Design And Implementation Of Medical Insurance Fraud Detection System Based On Deep Learning

While the popularity and development of medical insurance has brought basic guarantee to people’s health,the process has gradually given rise to some medical insurance frauds and irregularities.Fraudsters have been making illegal profits by fraudulently obtaining medical insurance funds through fictitious medical consultations,which seriously undermines the interests of participants and the safety of medical insurance funds.Therefore,we need to effectively and efficiently strengthen the regulation of medical insurance funds,promptly and accurately uncover potential fraud patterns and bring them to justice,and return the people’s medical insurance funds to the people.Unlike finance and e-commerce,medical insurance anti-fraud requires strong professional knowledge and is characterized by high concealment,high data dimension and many entities involved,which makes medical insurance fraud detection a challenging task.Based on this background,this topic is based on deep learning technology to research and explore different granularity of medical insurance fraud detection tasks,which mainly contains the following three aspects of work.First,a record-granularity medical insurance fraud detection algorithm based on graph and contrastive learning is proposed.First,a medicine graph is constructed based on the co-occurrence frequency of drugs,and the embedding of drugs in the treatment records is achieved by the powerful graph neural networks.Second,we introduce self-supervised method based on the idea of "same diagnosis similar treatment" and construct positive and negative sample pairs to calculate the contrastive learning loss function,and then we jointly train the model.The effectiveness of the algorithm is demonstrated by comparison experiments,analysis experiments and ablation experiments.Second,a dynamic graph-based algorithm for patient-granularity medical insurance fraud detection is proposed.The insurance record is modeled as a dynamic heterogeneous graph containing four entities and five relationships,and a representation learning model of dynamic graph based on attention mechanism is designed.In terms of spatial features,this model designs a static layer to divide each snapshot graph into static subgraphs by edge type,and feature aggregation is performed on the neighbor nodes of the subgraphs and on each subgraph.In terms of temporal features,this model designs a dynamic layer to aggregate the features of each snapshot graph.The effectiveness of the algorithm is demonstrated by the experimental results.Third,a prototype medical insurance fraud detection system is designed and implemented.The system encapsulates deep learning algorithms,achieve the function of user management,data management,data analysis,and anomaly detection modules.It reduces the user’s access threshold,and provides a convenient and efficient anti-fraud visualization interface for users.