A compressed large 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 essential for healthcare but lack inherent numerical representation for advanced analysis.
• Existing methods for utilizing ICD-10-CM codes in machine learning are limited by their discrete nature and lack of contextual information.

Purpose of the Study:

• To generate novel datasets of numerical representations for ICD-10-CM codes.
• To create informative input features for machine learning models by capturing semantic relationships and context.
• To enable more advanced analyses in biomedical informatics using readily available datasets.

Main Methods:

• Utilized a large language model to generate description embeddings for ICD-10-CM codes.
• Applied dimension reduction via an autoencoder to compress the embeddings.
• Validated the dimension reduction using the autoencoder and a supervised model for hierarchical category estimation.

Main Results:

• Successfully reduced the dimensionality of ICD-10-CM code embeddings to as few as 10 dimensions.
• Maintained the ability to reproduce original embeddings with high fidelity at reduced dimensions.
• Provided multiple compression levels for user-selectable requirements.

Conclusions:

• The generated numerical datasets of ICD-10-CM codes significantly enhance their utility in biomedical informatics.
• This approach facilitates more advanced machine learning applications by providing context-aware, dimension-reduced features.
• The readily available datasets require no additional setup, promoting wider adoption and research.