Exploiting hierarchy in medical concept embedding.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Medical codes like ICD-10 lack explicit hierarchical structure in standard embeddings.
• Existing embeddings may not fully capture the nuanced relationships between medical concepts.

Purpose of the Study:

• To create and release medical concept embeddings incorporating ICD-10 code hierarchy.
• To improve the representation of medical codes for downstream tasks.

Main Methods:

• Extended Word2Vec algorithm to include hierarchical information from medical codes and CCSR categories.
• Trained embeddings on a dataset of ~600,000 patients.
• Compared performance against standard Word2Vec and other pretrained embeddings.

Main Results:

• Hierarchically-informed Word2Vec models outperformed standard models (61.4% vs. 57.5% NMI).
• Improved classification performance by 96.2% and 66.7% in different outcome models.
• All developed models surpassed competitive benchmarks.

Conclusions:

• Proposed algorithms effectively capture the hierarchical structure of medical codes.
• Enhanced embeddings improve performance on classification tasks.
• Released ICD-10 pretrained embeddings outperform existing popular vectors.