Explainable clinical coding with in-domain adapted transformers.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Artificial Intelligence

Background:

• Automatic clinical coding from Electronic Health Records (EHR) is essential for information extraction.
• Existing methods often function as "black boxes," lacking transparency in their coding assignments.
• This limits the real-world applicability of automated clinical coding systems.

Purpose of the Study:

• To develop explainable clinical coding using transformer-based models.
• To enable models to assign clinical codes and identify supporting text references.
• To enhance the transparency and reliability of automated clinical coding.

Main Methods:

• Evaluated 3 transformer architectures on 3 explainable clinical coding tasks.
• Compared general-domain vs. in-domain (medical) adapted transformer models.
• Addressed explainable clinical coding as dual medical named entity recognition (MER) and normalization (MEN) tasks.
• Developed multi-task and hierarchical-task strategies for MER and MEN.

Main Results:

• Clinical-domain transformers significantly outperformed general-domain models.
• The hierarchical-task approach showed superior performance over the multi-task strategy.
• An ensemble of clinical-domain transformers with the hierarchical approach achieved state-of-the-art results (e.g., F1-score of 0.852 on Cantemist-Norm).

Conclusions:

• The hierarchical-task approach simplifies explainable clinical coding by separating MER and MEN tasks.
• This method establishes new state-of-the-art performances for the studied predictive tasks.
• The methodology can be extended to other clinical tasks involving medical entity recognition and normalization.