Disease Diagnosis Prediction Based On Medical Knowledge And Patient Identity Information

In recent years,with the development of deep learning,the medical field has shown a trend of informatization and intelligence,and diagnosis prediction based on medical big data has gradually become a hot topic in research.Accurate diagnostic prediction can help improve the quality of medical services in medical institutions and reduce the cost of medical treatment for patients.At present,most of the existing studies are based on the patient visit records in the electronic health record(EHR)to analyze the evolution of the disease,but the influence of medical field knowledge and patient background information on the development of the patient’s condition is not fully considered,resulting in inaccurate prediction results and insufficient interpretability of the model.Therefore,an improved RNN-based diagnostic prediction model is proposed,which introduces external medical knowledge to better explore the correlation between different types of diseases,and integrates patient identity information to better pay attention to the differences in health status among patients.The main research contents include:Firstly,in response to the fact that most existing disease feature representation methods only consider the connection between subcategory diseases and their immediate parent diseases,which fails to capture the multi-level medical associations between diseases,this paper proposes a disease feature representation method that integrates multi-level disease classification information in medical knowledge.On the one hand,it supplements medical knowledge with word frequency information of diseases,and generates preliminary feature representations of diseases based on the GloVe model;On the other hand,extracting the category relationships between diseases and constructing a medical knowledge graph with diseases as nodes,enhancing the features of subcategory diseases by replacing single paths with multiple paths,thus better mining the correlation between subcategory diseases.The testing on the CCS disease classification knowledge base shows that the disease feature vectors obtained by this algorithm are more in line with medical concept relationships.Secondly,in response to the insufficient utilization of patient identity information in existing disease diagnosis and prediction models,and the inability to quantify the differences in disease distribution among different population types of patients,this paper proposes an optimized patient background embedding method based on RNN based disease diagnosis and prediction models,and constructs a key-value pair attention network based on the feature vectors of parent diseases,Based on the multi head attention mechanism,the multi-dimensional identity information of patients is fused,and combined with the feature vectors of parent diseases,to quantitatively evaluate the impact of individual differences on the development trend of their diseases.Ultimately,based on the patient’s medical records,learning to capture their short-term and long-term health status and predict their future disease risks is achieved.Testing on the MIMIC Ⅲ dataset shows that the final results obtained by this model have improved prediction accuracy and medical interpretability.