Hyperbolic graph convolutional neural network with contrastive learning for automated ICD coding.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Machine Learning

Background:

• Manual International Classification of Diseases (ICD) coding is time-consuming and expensive.
• Automated ICD coding using deep learning faces challenges including imbalanced label distribution, code hierarchy, and noisy text.
• Existing methods struggle with ineffective and redundant label representation, particularly concerning code hierarchy.

Purpose of the Study:

• To introduce a novel Hyperbolic Graph Convolutional Network with Contrastive Learning (HGCN-CL) model for automated ICD coding.
• To address the limitations of previous methods in handling imbalanced label distribution and code hierarchy.
• To improve the accuracy and efficiency of automated ICD coding.

Main Methods:

• Utilized a Hyperbolic Graph Convolutional Network (HGCN) to effectively capture the hierarchical structure of ICD codes, mitigating embedding distortions.
• Integrated contrastive learning by injecting code features into the text encoder to generate hierarchical-aware positive samples.
• Trained and evaluated the HGCN-CL model on public MIMIC-III and MIMIC-II datasets.

Main Results:

• The HGCN-CL model demonstrated superior performance compared to state-of-the-art methods on the MIMIC III dataset.
• Achieved a 2.7% and 3.6% improvement over previous best results (Hypercore) for automated ICD coding.
• Ablation experiments and hierarchy visualization confirmed the effectiveness of the model's components.

Conclusions:

• The proposed HGCN-CL model offers a significant advancement in automated ICD coding.
• The model's ability to leverage hyperbolic geometry and contrastive learning effectively addresses key challenges in the field.
• HGCN-CL provides a more accurate and efficient solution for disease classification and medical data analysis.