Construction And Application Of Task-specific Information Extraction Framework Based On Medical Language Processing

Medical information technology aims to improve the quality of health care and reduce the costs of health care, which largely depends on the effective analysis and the use of clinical data. However, in actual clinical settings, complete structured data for clinical information system are quite limited, not to mention the fact that massive useful information exists only in narrative documentations. Therefore, it gains much popularity to utilize structured information extraction based on medical language processing.In China, although the natural language processing technology has been extensively used for various applications, which however, has not yet been applied to the clinical settings. To be more specific, China is faced with the lack of integrity, unity and standardized sets of medical terminology. Moreover, the various demands of clinical tasks cannot be easily fulfilled by a simple unified framework for text extraction. The three facts above give rise to the urgency of the effective use of medical language processing technology in the clinical decision support system. Focusing on such issues, the thesis carries out a comprehensive research on the design of the framework and the corresponding information extraction algorithms.Firstly, a task-specific information extraction framework is established for a wide variety of demands in the clinical practice. It is capable of monitoring the generation of clinical documents online and dynamically creating the corresponding information extraction tasks. A home-grown medical ontology which contains clinical concepts and relationships between them is maintained. The concepts required by the tasks that have not been enrolled in our home-grown medical ontology should be added through the framework platform, thus complementing this resource as a more comprehensive lexicon. Then the structured data of patients extracted from narrative documents can be smoothly fed into the CDS tasks.Secondly, three different extraction tasks based on the framework are implemented, namely the extraction of concept-value pair from clinical reports, drug-adverse event relation and symptom timeline both from progress notes. These tasks are assessed in real corpus. The results of the analyses are as follows. The precision of concept-value pair extraction from obstetric ultrasound examination report is 98.5%, the recall is 97.8%.3668 drug instructions and 39583 drug-adverse event relations have been collected in the drug-adverse event task. The result indicates that the extraction precision is 80.8%. The symptom extraction shows that the precision and recall is 97.2% and 58.3% separately.With the two parts above combined, the thesis establishes a task-specific information extraction framework and then implements the three tasks carefully. Now this framework has been deployed to a practical clinical environment, which lays a solid foundation for data utilizing.