Comparative effectiveness of medical concept embedding for feature engineering in phenotyping

Area of Science:

• Biomedical Informatics
• Computational Biology
• Health Data Science

Background:

• Feature engineering is a significant challenge in phenotyping.
• Medical Concept Embeddings (MCEs) can capture semantic relationships among medical concepts, aiding in feature retrieval.
• Comparing MCEs from different data sources is crucial for optimizing phenotyping.

Purpose of the Study:

• To compare the effectiveness of MCEs derived from knowledge graphs versus electronic healthcare records (EHR) for phenotyping tasks.
• To evaluate different embedding methods for learning MCEs.

Main Methods:

• Implemented five embedding methods: node2vec, SVD, LINE, skip-gram, and GloVe.
• Utilized two data sources: OMOP knowledge graphs and EHR data from CUIMC.
• Evaluated MCE performance using eMERGE phenotypes and concept relevance metrics (Hits@k%).

Main Results:

• MCEs learned using node2vec with knowledge graphs demonstrated superior performance.
• Both knowledge graph-based and EHR-based MCEs were evaluated, with node2vec (knowledge graphs) and GloVe (EHR) showing the best results respectively.
• Hierarchically structured knowledge graphs yielded better MCEs than EHR data.

Conclusions:

• MCEs facilitate scalable feature engineering for phenotyping.
• Knowledge graph-based MCEs, leveraging hierarchical concept relationships, are more effective than EHR-based MCEs for current phenotyping practices.