JAN: Joint Attention Networks for Automatic ICD Coding.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Computational Linguistics

Background:

• Manual International Classification of Diseases (ICD) coding is resource-intensive and prone to errors.
• Existing automatic ICD coding methods struggle with imbalanced datasets (long-tailed label distribution) and inter-label relationships (label correlation).

Purpose of the Study:

• To develop a novel end-to-end Joint Attention Network (JAN) for automatic ICD coding.
• To effectively address the challenges of long-tailed label distribution and label correlation in clinical document classification.

Main Methods:

• The proposed Joint Attention Network (JAN) utilizes document-based and label-based attention mechanisms.
• An Adaptive fusion layer and CorNet block are incorporated to manage attention weights and leverage label co-occurrence patterns.

Main Results:

• The JAN model achieved superior performance on MIMIC-III and MIMIC-II datasets compared to state-of-the-art methods.
• Key performance metrics include Micro-F1 of 0.553, Micro-AUC of 0.989, and P@8 of 0.735.

Conclusions:

• The developed Joint Attention Network (JAN) effectively handles long-tailed label distribution and label correlation for automatic ICD coding.
• Model visualizations confirm the effectiveness of attention mechanisms and label correlation, enhancing interpretability of deep learning models in healthcare.