Research On Missing Imputation For Medical Data

As the development Plan of The New Generation of Artificial Intelligence has clearly proposed the development of intelligent medical treatment,the application of modern medical information technology to mine valuable knowledge from medical data has become a current research hotspot.However,in the actual process of obtaining medical data,the collected data are often missing due to machine failures,environmental constraints,collection costs and other factors,and the incomplete data information brings severe challenges to medical data mining and follow-up tasks.Therefore,how to effectively complete the missing medical data to improve the effect of downstream analysis task has become a worthy topic of study.In order to eliminate the impact of missing medical data,this thesis studies the missing completion of different types of data,and has achieved some results.The main research results of this thesis are as follows:(1)A semi-supervised multi-modal medical image synthesis framework based on adversarial learning is proposed to solve the problem of modal loss in multi-modal medical images in unstructured medical data.The semi-supervised training method can integrate the distribution information of unpaired data and the sample details of paired data,which can effectively get rid of the dependence of traditional supervised synthesis method on a large number of paired training data.Due to the introduction of detailed supervision information,the visual effect of the synthesized image of the proposed model is obviously better than that of the unsupervised model.The effectiveness of the algorithm is verified by a large number of experiments on multimodal MR image data sets.(2)Aiming at the data loss of longitudinal electronic medical records in structured medical data,a context-aware medical time series data completion method was proposed.In view of the lack of data context may provide data characteristics of learning gain,by way of introducing the mixed fusion information to aid and downstream classification forecasting task completion task,and through the experiments on multiple data sets show that our model completion results have good generalization,can be better used in the subsequent analysis task.(3)Aiming at the problem of missing random blocks of electronic medical record data in structured medical data,a medical random block sequential data completion algorithm based on Transformer is proposed.The application of self-attention mechanism can make the model dynamically adjust the weight of input data,so as to better mine the correlation between missing data and existing data in the case of missing whole data.By adopting multi-task cooperative training,the model can learn more universal representations.The validity of the algorithm is verified by experiments on the constructed data set.