Research On Prediction Of Mortality In ICU Based On Gated Graph Neural Network

Predicting the mortality rate of patients in intensive care unit(ICU)accurately and reliably is of great significance.It enables early assessment of the patient’s condition and preliminary prediction of clinical outcome,which allows clinicians to implement early intervention in critically ill patients based on the predicted outcome and provide timely and effective treatment to prevent the deterioration of their condition.With the development of data mining technology,more and more researchers have carried out research on the prediction of patient mortality in ICU,and have made good progress.However,most of the patient data used in these studies is in a structured form.However,in real life,the relationship between patient data is often intricate,and it is difficult to express the potential relationship between patient data in a structured way.In addition,traditional mortality prediction methods can perform well when dealing with structured data,but when dealing with unstructured data,it is often difficult to model such data.In view of the above problems,this paper conducts an in-depth study.By using the graph structure to model the patient data,the data representation is more appropriate to the actual state of the patient.At the same time,it conducts training based on the graph neural network,so as to mine the complex structure information in the patient data.Specifically,this paper mainly does the following work:(1)Graph neural network is introduced into the patient mortality prediction research.At present,graph neural network has made good progress in many fields,such as recommendation system,molecular chemistry,social network,etc.,but it has not yet been applied to the direction of patient mortality prediction.This paper made an application innovation based on this starting point,using gated graph neural network to predict the mortality of patients in the ICU.(2)In this paper,a mortality prediction model GGNNAT based on attention gated graph neural network is proposed,which is improved on the basis of gated graph neural network.This model mainly takes into account the different nature of the topological structure of the nodes in the patient graph,so the adaptive attention mechanism is added to the propagation process of the gated graph neural network,by dynamically assigning weights to neighbor nodes that play different roles on the central node.To reflect the different importance of its neighbor nodes,so that the model can automatically learn more important node information in the process of information dissemination,so that it can better disseminate and update patient information and improve the model’s ability to predict patient mortality.(3)Based on MIMIC-Ⅲ database,data sets of five common diseases divided by ICD-9 code were extracted and the models were evaluated on the data sets of these five diseases.Experimental results show that the proposed model is superior to other baseline methods and can effectively improve the performance of predicting patient mortality in the ICU.