Metadata harmonization from biological datasets with language models.

Area of Science:

• Biomedical Informatics
• Data Science
• Natural Language Processing

Background:

• Biomedical data integration is challenging due to inconsistent metadata and researcher-specific terminology.
• Current metadata harmonization methods are often labor-intensive or disrupt existing workflows.
• Manual standardization consumes over 40% of data curation time, hindering research progress.

Purpose of the Study:

• To develop and evaluate a language model-based solution for automated biomedical metadata harmonization.
• To improve the accuracy and efficiency of mapping researcher-specific terms to standardized vocabulary.
• To reduce the manual effort required for data curation and accelerate downstream data integration.

Main Methods:

• Fine-tuning GPT-2 language models with realistic data augmentation to generate varied term representations.
• Developing domain-specific models for cancer, alcohol research, and infectious disease data.
• Evaluating model performance using in-dictionary and out-of-dictionary accuracy metrics.

Main Results:

• Achieved 96% in-dictionary accuracy, reducing manual effort by over 90% for known terms.
• Demonstrated 17% out-of-dictionary accuracy for novel standard terms, outperforming existing methods.
• Domain-specific models showed superior performance on specialized terminology compared to larger general models.

Conclusions:

• The proposed language model approach offers a scalable and low-burden solution for biomedical metadata harmonization.
• Automated harmonization significantly accelerates data integration by minimizing manual curation.
• The method enables effective standardization even in domains lacking comprehensive synonym sets.