Design And Implementation Of Systematic Review Citation Screening System Based On Literature Similarity

Systematic review is a very important evidence basis for Evidence-based medicine(EBM),which has important reference value for clinical treatment,diagnostic test and risk analysis.Filtering references is a step in systematic review,which usually requires at least two experts to review manually.Due to the rapid growth of the number of medical documents and the poor specificity of medical database search engines,screening references has become a time-consuming and labor-consuming process.Many studies have shown that the use of automatic text classification technology can greatly reduce the workload of experts' manual review of documents.In the present study,a citation screening algorithm for systematic review based on document similarity is proposed.This algorithm aims at the problems of unbalanced data sets in the current automatic text classification algorithm based on machine learning and insufficient data in the early stage of training of the machine learning classification algorithm combined with active learning.The similarity between documents is calculated by using information such as title,abstract and publication type of documents,and the documents are prioritized according to the document screening situation of experts and the similarity between documents to accelerate the citation screening process.In this paper,LDA(Latent Dirichlet Allocation)topic model is used to construct the feature vectors of titles and abstracts,respectively,and bag-of-words model is used to construct the feature vectors of publication types.Through calculating cosine values between the feature vectors,the similarity between documents is measured.In order to verify the effectiveness of the algorithm in this paper,based on the data provided by Cochrane Library,this study constructs a system review citation set of ten topics as test data,compares the other two classical algorithms in the field of systematic review citation screening,that is,the support vector machine classification algorithm combined with active learning and the Naive Bayes classification algorithm combined with active learning Results show that the WSS95 score of this algorithm is 20.09% higher than that of other algorithms when the title and abstract are used as input respectively,while the WSS95 score of this algorithm is only 1.93% lower than that of other algorithms when the publication type is used as input,which proves that the algorithm proposed in this paper can effectively reduce the workload of experts in citation screening for systematic review.In the present study,a citation screening system for systematic review is designed and implemented.The main function of this system is to assist users in systematic review citation screening,and to give the prioritisation score of documents as a reference value for users to stop document screening.In order to facilitate users to find documents,the system has realized the function of document retrieval,which can use different element information of documents to find documents with similar topics.The system also carries out the same storage management for users' documents,which is convenient for users to use the system to save and manage documents.