Automated Data Harmonization in Clinical Research: Natural Language Processing Approach.

Area of Science:

• Computational biology and bioinformatics
• Health informatics
• Machine learning in healthcare

Background:

• Data integration is crucial for advancing clinical and epidemiological research.
• Harmonizing diverse variables across datasets is a significant bottleneck.
• Existing methods for variable harmonization are often inefficient and not scalable.

Purpose of the Study:

• To develop and assess a natural language processing (NLP)-based method for automated variable harmonization.
• To enable scalable integration of multiple research datasets.
• To improve the efficiency and accuracy of data harmonization in large-scale studies.

Main Methods:

• Developed a fully connected neural network (FCN) model enhanced with contrastive learning.
• Utilized domain-specific embeddings from Bidirectional Encoder Representations from Transformers for Biomedical Text Mining.
• Trained and validated the model on three cardiovascular datasets (Atherosclerosis Risk in Communities, Framingham Heart Study, Multi-Ethnic Study of Atherosclerosis).
• Framed the harmonization task as a paired sentence classification problem using metadata descriptions.

Main Results:

• The FCN method achieved a top-5 accuracy of 98.95% and an AUC of 0.99.
• This significantly outperformed the logistic regression baseline (top-5 accuracy 22.23%, AUC 0.82).
• Contrastive learning enhancement also showed superior performance over the baseline.

Conclusions:

• The novel FCN approach offers a scalable solution for harmonizing metadata in large cohort studies.
• This method accurately categorizes harmonized concepts for cardiovascular disease and stroke research.
• The NLP-based strategy substantially enhances data integration capabilities over traditional methods.