The PICO Element Recognition In The Case Of Insufficient Annotation Data

Evidence-based medicine(EBM)is an evidence-based medical practice approach that requires medical practitioners to develop treatment plans within the best available research evidence carefully,accurately,and wisely.In EBM,in order to formulate clinical question,the PICO framework is widely used,which divides clinical problems into four elements:Population/Problem(P),Intervention(I),Construction(C)and Outcome(O).When searching evidence for clinical question,medical practitioners hope to retrieve medical literature abstracts according to PICO elements.However,most of the current medical literature abstracts do not specify the PICO elements,resulting in the existing retrieval tools can not support the structured retrieval of PICO elements.Therefore,recognizing PICO elements in the medical literature abstracts automatically,which has become an important task in medical informatics.In recent years,deep learning has become the mainstream method of PICO element recognition,that can capture deep semantic information automatically and improve the effect of PICO element recognition greatly,but it requires a large amount of annotation data.However,the amount of manual annotation data is limited in the PICO element recognition task.Although some researchers have found that the element tags corresponding to PICO elements in some English medical literature abstracts can be used to obtain annotation data,but in some languages,such as Chinese,there is not structured element tags corresponding to PICO elements in most medical abstracts,which leads to the distant supervision method can not be used to obtain labeled data.Aiming at the problem that the effect of deep learning PICO element recognition method declines in the insufficiency of labeled data,improving the effect of PICO element recognition in the case of insufficient annotation data,one introduced a element continuity constraint and the other considered the element position information.The paper includes the following work.Firstly,the sentences describing the same element in medical literature abstracts usually appear consecutively,that is,there is element continuity.Therefore,using the characteristic of element continuity,this paper proposed a PICO recognition method that introduces an element continuity constraint to guide the deep learning model to capture element continuity information.In order to verify the effectiveness of the proposed method in the case of insufficient labeled data,this paper conducted experiments on the small artificially dataset NICTA-PIBOSO,and at the same time,conducted experiments on training sets of different scales in the distant supervision dataset Pub Med-PICO.The experimental results show that the proposed PICO recognition method that introduced a element continuity constraint is better than the latest existing methods in the case of insufficient annotation data.Secondly,medical literature abstracts often follow a certain logical structure,and the position distribution of sentences with different elements in the abstract is regular.Therefore,this paper proposed a PICO recognition method that considered element position information,which using the location-aware conditional random field(location-aware CRF)for sequence labeling,and learn different state transition matrices for different elements’ position range.In order to verify the PICO element recognition effect of the proposed method in the insufficiency of labeled data,this paper conducted experiments on the small artificially dataset NICTA-PIBOSO,and at the same time,conducted experiments on training sets of different scales in the distant supervision dataset Pub Med-PICO.The experimental results show that the proposed PICO recognition method that considered the element position information can improve the effect of PICO elements recognition in the case of insufficient annotation data.