Medical code prediction via capsule networks and ICD knowledge.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Natural Language Processing

Background:

• Clinical notes are vital for patient health status documentation.
• International Classification of Diseases (ICD) codes are crucial for electronic health records.
• Current neural network methods struggle with the complexity and length of clinical notes for accurate code assignment.

Purpose of the Study:

• To propose a hybrid capsule network model for automatic medical code assignment.
• To improve the efficiency and accuracy of extracting features from lengthy clinical notes.
• To leverage label information for enhanced medical code prediction.

Main Methods:

• A hybrid capsule network model incorporating bi-directional Long Short-Term Memory (Bi-LSTM).
• A label embedding framework to integrate text and label information.
• A dynamic routing algorithm within the capsule network for feature extraction.

Main Results:

• The model achieved a micro F1-score of 67.5% on the MIMIC-III dataset.
• Outperformed existing state-of-the-art methods in automatic medical code assignment.
• Demonstrated effective feature capture from complex clinical note texts.

Conclusions:

• The proposed model effectively captures important features across sentences using dynamic routing and label embedding.
• Capsule networks and domain knowledge are beneficial for medical code prediction.
• The hybrid approach offers a significant advancement in automated medical coding.