Deep graph convolutional network for US birth data harmonization.

Area of Science:

• Data Science
• Bioinformatics
• Computer Science

Background:

• Fragmented databases hinder comprehensive analysis of long-term trends.
• Integrating historical datasets requires significant manual effort and expertise.
• Developing automated methods for database harmonization is crucial for research.

Purpose of the Study:

• To develop a deep learning framework for harmonizing US natality data.
• To create a consistent database from fragmented sources spanning five decades.
• To reduce manual effort in data integration and improve research efficiency.

Main Methods:

• Constructed a graph database representing variables and their properties.
• Employed a graph convolutional network (GCN) to learn variable similarity embeddings.
• Devised a novel loss function with a slack margin and banlist mechanism.
• Integrated an active learning mechanism for iterative harmonization with human review.

Main Results:

• Successfully harmonized 9,321 variables from 49 databases (1970-2018).
• Achieved 87.56% recall and 57.70% precision in the first harmonization round.
• Identified 323 hyperchains of variables, revealing inter-database relationships.
• Demonstrated the efficacy of the harmonized graph neural network (HGNN) method.

Conclusions:

• The HGNN method offers a feasible and efficient approach to meta-level database connection.
• The framework effectively learns global patterns and discovers variable similarities across databases.
• This approach significantly reduces manual effort in database harmonization and data integration.