Drug Prediction Based On Graph Neural Network Methodology And Application Research

With the advancement of medical information technology,electronic medical record system has been rapidly popularized and developed,and accumulated a huge amount of medical data,which contains a lot of valuable information.The analysis and mining of electronic medical record data and the drug prediction algorithm modeling patient information can predict the drugs needed by patients based on their clinical information and assist doctors in diagnosis and treatment,which has important research significance and application value and is a hot research topic in the medical field.However,the task of patient drug prediction based on electronic medical record datasets still faces many challenges.The electronic medical records of patients contain various types of data such as disease diagnosis codes,surgical treatment codes,drug codes,physiological indicators,etc.,which are characterized by many data types,complicated relationships between concepts and high dimensionality,and the representation of these medical concepts by traditional deep learning methods has sparsity.The degree of drug interactions in the predicted drug combinations is serious,and the safety of drug use is missing.Meanwhile,the medical records of patients with multiple visits show strong non-equilibrium,and most methods are missing the extraction of temporal features from medical records.To address these problems,this thesis makes further improvements based on the existing work,and the main innovations are as follows.In this thesis,we propose a drug prediction model based on graph embedding representation combined with memory networks,which reduces the prediction of drug interactions by constructing medical concepts into graphs,introducing specialized medical ontology knowledge and drug interaction knowledge to construct relationships,respectively,using graph neural networks to learn the association between medical concepts.It also uses a memory network to store historical patient medication information to reduce information loss due to long medical record sequences.Experiments on a real public electronic medical record dataset confirm that the model in this thesis is effective in reducing drug interactions and maintaining a high accuracy rate.Since the proposed first model only constructs knowledge graphs for the representation of diagnosis,surgery,and medicine separately,it ignores the potential complex connections among these medical concepts.Therefore,this thesis further introduces more specialized medical relationships and medical semantics,and proposes a prediction model based on heterogeneous graph representation for its characteristics of multiple data types and complicated relationships between concepts,constructing medical heterogeneous graphs of co-occurrence relationships between all medical concepts and medical ontology relationships,and the model uses two different levels of graph attention networks designed to jointly learn medical concepts and their relationships.The time interval information of patient medical record sequences is also introduced into the model,and the time-interval-aware attention mechanism is used to obtain more temporal features for the time-dependent problem of patient medical records.Multiple experiments on publicly available electronic medical record datasets confirm that this model can improve the prediction quality and confirm the effectiveness of this model.Finally,requirements analysis,system design,and system testing are performed for the medical assistance system.The improved model was also explored for ground application,and the usability of the algorithm was verified by applying it to the drug recommendation module of the medical assistance system.