[A customized method for information extraction from unstructured text data in the electronic medical records]

Abstract

Objective: There is a huge amount of diagnostic or treatment information in electronic medical record (EMR), which is a concrete manifestation of clinicians actual diagnosis and treatment details. Plenty of episodes in EMRs, such as complaints, present illness, past history, differential diagnosis, diagnostic imaging, surgical records, reflecting details of diagnosis and treatment in clinical process, adopt Chinese description of natural language. How to extract effective information from these Chinese narrative text data, and organize it into a form of tabular for analysis of medical research, for the practical utilization of clinical data in the real world, is a difficult problem in Chinese medical data processing.

Methods: Based on the EMRs narrative text data in a tertiary hospital in China, a customized information extracting rules learning, and rule based information extraction methods is proposed. The overall method consists of three steps, which includes: (1) Step 1, a random sample of 600 copies (including the history of present illness, past history, personal history, family history, etc.) of the electronic medical record data, was extracted as raw corpora. With our developed Chinese clinical narrative text annotation platform, the trained clinician and nurses marked the tokens and phrases in the corpora which would be extracted (with a history of diabetes as an example). (2) Step 2, based on the annotated corpora clinical text data, some extraction templates were summarized and induced firstly. Then these templates were rewritten using regular expressions of Perl programming language, as extraction rules. Using these extraction rules as basic knowledge base, we developed extraction packages in Perl, for extracting data from the EMRs text data. In the end, the extracted data items were organized in tabular data format, for later usage in clinical research or hospital surveillance purposes. (3) As the final step of the method, the evaluation and validation of the proposed methods were implemented in the National Clinical Service Data Integration Platform, and we checked the extraction results using artificial verification and automated verification combined, proved the effectiveness of the method.

Results: For all the patients with diabetes as diagnosed disease in the Department of Endocrine in the hospital, the medical history episode of these patients showed that, altogether 1 436 patients were dismissed in 2015, and a history of diabetes medical records extraction results showed that the recall rate was 87.6%, the accuracy rate was 99.5%, and F-Score was 0.93. For all the 10% patients (totally 1 223 patients) with diabetes by the dismissed dates of August 2017 in the same department, the extracted diabetes history extraction results showed that the recall rate was 89.2%, the accuracy rate was 99.2%, F-Score was 0.94.

Conclusion: This study mainly adopts the combination of natural language processing and rule-based information extraction, and designs and implements an algorithm for extracting customized information from unstructured Chinese electronic medical record text data. It has better results than existing work.