Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CDE-Mapper: Using Retrieval-Augmented Language Models for Linking Clinical Data Elements to Controlled Vocabularies

Komal Gilani, Marlo Verket, Christof Peters, Michel Dumontier, Hans-Peter Brunner-La Rocca, Visara Urovi

TL;DR

This work tackles the challenge of inconsistent clinical data element (CDE) representation by introducing CDE-Mapper, a Retrieval-Augmented Generation framework that links CDEs to multiple controlled vocabularies. The system is built as a modular pipeline with query decomposition, ensemble retrieval, knowledge filtering, a knowledge reservoir, and a two-step reranking mechanism, augmented by a human-in-the-loop validation for knowledge reservoir entries. Evaluated on four diverse datasets, CDE-Mapper achieves substantial accuracy gains over strong baselines and demonstrates robust handling of atomic, dependent, and composite CDEs, including complex units and historical contexts. The approach advances data interoperability and clinical decision support, while acknowledging limitations in vocabulary coverage, prompt design quality, and computational cost, and outlining directions for efficiency and vocabulary expansion. Overall, the study highlights the potential of RAG-based concept linking to improve harmonization of heterogeneous healthcare data at scale.

Abstract

The standardization of clinical data elements (CDEs) aims to ensure consistent and comprehensive patient information across various healthcare systems. Existing methods often falter when standardizing CDEs of varying representation and complex structure, impeding data integration and interoperability in clinical research. We introduce CDE-Mapper, an innovative framework that leverages Retrieval-Augmented Generation approach combined with Large Language Models to automate the linking of CDEs to controlled vocabularies. Our modular approach features query decomposition to manage varying levels of CDEs complexity, integrates expert-defined rules within prompt engineering, and employs in-context learning alongside multiple retriever components to resolve terminological ambiguities. In addition, we propose a knowledge reservoir validated by a human-in-loop approach, achieving accurate concept linking for future applications while minimizing computational costs. For four diverse datasets, CDE-Mapper achieved an average of 7.2\% higher accuracy improvement compared to baseline methods. This work highlights the potential of advanced language models in improving data harmonization and significantly advancing capabilities in clinical decision support systems and research.

CDE-Mapper: Using Retrieval-Augmented Language Models for Linking Clinical Data Elements to Controlled Vocabularies

TL;DR

This work tackles the challenge of inconsistent clinical data element (CDE) representation by introducing CDE-Mapper, a Retrieval-Augmented Generation framework that links CDEs to multiple controlled vocabularies. The system is built as a modular pipeline with query decomposition, ensemble retrieval, knowledge filtering, a knowledge reservoir, and a two-step reranking mechanism, augmented by a human-in-the-loop validation for knowledge reservoir entries. Evaluated on four diverse datasets, CDE-Mapper achieves substantial accuracy gains over strong baselines and demonstrates robust handling of atomic, dependent, and composite CDEs, including complex units and historical contexts. The approach advances data interoperability and clinical decision support, while acknowledging limitations in vocabulary coverage, prompt design quality, and computational cost, and outlining directions for efficiency and vocabulary expansion. Overall, the study highlights the potential of RAG-based concept linking to improve harmonization of heterogeneous healthcare data at scale.

Abstract

The standardization of clinical data elements (CDEs) aims to ensure consistent and comprehensive patient information across various healthcare systems. Existing methods often falter when standardizing CDEs of varying representation and complex structure, impeding data integration and interoperability in clinical research. We introduce CDE-Mapper, an innovative framework that leverages Retrieval-Augmented Generation approach combined with Large Language Models to automate the linking of CDEs to controlled vocabularies. Our modular approach features query decomposition to manage varying levels of CDEs complexity, integrates expert-defined rules within prompt engineering, and employs in-context learning alongside multiple retriever components to resolve terminological ambiguities. In addition, we propose a knowledge reservoir validated by a human-in-loop approach, achieving accurate concept linking for future applications while minimizing computational costs. For four diverse datasets, CDE-Mapper achieved an average of 7.2\% higher accuracy improvement compared to baseline methods. This work highlights the potential of advanced language models in improving data harmonization and significantly advancing capabilities in clinical decision support systems and research.
Paper Structure (29 sections, 10 equations, 5 figures, 8 tables, 1 algorithm)

This paper contains 29 sections, 10 equations, 5 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Clinical Data Elements (CDEs)
  4. Controlled Vocabularies
  5. Retrieval-Augmented Generation (RAG)
  6. Embedding Models
  7. Related Work
  8. datasets
  9. Proposed Framework
  10. Knowledge Base Creation
  11. Linking Rules
  12. Query Decomposition
  13. Knowledge Retriever
  14. Knowledge Filter
  15. Knowledge Reservoir
  16. ...and 14 more sections

Figures (5)

  • Figure 1: Atomic and Composite terms for for 4 example clinical data elements (sex, GGT, FHH, Caisson). The CDEs are described by their common label, the encoding datatype (e..g int), the nature (e.g. discrete or continuous), associated unit, point in time, method, or formula.
  • Figure 2: Modular Framework Stepwise Workflow for Concept Linking of Common Data Elements Using RAG
  • Figure 3: Detailed Overview of Proposed Modular RAG Framework for normalization and linking CDEs in controlled vocabularies (e.g. SNOMED, LOINC, ATC, RxNorm)
  • Figure 4: Two-step reranking to select a candidate using a combination of classification and relevance score.
  • Figure 5: Comparison of NCGD@K for Different Models Across datasets