JMIR Publications

ABSTRACT

Background:

For automatic multi-label text classification of the tenth revision of the International Classification of Diseases (ICD-10), this research aims to build a coding system using a contextual language model and rule-based preprocessing algorithm to improve the correctness of disease coding.

Objective:

This study aimed to establish a combination of natural language processing (NLP) and rule-based approaches for developing a more accurate and explainable ICD-10 auto-coding.

Methods:

We retrieved electronic health records (EHR) from a tertiary referral center and applied data mining and NLP techniques, including Word2Vec, XLNet, BERT, and AttentionXLM, to implement ICD-10 auto-coding system. Furthermore, we optimize the original contextual language model through the rule-based preprocessing approaches established by coding rules of disease classifiers.

Results:

The performance of the contextualized language model is better than the non-contextualized language model for the multi-label classification task. Our predicting result could achieve an F1-score of 0.77 and 0.69 on predicting ICD-10 clinical modification (ICD-10-CM) and procedure classification system codes (ICD-10-PCS), respectively, through a combination of the BioBERT pretrained model and rule-based preprocessing method. We improved the predication of ICD-10-CM with a keyword extractor and a combination code filter and the prediction for ICD-10-PCS with a keyword extractor and including surgical methods and examination reports.

Conclusions:

The performance of our model with the combination of the pretrained contextualized language model and rule-based preprocessing method is better than the model with only contextualized language in the multi-label classification task. This work highlights the importance of coding rules of disease classifiers to implant the rule-based algorithm into the coding system.