Automated ICD coding for coronary heart diseases by a deep learning method.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Data Science

Background:

• Automated ICD coding is a significant area of research, yet coronary heart disease (CHD) has been understudied.
• Existing methods often lack specific focus on CHD due to data limitations.

Purpose of the Study:

• To develop and evaluate a deep learning model for automated CHD coding.
• To address the challenge of long clinical text sequences in CHD coding.

Main Methods:

• Utilized Fuwai-CHD and MIMIC-III-CHD datasets.
• Developed a deep learning model (BW_att) integrating BERT variants for clinical text encoding, word2vec for code titles, and a label-attention mechanism.
• Implemented a truncation method to handle sequences longer than 512 tokens.

Main Results:

• BW_att achieved superior performance compared to baseline methods.
• On Fuwai-CHD, BW_att reached Macro-F1 of 96.2% and Macro-AUC of 98.9% for top codes.
• On MIMIC-III-CHD, BW_att achieved Macro-F1 of 40.5% and Macro-AUC of 66.1% for top codes.
• The model demonstrated interpretability by locating informative clinical text tokens.

Conclusions:

• The BW_att model accurately suggests CHD codes and offers robust interpretability.
• This deep learning approach shows significant potential for practical application in facilitating CHD coding.