Automatic ICD-10 coding: Deep semantic matching based on analogical reasoning.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Manual International Classification of Diseases, 10th Revision (ICD-10) coding is labor-intensive and costly.
• Existing automatic coding methods struggle with the discrepancy between clinical notes and ICD-10 classifications, especially for non-English records.
• There is a critical need for automated ICD-10 coding to support medical billing, resource allocation, and research.

Purpose of the Study:

• To develop and evaluate an automatic ICD-10 coding method for diagnoses extracted from Chinese discharge records.
• To address the limitations of current automatic coding techniques in handling language and classification differences.

Main Methods:

• Utilized BERT for initial word representations.
• Incorporated bidirectional Long Short-Term Memory (LSTM) for contextual information.
• Implemented a deep semantic matching layer with element-wise operations and a neural network, followed by a convolutional neural network and sigmoid output.

Main Results:

• The proposed method achieved high performance on a dataset of 1,003,558 primary diagnoses.
• Outperformed popular deep semantic matching algorithms (DSSM, ConvNet, ESIM, ABCNN).
• Demonstrated state-of-the-art single text matching accuracy (0.986), precision (0.979), recall (0.983), and F1-score (0.981).

Conclusions:

• The proposed deep semantic matching approach successfully enables automatic ICD-10 coding for Chinese diagnoses.
• The method's analogical reasoning capability is key to its effectiveness.
• This advancement has significant implications for improving the efficiency and accuracy of medical coding in Chinese healthcare settings.