Representation Learning Methods Of Medical Events Based On Graph Mining

As an important part of electronic health records,medical events mainly refer to a series of treatment data generated by patients during their admission,including various types of data such as disease diagnosis codes,drug codes,and physiological indicators.Representation learning research on medical events using artificial intelligence technology is a hot research topic nowadays,and is also the basic work for disease prediction tasks.However,due to the heterogeneous and high-dimensional nature of electronic health record and the presence of a large number of missing values,which are relatively time-dependent.It is challenging to obtain a good representation of medical events.To address the above problems,this paper develops a study on the representation learning method for heterogeneous medical events,and establishes a medical event representation method based on temporal heterogeneous graphs and temporal prediction networks,which fully explores the relationships between events while representing medical events,and fuses the temporal characteristics of events to achieve multiple disease prediction.The innovations of this paper are as follows:(1)A medical event representation learning method based on Mogrifier-Time Long Short Term Memory is proposed,which possesses the ability to handle variable-length interval sequences and eliminate the strong dependence of clinical data on timestamps.Firstly,considering the non-uniform time interval of medical events in a patient’s treatment sequence,different types of medical events in the same time period are vectorially aggregated;secondly,the temporal gating module is fused with the forget gate and memory cell in the Long Short-Term Memory Network respectively,to indirectly characterize the patient’s treatment time and unequal visit interval;finally,the Mogrifier-Long Short Term Memory Network is introduced to enhance the contextual modeling capability of the model.(2)A medical event representation learning method based on dilated convolution and attention is proposed,which possesses the ability to fully extract semantic features and improve the computational efficiency of the model.Firstly,the electronic health record is modeled as a sequence of equally interval events;secondly,the dilated l convolution and attention mechanism are introduced to extract semantic features from the event sequence,realize data dimensionality reduction while screening important features;finally,introducing two different temporal fusion mechanisms to solve the problem of partial temporal information loss during the convolution process and capture the temporal characteristics of patient data for disease prediction.(3)A learning method for medical event representation based on relation-attention graph is proposed,which integrates the association information between heterogeneous medical events.Firstly,different correlation degree functions are set to fully explore the relationships between different types of nodes;secondly,constructing a graph attention network model by attention and assign attention weights to the obtained relationship sequences and filter important features;finally,temporal features are combined with event relationships to achieve disease prediction using a temporal prediction network.Finally,by conducting a large number of comparative experiments on the MIMIC-III public database and evaluating several metrics of the model comprehensively,the results show that the model in this paper can achieve higher accuracy in medical event representation learning and effectively realize the diagnosis of clinical diseases.