Knowledge Fusion Research For Clinical Decision Support

With the advent of aging population,people’s demand for medical and health services is becoming more and more frequent,and doctors,as the leading providers of medical services,are receiving and needing to call on an increasing amount of information,so it is very easy to produce medical errors such as missed diagnosis and misdiagnosis.Clinical decision support services can not only help physicians reduce errors in clinical diagnosis and treatment plans and improve medical quality and efficiency,but also improve medical outcomes,thereby indirectly controlling medical expenses and reducing medical costs.Clinical decision-making is a knowledgeintensive and specialized task that requires the effective and comprehensive use of health information resources.Health information resources include patient health information and disease knowledge.Patient health information is scattered in many systems,including hospital testing systems and electronic medical record systems,and is mostly in the form of natural language that is difficult to be processed directly and there are often many different expressions of the same knowledge concept;at the same time,relevant disease background knowledge is also scattered in various aspects,such as medical textbooks and clinical guidelines.Therefore,the multi-source and heterogeneous nature of health information resources seriously limits the efficiency of clinical knowledge utilization and the effectiveness and quality of clinical knowledge services.Knowledge fusion provides a good solution for this purpose.Facing the great challenges brought by multi-source heterogeneous data in the clinical setting,the relevant methods,techniques and tools of knowledge fusion can realize the deep-level interactive processing of health information resources at the semantic level,so as to fuse them into new and integrated knowledge objects,build a more comprehensive knowledge system,and provide users with accurate and perfect knowledge services and help decision makers make more scientific decisions.It helps decision makers to make more scientific decisions.Through summarizing and sorting out relevant research results at home and abroad,this study considers the following problems in the current research:(1)The current research on the framework of knowledge fusion mostly favors the framework construction under a single perspective,and the data characteristics of target information resources and the application purposes of domain services both put forward different requirements on the fusion process,and there is a lack of a fusion framework that starts from the combination of target resources and application purposes.information resources are rich in structure types and highly specialized,it is difficult for the knowledge fusion framework of general domain to produce effective guidance for clinical and health domains.(2)At the present stage,the pre-training models represented by BERT have greatly improved the performance of knowledge extraction,but the existing knowledge extraction models are mostly based on the extraction models of single oriented flat entities or nested entities.In the biomedical field,the entity types are very rich and contain many types of entities such as flat entities,nested entities and discontinuous entities,and there is no joint extraction model for many types of entities.The problem of "imbalance" caused by the number of different entity types and the difficulty of extraction in the process of entity extraction has not received much attention from scholars.However,the existing graph convolutional neural networks mostly use word co-occurrence information and sequence feature information to construct text graphs,and fail to make full use of more diverse and deep text information;at the same time,in the process of knowledge association,the research has to a certain extent ignored the utilization of the associated knowledge during the operation of the model,thus limiting the performance of the model.The performance improvement of the model is thus limited.In the clinical decision support research,for the multi-label classification research for multi-patient decision making,the correlation between the classification labels in the classification process is still not sufficient,meanwhile,the application of the existing three-branch decision making research in the medical field mainly focuses on improving the accuracy of decision making through the three-branch decision making model,seeking to get higher performance of evaluation indexes,while ignoring the correlation between the "uncertain" information in the original data and the "uncertain" information in the original data."In addition,most of the studies have focused on the combination of three-branch decision making and single-label classification,but not on the combination of three-branch decision making and multi-label classification.To address the above problems,this paper aims at the clinical decision support problem,with the ultimate goal of satisfying users with high-quality and high-efficiency knowledge services,and under the guidance of multidisciplinary theories,starts from the inner logic and overall framework of knowledge fusion,utilizes text mining,natural language processing,deep learning and other methods and technologies,and follows the research of knowledge extraction-knowledge The study investigates the feasible solution of knowledge fusion for clinical decision support along the research line of knowledge extraction-knowledge association-knowledge application.The specific research contents include:(1)A knowledge fusion framework is constructed from multiple perspectives,and the relevant theoretical results of knowledge fusion are enriched.Based on the previous work,this study analyzed the requirements of knowledge fusion from two aspects of the functions of knowledge fusion and the psychological needs of users,clarified the objectives,principles and challenges of knowledge fusion,and then,under the guidance of system theory,knowledge space theory and DIKW hierarchy theory,divided into three perspectives of logic,dimension and process from the level of concept and logic,the level of structure and function,and the level of method technology and process The knowledge fusion framework is constructed respectively,which enriches the relevant theoretical results of knowledge fusion and provides certain theoretical reference and support for the subsequent knowledge fusion research.(2)A unified named entity recognition model UWPR-NER,which can simultaneously recognize many different types of entities,is proposed to address the characteristics of rich types of clinical entities.The new loss function is designed by gradient coordination mechanism,category similarity and category number ratio to further tune up the performance of knowledge extraction for the imbalance between the number and classification difficulty generated in the classification process.(3)A dynamic knowledge association model based on graph neural network is proposed.The study constructs a text graph from three dimensions of sequence,syntax and topic,and uses the gravitational model GM algorithm to weight the nodes,utilizes richer and more diverse text information,and performs local association based on BERT-whitening combined with cosine similarity algorithm,and finally constructs a dynamic graph neural network containing spatial convolution layer and time-domain convolution layer based on the attention mechanism to generate the final association information,which further optimizes the model structure,makes full use of the knowledge that has been associated during the model operation,and improves the accuracy of knowledge association.(4)A three-way decision model for multiple diseases is proposed.The study weighted the diagnostic indicators involved in the decision-making process by the method of superior order diagram,distinguished the importance of common clinical manifestations,episodic clinical manifestations and auxiliary examination results in the disease diagnosis process,and provided theoretical and mathematical support for the utilization of "uncertain information" based on decision rough set theory.Finally,the Node Rank algorithm is used to quantify the contribution of different diagnostic indicators to disease diagnosis and construct the relationship nodes of "standard disease(GS)-gold standard" and "standard disease(TS)-typical symptoms" to better control the output of the model.The problem of coarse decision granularity in traditional twobranch decision making studies and multi-disease decision making in clinical decision support is solved.For the optimization of methodological techniques in the above content,data experiments,comparison experiments and ablation experiments were conducted on selected relevant data sets,and the different methodological aspects of the study were evaluated accordingly by different evaluation index settings and expert evaluation methods to highlight the feasibility,superiority and necessity of the optimized methods.