Research And Implementation On Knowledge-guided Intelligent Consultation-based Diagnosis

As an important method of keeping people healthy,healthcare is developing with experts’ practice and accumulation.Whether past or present,it always plays an irreplaceable role.As one of the diagnosis methods in modern medicine,consultation-based diagnosis offers important support for final diagnosis and prescription.It identifies diseases according to patients’ self-descriptions and doctors’ inquiries,being an important method to understand patients’ conditions.Recently,the techniques in artificial intelligence have been developed rapidly,and the methods based on Neural Network(NN)show excellent performance in the healthcare domain.Applying NN-based methods to intelligent consultation-based diagnosis helps to promote the diagnosis efficiency and to lighten the burden on doctors.Through investigation of the related works,we found the following problems existing in intelligent consultation-based diagnosis:(1)Current studies are mainly based on deep reinforcement learning and adopt one-hot for state representation that is unable to reflect the relationship between actions.(2)With limited data,it is needed to study how to effectively mine features.(3)The existing knowledge graph based methods take advantage of features of nodes but ignoring the information from the graph structure.To deal with the above problems,this article proposes the following methods:(1)building a basic Deep Q-network(DQN)and proposing an embedding-based method for state representation which provides a more reasonable representation for the state;(2)proposing a self-supervised learning task and a self-supervised learning model called DOM to enrich the embedding and make DQN focus on the association of actions;(3)based on DOM,proposing a self-supervised learning model taking advantage of knowledge graph and graph convolutional network,which is named as GBM and further improves the embedding with the structural information from the graph.We conducted experiments on MZ dataset and DX dataset to verify the effectiveness of the embedding-based method for state representation and the self-supervised learning models including DOM and GBM.Besides,the DQN based on our methods is compared with the related methods proposed in recent years.Experimental results show that our methods contribute to enrich and improve the embedding,improving the performance of DQN on the intelligent consultation-based diagnosis.