Research And Implementation Of Diabetes Assisted Diagnosis Model Based On Knowledge Graph And Deep Learning

Diabetes is one of the most important chronic diseases that threaten human health,and if not controlled in time,it may lead to various complications and directly endanger lives.Currently,diabetes is still not completely curable,but early detection and prevention can effectively improve the quality of life and reduce the rate of disability and death.However,due to the limitations of medical conditions and physicians,nearly half of the diabetic patients in China are undiagnosed,making diabetes prevention a serious challenge.In order to improve this situation,the field of smart medicine has started to widely apply deep learning technology to diabetes assisted diagnosis,even surpassing human experts in some aspects,but the lack of medical data quality and model interpretability has limited its clinical application and development.In response,this thesis considers the introduction of external medical expertise and constructs a diabetes-assisted diagnosis system based on knowledge graph and deep learning,with the main work divided into the following three parts.(1)To address the problems that existing diabetes prediction models lack practical medical knowledge support and are weakly interpretable,this thesis proposes a diabetes prediction model(TH-SAC)that integrates knowledge representation and deep learning.the model first constructs a relationship graph between physical examination indicators and test values based on the normal range of physical examination indicators,and encodes the relationship graph by a knowledge representation learning model,and then represents patient physical examination data Then,the patients’ physical examination data are represented as vectors and input to the classifier constructed by the self-attentive mechanism and convolutional neural network to achieve diabetes prediction.The experimental results show that the accuracy and recall of the model in this thesis outperform the traditional diabetes prediction model in the diabetes prediction task.This also indicates that the model in this thesis well applies knowledge representation learning and deep learning techniques to diabetes prediction,which is important for early detection and health management of diabetes.(2)To address the problem of long latency period of diabetic complications and difficulty in early diagnosis,this thesis proposes a deep neural network(KGDCP)based on knowledge graph for the auxiliary diagnosis of diabetic complications,which enables the diagnosis of complications to be jointly linked with patient physical examination data and treatment information(including patient symptoms and oral medications).The model is based on physical examination knowledge embedding and diabetes knowledge mapping to characterize patient physical examination data and diagnosis and treatment information.In order to explore the intrinsic association of physical examination indicators and measure the importance of medical entities in the diagnosis and treatment information,different attentions are introduced to encode them in the model to obtain the features of patients’ physical examination indicators and diagnosis and treatment information.Finally,the two features are fused so as to perform the auxiliary diagnosis of complications.The experiments show that the model is more accurate and more recalcitrant than other traditional prediction models,proving the effectiveness of the inclusion of potential expressions of complications in the patient’s diagnosis and treatment information and the introduction of the diabetic knowledge map.(3)In this thesis,a diabetes assisted diagnosis system is designed and implemented based on the proposed fused knowledge representation and deep learning diabetes prediction model and knowledge graph based diabetes complication prediction model.The general design,detailed design,and database design are carried out from the basic requirements.Based on the design concept of front-end and back-end separation,the system is developed using the React framework(front-end)and the Express framework(back-end).The system allows users to analyze the risk of diabetes and related complications based on information such as physical examination indicators,symptom response or oral medication.The implementation of the system demonstrates the practical feasibility of this thesis’ s diabetes and complication prediction model,and also brings convenience to both patients and healthcare professionals.