Research On Patient Similarity Based On Network Representation Learning

With the continuous advancement of the informatization process in the medical and health field,a huge amount of medical data has been formed.These medical data are stored in various forms,how to dig out valuable information from it and apply it in the medical and health field has become a research hotspot.In recent years,network representation learning has proved its effectiveness in tasks such as network processing and analysis.Network representation learning is widely used in medical data analysis such as disease risk prediction and drug interaction prediction.Patient similarity is to learn a clinically meaningful metric based on patient-related medical records to measure the similarity between patient pairs.It is an important foundation and support for improving medical services and precision medicine.Compared with traditional statistical methods or machine learning methods for patient similarity research,the patient similarity method based on network representation learning can effectively improve the accuracy of patient similarity measurement.The work of this thesis focuses on network representation learning to solve the patient similarity problem.The main research works are as follows:First,time sensitivity is an important factor in the study of patient similarity,which can improve the accuracy of patient similarity measurement,but the time context of medical entities is often ignored in existing research.In response to this problem,we propose a novel time-aware patient similarity framework based on network representation learning,named T-PS.Applying network representation learning to the study of patient similarity,taking full account of the timeliness of the medical entity,the information of patients and their medical entities is extracted from EHRs,and the temporal medical entity association graph is created.The topological structure information in the graph is learned by using the network representation learning algorithm to obtain the feature vector representation of patients and their medical entities,and the similarity of patient pairs is further calculated by using the feature vectors of patient nodes.Experimental results on real-world ICU dataset MIMIC-III showed the accuracy of T-PS.Second,when conducting medical network analysis,fully considering the rich edge relationships and the attribute information of the nodes in the target network can improve the accuracy of node vector representation in the medical network.However,the existing research often ignores the rich edge relationships and nodes attribute information in the network.In response to this problem,a patient similarity framework based on medical attributed heterogeneous network representation learning is proposed,which named MAHIN-PS.At present,most of the work is only based on the topology training vector of the network,ignoring the attribute information of the nodes.The framework fully considers the topology structure and attribute information in the medical attributed heterogeneous network,maps the medical entity nodes in the medical attributed heterogeneous network to a low-dimensional vector space,and makes the vector representation of nodes in the medical attributed heterogeneous information network satisfy the consistency of network topology information and node attribute information.The learned patient vector representations are input into the patient similarity measurement function to calculate the similarity between patient pairs.Through extensive experiments on the real-world ICU dataset MIMIC-III,the experimental results show the effectiveness of MAHIN-PS in patient similarity.