Explainable automated coding of clinical notes using hierarchical label-wise attention networks and label embedding...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Natural Language Processing for Clinical Documentation

Background:

• Automated medical coding from clinical notes aims to improve efficiency and accuracy over manual processes.
• Current deep learning models for automated coding lack explainability and ignore complex correlations among medical codes.
• Poor model interpretability hinders confident adoption in clinical practice.

Purpose of the Study:

• To develop an explainable automated medical coding system addressing limitations of current deep learning models.
• To enhance deep learning models by incorporating label correlations for improved performance.
• To quantify the importance of words and sentences for each medical label using attention mechanisms.

Main Methods:

• Proposed a Hierarchical Label-wise Attention Network (HLAN) for interpretable medical coding.
• Introduced a label embedding (LE) initialization approach to capture correlations among medical codes.
• Evaluated HLAN and LE on MIMIC-III discharge summaries, comparing against state-of-the-art CNN and RNN models.

Main Results:

• HLAN achieved top performance in Micro-level AUC and F1 score for top-50 code prediction (91.9% and 64.1%).
• HLAN provided more meaningful and comprehensive model interpretations compared to baseline models.
• Label embedding initialization significantly boosted a state-of-the-art model's performance to 52.5% Micro-level F1 for full code prediction.

Conclusions:

• HLAN offers comparable or superior automated coding results with enhanced explainability.
• Label embedding initialization effectively improves deep learning model performance in multi-label classification for medical coding.
• The developed methods show potential for deployment in hospitals, with further research areas identified.