Research On Hospital Readmission Prediction Based On Feature Combination And Medical Graph Of Electronic Health Records

Hospital readmission prediction is an important research direction in the field of health care,and has gradually attracted the attention of scholars and industry.Accurate and personalized hospital readmission prediction has great application value for the promotion of the overall people's health care,medical public service level and national medical insurance overall planning.With the continuous improvement of medical informatization,the medical field has gradually accumulated a large amount of data,which provides a better data foundation for hospital readmission prediction research.Electronic Health Records(EHRs)collect medical health data from individuals at different times in an electronic format,whichserves as an important vehicle for data-driven medical research.Electronic health records have a variety of properties,such as a wide range of sources,multiple data types,and high dimensions,which contain many information related to hospital readmission prediction.The traditional hospital readmission prediction method firstly extracts features related to the hospital readmission prediction by the expert,which is used to represent the patient's hospital readmission environment.Due to the high dimensionality of electronic health records,feature explosions can occur when features are extracted,and the correlation between a large number of features and hospital readmission cannot be judged.If the number of extracted features is too small,the accuracy of the prediction method cannot be guaranteed.In order to solve the above challenges,it is necessary to accurately find the feature combinations mostrelevant to the hospital readmission behavior in the large and complex electronichealth records for hospital readmission prediction,so as to ensure the accuracy of theprediction and improve the performance of the method.In addition,electronic healthrecords have the nature of temporality and relevance,and medical events can berepresented as medical event sequences.People's health is not only related to thecurrent physical condition,but also closely related to the previous diseases,treatmentand medication.The development path and treatment path of the disease jointly determine the direction of health,such as the possibility of suffering from a certain disease,hospital readmission,and taking certain drugs in the future.Therefore,medical health records can be expressed in the form of a medical graph,which can more vividly express the patient's medical history.The medical graph not only reflects the temporality and relevance of electronic health records,but also solves problems such as data sparsity.In view of the above problems,this paper studies the patient's hospital readmission prediction from two aspects of the optimal feature combination and the medical graph,and proposes two patient hospital readmission prediction models.The main work of this article includes:1.A hospital readmission prediction method based on optimal feature combination(A Hospital Readmission Prediction Method Based on Optimal Feature Combination,MulFeature).This method uses a genetic algorithm:the adapted survives,and the unfit is eliminated,to search for the optimal feature combination for hospital readmission prediction.Specifically,this article first extracts and selects multiple features for each patient to represent the patient's hospital readmission environment;then each patient's features are used as input to select the optimal feature combination of hospital readmission based on a multi-objective fitness function of genetic algorithm;finally,ensemble learning algorithms are used to classify patients and determine whether patients will be hospital readmission within a certain time interval.This method is validated in a specific clinical scenario based on real-world data sets.The experimental results show that the method is effective compared to the traditional method.2.A hospital readmission prediction method based on patient medical graph(Predicting Hospital Readmission From Longitudinal Healthcare Data Using Graph Pattern Mining Based Temporal phenotypes,Tephe).The method is based on the graph theory and gets the temporal phenotypes of each patient,and uses temporal phenotypes to predict whether each patient will be hospital readmission.In particular,each patient's sequence of medical events is first represented by a medical graph that captures the temporal relationship between medical events and makes the original data more intuitive;based on graph-pattern mining,we define more prominent frequent subgraphs as a temporal phenotypes,and this allows us to better understand the evolutionary patterns and pathways of the disease.In addition,an improved greedy algorithm is designed to find the optimal expression coefficient for frequent subgraphs for each patient.Finally,random forest were used for prediction tasks based on the optimal expression coefficient.The experimental results show that our proposed method is more accurate in predicting tasks than the baseline method.This paper uses the real electronic health records of patients collected in several general hospitals in the period from 2011 to 2016 to evaluate the proposed two methods for hospital readmission.The electronic health record includes three parts of data:clinical diagnosis and treatment data,health examination data and medical insurance data.In order to verify the hospital readmission prediction model based on the optimal feature combinations,the electronic health records of schizophrenia patients and patients with coronary heart disease were taken as experimental data.Experimental results show that this method is effective compared to the baseline method.In this paper,the hospital readmission model based on the patient's medical image is validated in a real clinical scenario.The electronic health records of patients with coronary heart disease are taken as experimental data.Compared with the baseline method,the proposed method in this clinical scenario is more competitive in prediction tasks.