Breaking barriers in ICD classification with a robust graph neural network for hierarchical coding.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Documentation Improvement

Background:

• Accurate International Classification of Diseases (ICD) code classification is crucial for clinical documentation.
• Existing automated methods struggle with the complexity and nuances of medical text.
• Traditional models lack flexibility and robustness in handling sparse medical data.

Purpose of the Study:

• To propose an advanced adversarial learning framework, LGG-NRGrasp, for automated ICD coding.
• To address limitations of existing methods in flexibility, robustness, and handling complex medical text.
• To improve the accuracy and dependability of diagnostic code assignment to medical discharge summaries.

Main Methods:

• Developed Labeled Graph Generation with Node Representation Grasp (LGG-NRGrasp), an adversarial learning framework.
• Modeled ICD coding as a labeled graph generation problem, incorporating a hierarchical structure for feature learning.
• Integrated adversarial reinforcement learning and domain adaptation techniques for enhanced generalization.

Main Results:

• LGG-NRGrasp demonstrated superior performance compared to leading models on benchmark datasets.
• The framework effectively addresses over-smoothing issues in deep graph neural networks.
• Achieved enhanced performance and dependability in automated ICD code classification.

Conclusions:

• LGG-NRGrasp offers a robust and flexible solution for automated ICD coding.
• The proposed method significantly improves the accuracy of diagnostic code assignment.
• This framework advances the field of clinical documentation improvement through AI.