Clinical Evidence Attribute And Relationship Extraction In Pico Framework

With the increasing of clinical medicine literrature,evidence-based medicine practitioners need to read a large number of medical literature to obtain the latest research results.Although field experts summarize clinical evidence for the reference of relevant personnel,manual summary of clinical evidence is time-consuming and laborious.Therefore,the automatic generation of structured clinical evidence is important for fasting the practice of evidence-based medicine.According to analysis,the existing studies on structured clinical evidence are usually sentence level literature abstracts or the identification of a few important attributes,while the fully structured clinical evidence studies are rare.However,sentence level summary still belongs to unstructured information,which cannot directly construct structured clinical evidence,and identifying a few important attributes is not enough to fully describe clinical evidence.Based on the above background,this paper studied the clinical evidence attributes and relationship extraction methods for PICO framework,and extracted as much as possible rich structured information.The main research contents are as follows:(1)Ahierarchical attribute and relationship extraction for PICO(HARE)framework is proposed.This method consists of two steps: sentence division and phrase classification.Meta Map is used for sentence division,and phrase classification is used for hierarchical multi-category classification method based on PICO framework.HARE proposes two new sample features for phrase classification: the phrase based shortest dependency path,which obtains the important information that is not adjacent to the phrase in a long sentence;the attention of the tag to the sample,which obtains the keywords of the description tag according to the sample statistics,and emphasizes the key information of the tag.HARE eventually outputs a set of candidate attributes and relationships for clinical evidence.By feature ablation and comparison experiment,this paper proves the effectiveness of feature extraction and classification method in HARE.(2)Rule based Clinical Evidence Generation(RCEG)is proposed,which includes two steps: removing redundancy of candidate sets and generating triples based on rules.First,duplicate information in the candidate set is removed based on textual similarity,and then the text describing attributes and relationships is transformed into a triad of structured clinical evidence.Rule-based approach is used in the triplet generation process,including generation of age,exclusion criteria and dose of drug based on regular matching,and generation of disease or symptoms based on semantic rules.The quality of structured clinical evidence generated by the RCEG was demonstrated through a customized evaluation strategy and manual evaluation.(3)This paper designed and implemented a clinical evidence generation and visualization system(Aceso Evidence,Aceso E).Compared with the existing studies,this system has comprehensively identified the attributes and relationships in structured clinical evidence based on PICO framework.Aceso E integrates HARE and RCEG,and its main functions include uploading literature abstracts,generating structured clinical evidence,visualized clinical evidence,and downloading results.In summary,this paper studies the PICO oriented framework structured clinical evidence attribute and relation extraction methods: Firstly,the candidate set of attributes and relationships was identified based on deep learning,and then structured clinical evidence was generated based on rules.Finally,a clinical evidence generation and visualization system was designed and realized.