Interpretable deep learning to map diagnostic texts to ICD-10 codes

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Automatic extraction of disease information from death certificates is crucial for various applications, including billing and epidemiology.
• Natural language in clinical documents often deviates from standardized terminologies like the International Classification of Diseases (ICD), complicating automated coding.
• Developing a general and multilingual approach is necessary for consistent disease coding across different regions and languages.

Purpose of the Study:

• To propose a general and multilingual method for mapping diagnostic terms to the International Classification of Diseases (ICD) framework.
• To evaluate the proposed approach on clinical texts in French, Hungarian, and Italian.

Main Methods:

• The study frames ICD-10 encoding as a sequence-to-sequence task, leveraging neural networks for multi-class classification.
• Different neural network architectures were tested on datasets linking diagnostic terms to their corresponding ICD-10 codes.
• The approach was evaluated on multilingual datasets to assess its generalizability.

Main Results:

• The proposed method achieves state-of-the-art results in multilingual ICD-10 coding.
• High F-measures were obtained: 0.838 for French, 0.963 for Hungarian, and 0.952 for Italian.
• The model provides interpretable results by showing text-code alignments, aiding expert review.

Conclusions:

• The developed approach demonstrates the feasibility of automatic ICD-10 prediction in a multilingual context.
• This method offers a significant improvement over existing approaches for disease coding from clinical text.
• The interpretability of the model enhances its utility for clinical and research purposes.