Building Structured Patient Follow-up Records from Chinese Medical Records via Deep Learning

Abstract

Employing deep learning (DL) method to process and analyze Chinese medical records to build patient follow-up records (PFRs) has been a very valuable task. In recent years, the identification and classification of clinical terms in electronic medical records has received increased attention. However, electronic medical records are difficult to access because of their exceedingly high privacy, so it has become more feasible to extract information from paper medical records. This study proposed a DL approach that extract text information from the pre-processed images of Chinese medical records by optical character recognition (OCR) model base on CRNN first, and then identify the clinical entities using named entity recognition (NER) model based on BERT-CRF. The experimental results of this study demonstrate that the proposed method achieves precision over 75%, which is more than 90% for some specific entities. In addition, the proposed method can be extended as a universal approach to other diseases that require the establishment of the structured patient follow-up records (PFRs).