Research On Relevant Algorithms For Automated Risk Of Bias Assessment In Systematic Reviews

Systematic reviews,mainly done by professional reviewers and doctors manually,were regarded as the source of evidence with the highest quality in evidence-based medicine.Risk of bias assessment needs to be done on clinical trial reports(Randomized Controlled Trials,RCTs)that were recruited in the systematic reviewers.It is very time-consuming and error-prone.Meanwhile,the publication of biomedical literature soars recently,which made the systematic reviews,especially the risk of bias assessment,more difficult.This study proposed a method to complete the risk of bias assessment automatically via introducing algorithm in artificial intelligence.N-gram and TF-IDF(Term Frequency-Inverse Document Frequency)were combined in this study to obtain the features,and the Linear Support Vector Machine(Linear SVM)classifier was utilized to complete the risk of bias assessment automatically.The high sparsity and latitude of the BoW(Bag of Word)was effectively settled The results of experiment showed that the F1 value of the model on the document classification task of the automated risk of bias assessment task was 62.8%-80.0%,and the F1 value of the sentence classification task was 67.7%-74.7%.Furthermore,an automated bias risk assessment model based on BERT(Bidirectional Encoder Representations from Transformers)was proposed since the lack of contextual semantic in traditional feature engineering.Previous studies showed that the BERT performed better because it used the Transformer structure as a feature extractor,while learning contextual information with a bidirectional language model.Experimental results showed that the F1 value of the model on the document classification task was about 14.7% higher than that of the traditional method(combining the TF-IDF feature engineering and SVM classifier method),and the F1 value on the sentence extraction task was higher 18.2 % around.Datasets used in this study were the original 3802 RCTs and their corresponding assessment data obtained from the Cochrane Database of Systematic Reviews(CDSR),Pubmed and other databases.Firstly,the pdfminer3 k toolkit was used in the preprocessing of the original RCT.Secondly,the similarity calculation of text was proposed to obtain the corresponding risk of bias assessment.Finally,the data can be used in the training of the automated risk of bias assessment was obtained,which is the largest in the relevant research.In the following,an automated risk of bias assessment system was designed and implemented in this study.The system was developed based on a lightweight framework-the Flask,and was deployed on the local server and the Alibaba Cloud server by using Docker container technology.The results of the experiments showed that the time consumed in the risk of bias assessment can be largely reduced with high accuracy.