Clinical Assisted Diagnosis Of Electronic Medical Records Based On Heterogeneous Information Network

The continuous accumulation of electronic medical records and the increasingly mature data analysis technology have laid the foundation for the realization of intelligent healthcare.Deep learning technology has also been widely applied to clinical assistant diagnosis tasks and has achieved promising predictive results.However,existing prediction methods generally rely on patients’ historical medical records,and seldom consider the potential association between entities contained in electronic medical records.To address this challenge,we construct a deep learning prediction model by combining graph neural networks with electronic medical records,to assist doctors in improving diagnostic.We have mainly completed the following research:Firstly,we propose a patient feature embedding method based on relational representation learning(RRL).To address the problem of poor predictive performance due to the imbalanced distribution of symptom data in electronic medical records,we extract "patient-symptom" relationships from electronic medical records and construct a patient medical records network.An external medical knowledge graph is constructed by utilizing "disease-symptom" relationships extracted from an external medical knowledge base to solve the problem of imbalanced electronic medical records.The thesis effectively integrates contextual information such as patient medical condition descriptions and physiological records into an information matrix,which is used as partial features of patients.We design an aggregation function to aggregate neighbor nodes information and obtain a more comprehensive representation of patient node embedding.Compared with graph representation learning methods such as Deep Walk and LINE,the RRL method achieves 6.7% improvement in F1 in the classification task.Secondly,a clinical assistant diagnosis based on heterogeneous graph medical records attention network(HCAD)is proposed.To address the heterogeneity of electronic medical records,this thesis constructs a heterogeneous graph medical record network and designs a hierarchical attention mechanism to identify the importance of nodes and different semantic relationships.The highly representative patient node embedding representation is obtained through weighted fusion to accurately predict diseases.The effectiveness of the HCAD model is verified through experiments in classification and clustering tasks.Compared with baselines such as HIN2 vec and DHNE,the HCAD achieves at least a 4.86% improvement in F1 and 8.23% improvement in NMI.Thirdly,The PM-GSL model,based on heterogeneous graph structure learning,is proposed for clinical assistant diagnosis of diabetes.Noise interactions and missing data in heterogeneous graph can propagate erroneous information throughout the network via the cascading effect of iterative mechanisms,affecting the overall prediction performance of the model.Therefore,this thesis proposes the PM-GSL to model complex interactions in electronic medical records,construct a patient multi-relational heterogeneous graph,and generate three candidate graphs from two aspects of node features and higher-order semantics heterogeneous interactions,which are fused to generate a new heterogeneous graph.Under the supervision of disease prediction tasks,the graph structure is jointly optimized with GNN.Compared to the optimal benchmark model,the prediction results of PM-GSL on both datasets have improved,with an average AUC increase of 7.26%.