Research On Disease Prediction And Medication Recommendation Based On Electronic Health Record

With the continuous development of digital medical technology,the number of hospitals using electronic health record system is also growing.The emergence of electronic health record provides conditions for machine learning technology to play a role in the medical field.At present,electronic health record has been used for many research tasks,and disease prediction and medication recommendation are two typical tasks.The disease prediction aims to predict potential diseases of patients,so as to achieve the purposes of assisting doctors in diagnosis and reminding patients to prevent.Medication recommendation aims to recommend necessary drugs for patients,so as to achieve the purpose of assisting doctors to prescribe drugs and assisting patients in treatment.Both of these tasks are integral parts of the smart medicine.This thesis studies the task of disease prediction and drug recommendation separately.In recent years,the task of disease prediction and medication recommendation has attracted much attention,and various excellent models have been born.The existing disease prediction methods all assume that the patient’s historical information is accurate.However,in the real medical scene,doctors are likely to have missed diagnosi,resulting in errors in patients’ electronic health records.These erroneous historical information may make the model misunderstand the patient’s health status.Existing medication recommendation methods often encode patient information first,and then recommend drugs based on the encoded representation.However,these models ignore the relationship between drugs used by patients.For example,patients with chronic diseases will use the same class or even the same drug for a long time.To solve the above problems,this thesis propose a disease prediction method based on joint learning and a drug recommendation method based on replication mechanism.In the task of disease prediction,since the patient’s medical history,symptoms,drugs and other information are recorded in detail in the text electronic health record,and the information such as symptoms is often directly related to the disease,this thesis considers introducing the text electronic health record into the process of disease prediction to supplement the historical diagnostic information,so as to achieve more accurate disease prediction.In the drug recommendation task,it is found through analysis that a large proportion of the drugs currently needed by patients have been used before,while the existing work ignores this correlation.Therefore,in this thesis,we consider using the historical information from the perspective of drugs and modeling the whole task as a sequence generation.The main contributions of this thesis can be divided into the following three parts:·In this thesis,a disease prediction method based on joint learning is proposed.The method comprises the following steps of: firstly,extracting symptoms from historical text health record;secondly,constructing a heterogeneous graph consisting of symptoms,diseases and patients according to the extraction result,and carrying out information transmission based on the graph;next,based on the representation of disease and symptom nodes after information transmission,predicting the potential missed diagnosis disease of the patient,and obtaining the representation of the patient according to the prediction result;finally,the historical diagnostic information of the patient is encoded,and the encoded result is combined with the patient representation obtained in the previous step to carry out the final disease prediction.· This thesis proposes a meidcation recommendation method based on copy mechanism.The method can be divided into two parts: a basic model and a copy module.The basic model performs medication recommendation based on the current information of the patient,which first encodes the current diagnosis and procedure information of the patient,then encodes the relationship between drugs,and finally iteratively decodes according to the encoding result and the recommended drugs to generate the recommendation.The copy module introduced the patient’s historical information on the basis of the module,and designed a hierarchical selection mechanism to determine whether each historical drug could be reused,and combined the results of the basic model to obtain the final recommended drug.·In this thesis,comprehensive experimental analysis is performed on real electronic health record data set,and the effectiveness of the disease prediction method and meication recommendation method proposed in this paper is verified.