Research On Assisted Medical Decision-making Method Based On Medical Knowledge

With the rapid growth of medical data and precision medicine,using Electronic Medical Record(EMR)mining technologies to find potential diagnosis and treatment patterns,and further assisting doctors have gained more and more attention.Although the EMR contains abundant patient information,the ideal decision effect can not be obtained only based on it.This is because that much of the medical knowledge,beneficial for diagnosis and treatment,is difficult to obtain through EMR data mining,such as the relationships between diseases and the interactions between drugs.Due to the lack of medical knowledge,current diagnosis and treatment methods based on EMRs suffer from large-scale disease space,rare diseases,and the safety of drug combinations.To this end,a diagnosis method based on the hierarchical tree of diseases is proposed for solving the dilemmas of large-scale disease space and neglected disease association.Benefitting from the knowledge of disease associations contained in the hierarchy tree,the proposed method can narrow down the candidate space of diseases and further reduce the diagnosis difficulty.Correspondingly,the diagnosis effect can be improved.Aiming at the poor diagnosis accuracy on rare diseases caused by the imbalanced distribution of diseases,we design a fewshot diagnosis method based on the heterogeneous EMR graph.By constructing the heterogeneous EMR graph and fusing medical entities and relations into it,the richness of patient information can be enhanced.To ensure the safety of drug combination prediction,one treatment decision-making assist method based on a dynamic drug knowledge graph is probed.Utilizing the synergy and antagonist relationships between drugs,effective and safe drug combinations can be generated to assist doctors in making safe treatment decisions.Consequently,this research presents several assist methods based on medical knowledge to help doctors make decisions,from the practical problems faced by medical decision-making.Overall,the main contents of this research are as follows:Firstly,we research one method based on the hierarchy tree of disease to assist the diagnosis decision-making.In order to narrow the candidate diseases space and reduce the difficulty of disease selection,a thick-to-fine disease path generation algorithm based on reinforcement learning is proposed.It performs multi-step disease selection,continuously narrowing the candidate disease space until finding the most fine-grained disease codes as the diagnosis results.The correlations can also be captured through the path message passing mechanism for father-child,and sibling-sibling diseases.The experimental results show that the proposed method can refine the diagnostic procedure by exploiting the hierarchical association between diseases,which provides a new idea for assisted decision-making.Secondly,we research an assist diagnosis method for rare diseases based on the heterogeneous EMR graph.Owing to the unbalanced characteristics of disease distribution,most diagnostic methods have poor performance on rare diseases.Therefore,a few-shot learning method based on graph contrastive learning is proposed to improve the diagnostic ability of rare diseases.The method constructs a heterogeneous graph based on the EMR and injects medical entities and relationships into the graph.Meanwhile,a few-shot learning architecture is used to transfer the classification knowledge from frequent diseases to rare diseases.Validated on a real EMR dataset,the experimental results show that the proposed graph contrastive learning combined with few-shot learning can significantly improve the diagnostic effect on rare diseases.Thirdly,we research one drug combination prediction method based on the dynamic drug knowledge graph.To avoid producing adverse drug combinations,a drug selection strategy based on graph convolution reinforcement learning is proposed.The method adopts reinforcement learning to select drug combinations,and the synergistic and antagonistic relationships among drugs are taken into account adaptively.The relational graph convolutional neural network is used to ensure the effectiveness and safety of selected drug combinations,by seizing multiple relationships between drugs simultaneously.Validated on a real EMR dataset,the optimal strategy of drug combination prediction can be found,assisting doctors to make safe and effective treatment decisions.In terms of theory,the research technologies consist of a disease path generation method based on reinforcement learning,a patient representation method based on contrastive learning of EMR graph,a few-shot learning method for rare disease diagnosis,and a graph convolutional reinforcement learning method for drug combination prediction.By taking advantage of medical knowledge,several challenges faced in medical decision-making are addressed effectively.In terms of application,the proposed method can be integrated into the clinical decision support system to help doctors mine more efficient and accurate patterns for diagnosis and treatment.Furthermore,it can prompt doctors to make more informed decisions.