Multigranularity Label Prediction Model for Automatic International Classification of Diseases Coding in Clinical...

Area of Science:

• Medical Informatics
• Computational Linguistics
• Health Services Research

Background:

• International Classification of Diseases (ICD) coding is crucial for global disease statistics.
• Current ICD coding methods struggle with the vast, hierarchical nature of ICD codes.
• Existing studies often overlook the hierarchical relationships between ICD code levels.

Purpose of the Study:

• To develop an improved ICD coding prediction model.
• To leverage the hierarchical structure of ICD codes for enhanced accuracy.
• To address the limitations of models focusing only on subcategory prediction.

Main Methods:

• Proposed a multitask learning model with multiple classifiers for different ICD code levels.
• Incorporated a reinforcement mechanism to capture relationships between coarse and fine-grained labels.
• Evaluated the model on English and Chinese benchmark datasets using attention mechanisms.

Main Results:

• Achieved competitive performance compared to baseline models, especially in macro-F1 scores.
• Demonstrated effective utilization of ICD code hierarchy for improved prediction.
• Attention analysis revealed multigranularity attention captures crucial text features for explanation.

Conclusions:

• The proposed multitask learning approach effectively leverages ICD code hierarchy.
• The model shows promise for improving the accuracy and interpretability of automated ICD coding.
• This method offers a more robust solution for disease classification and statistical analysis.