EHR retrieval.
• DR.EHR models demonstrate strong generalizability across different EHR corpora and complex natural language questions.