Applying AI to Support Categorization of Heterogeneous Epidemiological Datasets

Area of Science:

• Health Informatics
• Data Science
• Biomedical Research

Background:

• The increasing importance of Findable, Accessible, Interoperable, and Reusable (FAIR) data in research.
• NFDI4Health aims to improve health data findability, reusability, and interoperability for epidemiological, clinical, and public health studies.
• Existing platforms like the German Central Health Study Hub and Maelstrom Catalog use standardized categorization for data reuse.

Purpose of the Study:

• To present an AI solution for automatic classification and annotation of health study variables.
• To integrate this AI solution into the NFDI4Health Metadata Annotation Workbench.
• To enhance data findability and reuse through accelerated and improved variable categorization.

Main Methods:

• Development of a BioBERT-based text classifier for automatic classification and annotation.
• Integration of the AI model into the NFDI4Health Metadata Annotation Workbench service.
• Evaluation of the model's performance using a weighted F1-score.

Main Results:

• The BioBERT-based text classifier achieved a weighted F1-score exceeding 92%.
• The AI solution demonstrated improved annotation performance, especially for non-expert users.
• Accelerated categorization of study variables was observed, enhancing data findability and reuse.

Conclusions:

• The AI solution significantly accelerates the categorization of study variables, boosting data findability and reuse.
• Further development of AI approaches is expected to reduce curatorial workload.
• The AI tool promotes the creation of semantically annotated, interoperable data catalogs, advancing FAIR data principles in health research.