ICDXML: enhancing ICD coding with probabilistic label trees and dynamic semantic representations.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Natural Language Processing

Background:

• Accurate clinical coding is essential for healthcare applications.
• Medical language complexity poses challenges for automated diagnosis and procedure coding.
• Existing methods struggle with the nuances of clinical text for precise International Classification of Diseases (ICD) assignment.

Purpose of the Study:

• To propose a novel model for enhancing International Classification of Diseases (ICD) coding from clinical text.
• To improve the accuracy and robustness of automated clinical code assignment.
• To leverage multimodal learning and hierarchical structures for richer semantic encoding.

Main Methods:

• Developed a model incorporating extreme multi-label classification for ICD coding.
• Utilized deformable convolutional neural networks to fuse representations from pre-trained language models.
• Integrated external medical knowledge embeddings using a multimodal approach.
• Constructed a probabilistic label tree based on ICD's hierarchical structure for structured prediction.

Main Results:

• The proposed model demonstrated competitive performance in medical code prediction tasks.
• Experiments were conducted on the MIMIC-III database, a widely used clinical dataset.
• The fusion of language model representations and medical knowledge embeddings proved effective.

Conclusions:

• The novel model offers a significant advancement in robust clinical code assignment.
• Incorporating extreme multi-label classification and hierarchical structures enhances ICD coding accuracy.
• This technique shows promise for improving healthcare applications reliant on precise clinical documentation.