Research On The Screening Of High Risk Individuals Of Cerebral Infarction Based On Multi-view Learning

Cerebral infarction is an acute cerebrovascular disease with high deformity and mortality.With the growth of living standard and average life span,the number of patients shows an explosive growth trend.Due to the lack of general and effective treatment,the clinical treatment of cerebral infarction generally adopts the idea of combining prevention and treatment,focusing on prevention.Among them,the screening of high-risk individuals of cerebral infarction aims to find the high-risk individuals of cerebral infarction in advance,and then delay or avoid acute clinical events through preventive intervention,so as to reduce the burden on individuals and families.Some existing screening studies of high risk individuals of cerebral infarction treat the collected risk factor characteristics as general single-view structured data,ignoring its multi-view characteristics.In addition,in the screening of high-risk individuals of cerebral infarction,there is correlation between multiple features in the same view,which jointly reflects a certain medical attribute of the sample.At the same time,there are connections between views,some information is shared by multiple views,and the existing methods do not verify the extracted view common information.In view of the above problems and shortcomings,the main work of this paper includes the following aspects:(1)Aiming at the deficiency of treating risk factors data as single-view data and neglecting its multi-view characteristics in the existing research work,the data was divided into views according to the medical background knowledge and reconstructs into the multi-view data.Based on the Bi LSTM,a multi-view learning model is designed,which integrates multiple view data and synthesizes the output of multiple cells.The results of a series of experiments on the relevant data sets prove the effectiveness of the approach and model.(2)Aiming at the problem that redundancy is not considered when extracting view specific information in the existing research work,the auto-encoder structure is used to extract the specific information and remove the redundant representation between the specific information and the original information.The middle layer representation of the auto-encoder is used as the low dimensional representation of the view data,and fusion prediction by using Bi LSTM.The results of a series of experiments on cerebral infarction data set and other two benchmark data sets prove the effectiveness of the approach and model.(3)Aiming at the problem that there is no verification when extract the view common information in the existing research work.A antagonistic network structure is designed to extract the view common information by referring to the design idea of Generative Adversarial Nets.The feature extractor is used to extract features that can confuse the discriminator,as common information of views to assist model prediction.In the case of more input information,attention mechanism is introduced to improve data utilization efficiency and prediction ability.The results of a series of experiments on cerebral infarction data set and other two benchmark data sets prove the effectiveness of the approach and model.