Multi-view representation learning for tabular data integration using inter-feature relationships.

Area of Science:

• Data Science
• Machine Learning
• Bioinformatics

Background:

• Harmonizing diverse data sources is a key challenge in data science, particularly in healthcare.
• Integrating data from multiple origins with unmapped features is essential for developing generalizable algorithms.
• Existing methods often rely on ambiguous or unavailable metadata, necessitating new approaches.

Purpose of the Study:

• To design and evaluate methods for mapping structured tabular datasets from electronic health records (EHRs) independent of metadata.
• To identify effective strategies for feature mapping when only a small set of features are initially known.

Main Methods:

• Comparison of contrastive learning, partial auto-encoders, mutual-information graph optimizers, and statistical baselines.
• Evaluation on simulated data, public datasets, MIMIC-III, and perioperative records.
• Performance assessment based on feature mapping accuracy and data reconstruction.

Main Results:

• Contrastive learning methods demonstrated superior performance in feature matching and reconstruction, especially on real-world data.
• Partial auto-encoders performed comparably to contrastive methods in many scenarios.
• A novel statistical method showed reasonable performance with less hyperparameter tuning.

Conclusions:

• Inter-feature relationships are effective for identifying matching features across tabular datasets lacking metadata.
• Decoder architectures can effectively impute features when exact matches are not found.