Research Of Traditonal Chinese Medicine Inquiry Modeling Based On Deep Learning And Conditional Random Fields

Traditional Chinese Medicine (TCM) inquiry refers to the physician asking questions of the patient or the patient’s companion. The purpose of the approach is to diagnose the disease by understanding the information related to the disease. There is a lot of subjectivity in this diagnostic process. Multi-labeled learning is used to solve the problem that objects are usually associated with multiple labels simultaneously in many real-world applications. During research in clinical practice in TCM, we found that patients may be strong relevance to multiple syndromes, and the syndrome often happen not a single. This means that the objectification and digitization studies of inquiry in TCM can be realized by employing multi-label learning methods.In this paper, we built syndromes classification model of Chronic Gastritis (CG) in TCM from the relationship of symptom and syndromes, and the relationship of syndromes and syndromes using the sample of CG data. During studying the relationship of studying symptom and syndrome, Severals multi-label learning methds based on Binary relevance and deep belief nets or deep machine Boltzmann Machine are modeled with building a process of mutual learning between multi-layer features by introducting some hidden layer that are some higher levels representation, representing more and more abstract functions of the raw input. During studying the relationship of studying syndrome and syndrome, a discriminant of conditional random fields (CRF) was employed in multi-label learning to exploit of the hidden correlation among syndromes of CG based on building an undirected graph between syndromes and to construct a syndromes diagnostic model for CG in TCM. Several feature construction methods about dimensionality reduction and classification are used to construct the node features and edge features of CRF. In addition, an undirected graph structure of syndromes and several confidence values on each edge are analysed rationality combinied with TCM theory. In the experiment, both multi-label learning methods show a good performance against other well-estabised methods with the average precion up to0.833and0.837respectively. It will provide reference for syndrome standardization and objection of...