An Advanced ICD-9 Terminology Standardization Method Based on BERT and Text Similarity

Abstract

The ICD-9 terminology standardization task aims to standardize the colloquial terminology recorded by doctors in medical records into the standard terminology defined in the ninth version of International Classification of Diseases (ICD-9). In this paper, we first propose a BERT and Text Similarity Based Method (BTSBM) that combines BERT classification model with text similarity calculation algorithm: 1) use the N-gram algorithm to generate a Candidate Standard Terminology Set (CSTS) for each colloquial terminology, which is used as the training dataset and test dataset for next step; 2) use the BERT classification model to classify the correct standard terminology. In this BTSBM method, if a larger-scale CSTS is taken as the test dataset, the training dataset also needs to maintain larger-scale. However, there is only one positive sample in each CSTS. Hence, expanding the scale will cause a serious imbalance in the ratio of positive and negative samples, which will significantly degrade system performance. While if we keep the test dataset relatively small, the CSTS Accuracy (CSTSA) will degrade significantly, which results a very low system performance ceiling. In order to address above problems, we then propose an optimized terminology standardization method, called as Advanced BERT and Text Similarity Based Method (ABTSBM), which 1) uses a large-scale initial CSTS to maintain a high CSTSA to ensure a high system performance ceiling, 2) denoises CSTS based on body structure to alleviate the imbalance of positive and negative samples without reducing the CSTSA, and 3) introduces the focal loss function to further promote a balance of positive and negative samples. Experiments show that, the precision of the ABTSBM method is up to 83.5%, which is 0.6% higher than BTSBM, while the computation cost of ABTSBM is 26.7% lower than BTSBM.