Study On EHR-oriented Knowledge Graph Based Implicate Disease Information Mining And Application

With the development of medical informatics in China,the electronic health record(EHR)is becoming more important in clinical practice.The EHR contains implicate disease information like measurement abnormality,variation trend of the test indexes and disease risk factors,which is significantly valuable for clinical application.However,the valuable information is not properly used during routine practice.Specialists focus mainly on disease-related information relevant to their specialties and may be insensitive or less concerned about disease information beyond their department.In the meantime,patients’ visiting multiple hospitals causes EHR data fragmentation.The disease information within a single hospital is not intact and the separated information is hard to achieve clinical value.These factors cause the lack of valuable disease information usage during clinical practice,which results in delayed diagnosis of cross-departmental disease risks,unproper clinical decisions and harming of healthcare quality.This study proposed an EHR-oriented knowledge graph method for implicate disease information mining and application.The method uses deduction knowledge graph reasoning to mine implicate disease information and identify disease risks that were ignored by clinicians.The method provides complementary decision support and improve medical data utilization.The main innovation of the study is:The study proposed an EHR-oriented knowledge graph construction method.The multidisciplinary medical knowledge and various EHR data domains are normalized and represented under a top-level ontology structure,based on standardized medical terminology,medical knowledge base and common data model.Through development of semantic mapping rules and tools,the medical knowledge metadata is transformed into relation network.The knowledge graph engine and application tools support the function of the knowledge graph system.The study designed 467 top-level classes,62 semantic mapping rules and 46 function queries.The method supported the construction of an EHR knowledge graph system with 4.79 million concepts and 35.31 million medical relationships.The study proposed an implicate disease Information mining method based on EHRoriented knowledge graph.The scattered table-based EHR data is created as a patientcentralized patient information model in semantic form,to strengthen the relationship between clinical elements in EHR data.The system performs deduction reasoning on patient information model to create region of interest in EHR data and identify critical clinical findings and disease risks buried in non-used medical information.The system provides traceable and explainable clinical decision support for clinicals to understand the importance of the information and make decisions in-time.The system was evaluated through an application study on unconsidered chronic kidney disease(CKD)warning for non-nephrology clinicians.The application study found 71679 patients with CKD risks in EHR data during March 2007 to May 2019 at a tertiary hospital.The follow-up study of 5439 patients showed 82.1%of patients meeting the CKD diagnosis criteria and 61.4%of patients requiring high attention were CKD positive.The system could identify the risk of CKD 2 years before the follow-up on average.The study proposed a method to utilize multicenter disease information in data safety environment based on EHR-oriented knowledge graph system.The study uses local knowledge graph reasoning combining with multicenter summary of intermediate reasoning results to utilize fragmented disease information in multiple hospitals,without sharing original EHR data.The intermediate reasoning results are synchronized among multiple systems through a blockchain network with data and identity encrypted.The system can utilize fragmented information buried in multiple hospitals for in-time decision support.The system was evaluation through an application study among 3 tertiary hospitals using EHR data from March 2008 to November 2020.The system identified 124 unconsidered CKD patients who were not able to be found using only single hospital data.The overall accuracy through clinician assessment is 85.71%.The proposed method of the study utilizes EHR-oriented knowledge graph to find implicate disease information from non-used and neglected medical data.The method identified important risks of cross-departmental diseases and help clinicians to realize and manage the diseases in time.The method helps resolving omission of EHR data usage and utilize EHR data for improving healthcare service quality.