Research On Hospital Readmission Prediction Based On Patient And Disease Bipartite Graph

Hospital readmission prediction research based on electronic health records is a hot medical research trend today.Accurate hospital readmission predictions not only help patients understand their own physical health and guide them healthy living behaviors,but also have great significance and applications for improving the future level of public medical services,improving the overall planning of national medical insurance,and reducing social medical costs.The application of medical big data and artificial intelligence methods provides an opportunity for more accurate hospital readmission predictions.There are many approaches to predicting hospital readmissions based on machine learning.In recent years,with the rise of deep learning,people use recurrent neural networks or its variants to use serialized medical health records to make hospital readmission predictions.However,most of these existing methods only use the patient's own feature information to make predictions,and less consider the potential relationship between patients,which makes the prediction effect not robust for similar populations.Especially for patients with a short medical history,that is,patients with fewer case data,it is difficult to predict hospital readmissionAiming at the problem that the medical history data of patients with few cases are difficult to predict hospital readmission,this paper mainly makes use of the feature information of other potentially relevant patients to assist in the prediction of hospital readmission and improve the prediction effect.Moreover,even for patients with a long history of medical treatment,the prediction effect can be improved by using the feature information of other related patients.Based on this idea,this paper proposes a method for predicting hospital readmission based on the patient and disease bipartite graph.This method uses the bipartite graph of patients and diseases to establish the relationship between patients and diseases,which can express the patient's historical disease information and the indirect connection between patients with the same disease,and can make full use of the patient's basic health information This method first constructs a patient and disease bipartite graph based on the patient's medical records.The patients' basic health information is used as the initial features of the patient nodes,and the disease representation vector is learned as the initial features of the disease nodes using the medical concept embedding method based on the patient medical temporal graph designed in this paper.Then,in this hospital readmission prediction method,a patient and disease bipartite graph embedding generation algorithm is designed to learn the representation the medical bipartite graph constructed.This algorithm aggregates node neighborhood information for each node layer by layer,and finally the aggregation information of patient nodes is used to predict hospital readmission.In this paper,the real electronic health records of patients collected from multiple general hospitals are used to evaluate the effectiveness of the proposed method.The data includes medical history information and patient health examination information.Experimental results show that the proposed method has better prediction effect than the baseline methods.