Projection Word Embedding Model With Hybrid Sampling Training for Classifying ICD-10-CM Codes: Longitudinal...

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Current International Classification of Diseases, Tenth Revision Clinical Modification (ICD-10-CM) code search models rely on word embeddings, but are limited by embedding quality and vocabulary diversity.
• Electronic Health Record (EHR) embeddings offer medical understanding but lack vocabulary breadth.
• A need exists for word embeddings balancing open-source vocabulary diversity with EHR-specific medical terminology.

Purpose of the Study:

• To propose a projection word2vec model and a hybrid sampling method for improved ICD-10-CM code searching.
• To validate the effectiveness of these novel methods through comprehensive experiments.

Main Methods:

• Compared projection word2vec with traditional word2vec using English Wikipedia and PubMed abstracts.
• Evaluated medical semantic understanding across seven datasets and applied embeddings for ICD-10-CM code identification in discharge notes.
• Utilized 94,483 labeled discharge notes for training and 24,762 for testing, with additional validation on 74,324 notes from seven other hospitals.

Main Results:

• Projection training improved Wikipedia embeddings' medical semantic understanding, though EHR and PubMed embeddings remained superior.
• The model using projection PubMed and Wikipedia embeddings achieved the highest F-measure (0.7362 and 0.6693) in ICD-10-CM coding tasks.
• The hybrid sampling method further enhanced model performance, yielding F-measures of 0.7371 and 0.6698.

Conclusions:

• EHR and PubMed embeddings demonstrate superior medical semantic understanding; projection word2vec enhances Wikipedia embeddings' medical capabilities.
• While projection word2vec's impact on ICD-10-CM coding was moderate, it aids in handling emerging diseases.
• The hybrid sampling method significantly improves model performance, mimicking human expert capabilities in medical coding.