| Electronic medical records(EMR)record the whole process of patients ’ medical diagnosis,which contains a lot of valuable information.Mining these information and applying it to patient similarity research can provide medical services for patients in a more timely and convenient manner.At the same time,it has positive significance for the improvement of medical health level and research in the medical field.Patient similarity studies learn a clinically meaningful distance metric based on patient-related clinical data to measure the degree of similarity between patients.However,electronic medical records contain a large amount of semantic information and there are complex correlations between data.These complex medical data have a natural graph structure,traditional deep learning methods are difficult to apply to these graph data,and how to better process such complex medical data to obtain more accurate patient representation is the key.The emergence of graph neural network effectively solves the complex and heterogeneous graph structure data,it can effectively capture the potential relationship between nodes and node representation in the graph.Therefore,how to better learn patient representation through graph neural network is particularly important to improve the accuracy of patient similarity.The works of this thesis are shown below:First,most patient similarity studies mainly utilize discrete medical entities embedding as patient’s feature representation,these structured data could be incomplete or wrong,and the structural and semantic information between medical entities is rarely considered,resulting in inaccurate patient representation.Therefore,we propose a patient similarity framework based on a medical attributed heterogeneous graph convolution neural network,named AHGCN-PS.Firstly,the framework leverages the patients’ medical entity and incorporates the patients’ medical text as the attributes of patients to obtain more integral patient information.Then,we construct a medical attributed heterogeneous information network from EMR,capturing the structural information in the network and the hidden semantic information between different nodes by selecting different meta-paths.Then,we adopt a graph convolutional neural network and a semantic attention mechanism to aggregate node neighbor information and meta-path semantic information.Finally,this thesis uses the obtained patient node feature representation for patient similarity calculation.The experiment is carried out in the MIMIC-III dataset,and the experimental results show the effectiveness of AHGCN-PS.Second,the patient views considered in the existing studies are relatively simple,only through a single patient medical information network to aggregate patient-related medical entities,the patient feature information learned is not sufficient.Therefore,this thesis establishes a patient similarity based on multi-view contrastive learning,named SCO4 PS.Firstly,this thesis generates drug views and procedure views from the constructed attribute medical heterogeneous information network,and then learns patient node representations under different views through a graph convolutional neural network.Secondly,introduce intra-view and inter-view contrastive learning mechanisms to optimize the feature representation of patient nodes in different views.Finally,aggregate the feature representation of patients in different views,and use the final patient feature representation to calculate the patient similarity.The experiment uses the MIMIC-III dataset and compares it with other baseline models.Experimental results show the effectiveness of SCO4 PS. |
