A Compressed Language Model Embedding Dataset of ICD 10 CM Descriptions

Area of Science:

• Biomedical Informatics
• Machine Learning
• Natural Language Processing

Background:

• International Classification of Diseases, 10th Revision, Clinical Modification (ICD-10-CM) codes are crucial for healthcare but lack inherent numerical representations for advanced analysis.
• Existing methods may not fully capture the contextual relationships and hierarchical structures within ICD-10-CM codes.

Approach:

• Generated numerical embeddings for ICD-10-CM code descriptions using a large language model.
• Applied autoencoder-based dimension reduction to create compact, informative feature sets.
• Validated the dimension reduction and the utility of embeddings for predicting hierarchical categories.

Key Points:

• Novel datasets of numerical ICD-10-CM code representations were created.
• Dimension reduction to as few as 10 dimensions preserves embedding fidelity.
• Multiple compression levels are available for user flexibility.

Conclusions:

• The generated ICD-10-CM code datasets enhance machine learning model input features.
• This approach facilitates advanced analyses in biomedical informatics.
• The method holds potential to significantly improve the utility of ICD-10-CM codes in research.