BDIViz: An Interactive Visualization System for Biomedical Schema Matching with LLM-Powered Validation

Area of Science:

• Biomedical Informatics
• Data Science
• Human-Computer Interaction

Background:

• Biomedical data harmonization is crucial for research but schema matching is difficult.
• Current automated methods struggle with accuracy on complex biomedical schemas.
• Manual schema matching is time-consuming and prone to errors.

Purpose of the Study:

• To develop a novel visual analytics system, BDIViz, to streamline biomedical data schema matching.
• To address scalability and semantic ambiguity challenges in schema matching.
• To improve the efficiency and accuracy of matching disparate biomedical datasets.

Main Methods:

• Developed BDIViz, an interactive visual analytics system incorporating user-centered design.
• Employed an ensemble approach combining multiple matching methods with LLM-based validation.
• Utilized interactive heatmaps and coordinated views for efficient attribute comparison.

Main Results:

• BDIViz significantly improves schema matching accuracy compared to baseline methods.
• The system reduces cognitive load and curation time for domain experts.
• User studies demonstrated the effectiveness of BDIViz in two biomedical case studies.

Conclusions:

• BDIViz offers an effective solution for the challenging task of biomedical data schema matching.
• Interactive visualization and ensemble methods enhance accuracy and usability.
• The system facilitates more efficient exploratory analyses and meta-studies in biomedicine.