A natural language processing approach to support biomedical data harmonization: Leveraging large language models.

Area of Science:

• Biomedical informatics
• Computational biology
• Data science

Background:

• Biomedical research necessitates large, diverse datasets for unbiased results.
• Retrospective data harmonization is crucial but labor-intensive.
• Automated variable matching methods are needed to accelerate this process.

Purpose of the Study:

• To develop and evaluate novel methods for automated variable matching.
• Leverage large language models (LLMs) and ensemble learning for variable matching.
• Improve the efficiency of biomedical data harmonization.

Main Methods:

• Utilized data from two GERAS cohort studies (European and Japan).
• Developed four natural language processing (NLP) methods using LLMs (E5, MPNet, MiniLM, BioLORD-2023).
• Implemented an ensemble learning method (Random Forest) integrating NLP methods.

Main Results:

• The ensemble Random Forest model outperformed individual LLM methods.
• The Random Forest model achieved an average HR-30 of 0.986 and MRR of 0.744.
• LLM-derived features were the primary contributors to the ensemble model's performance.

Conclusions:

• NLP techniques, particularly LLMs, show great potential for automating variable matching.
• Ensemble learning enhances the accuracy and efficiency of automated variable matching.
• These methods can significantly accelerate biomedical data harmonization for large-scale studies.