Research On TCM Prescription Recommendation Based On Graph Neural Network

As an outstanding achievement in ancient Chinese science,the development of tra-ditional Chinese medicine(TCM)in modern times faces disadvantages such as weak foundation,insufficient inheritance,and insufficient innovation.The current machine learning and artificial intelligence technologies can provide strong support for TCM development.An intelligent TCM auxiliary diagnosis and treatment system based on TCM prescriptions is helpful for efficient mining of TCM experiences.Besides,inte-grating TCM with modern medicine can help study the action mechanism of herbs and accelerate TCM modernization.The TCM clinical diagnosis and treatment follows the“syndrome induction and treatment determination” principle.In the syndrome induc-tion process,the doctor comprehensively analyzes the patient's symptoms collected by“watch,listen,ask and feel the pulse” to diagnose the cause as a certain syndrome.For the treatment determination,the corresponding treatment method is selected according to the syndrome to formulate the prescription.Thus,assisting doctors to induce syn-dromes based on symptoms and further recommend prescriptions are two critical tasks of the intelligent TCM diagnosis and treatment system.However,due to the ambigu-ity and complexity of syndrome induction,most real prescriptions lack the syndrome label.Therefore,the existing prescription recommendation approaches take a set of symptoms as input and output a set of herbs as the recommended prescription.The syn-drome is regarded as the hidden variable linking symptoms and herbs to capture their co-occurrence information in prescriptions.The aforementioned methods cannot ade-quately depict the heterogeneous relations among TCM entities,and ignore the set infor-mation of the co-occurred symptoms in the syndrome induction process.In summary,the prescription recommendation faces the following challenges:(1)data sparseness of TCM prescriptions.Each TCM prescription usually contains only a few symptoms and dozens of corresponding herbs.The sparse co-occurrence of entities limits the model generality.(2)the complexity and ambiguity of syndrome induction and treatment de-termination.TCM theories observe the human body from a macro perspective,which is highly abstract and generalized.Thus,there is a lack of standard answers to syndrome induction in real prescriptions.(3)heterogeneous TCM entity and relations.The TCM domain includes multiple entities such as symptoms,syndromes,treatments,herbs,and the heterogeneous relations among them,which increases the difficulty of prescription analysis.In recent years,graph representation learning(GRL)has been a hot research topic in machine learning.The core idea of GRL is to project the nodes,edges,or the whole graph into a low-dimensional feature space while keeping the original graph structure information.The application of GRL in recommendation systems helps improve rec-ommendation accuracy and user experiences.Inspired by this,this thesis transforms TCM prescriptions into heterogeneous graph data and adopts graph neural networks as the technical framework.Specifically,first this thesis integrates TCM and modern medicine knowledge graphs to depict the heterogeneous relationship among entities,and then learns the multi-view representations for TCM symptoms and herbs.Further,this thesis summarizes the embeddings of multiple symptoms to induce the syndromes,and generates appropriate prescriptions based on the syndrome embedding.This re-search has provided strong support for the research and innovation of the TCM experi-ence from multiple perspectives.The research content and technical contributions are summarized as follows:1.Syndrome-aware prescription recommendation with multi-graph convo-lution network.This thesis proposes a syndrome-aware multi-graph convolu-tional network model.First,this thesis constructs the symptom-herb graph,symptom- symptom graph,and herb-herb graph from prescriptions.It utilizes three graph convolution networks to learn the comprehensive representations for symptoms and herbs on the above three graphs.Considering the heterogeneity of the symptom- herb graph,this thesis designs a bipartite graph convolution network to distinguish the different type spaces of symptoms and herbs.Subsequently,given a set of symptoms,all the symptom embeddings in the symptom set are merged to induce the implicit syndrome embedding,and the syndrome embedding is used to select herbs into the recommended prescription.The experimental results on real TCM prescriptions show the superiority of graph neural networks in modeling high-level semantics among TCM entities,and the necessity of modeling the syndrome in-duction and treatment processes according to the holism theory.2.Knowledge-enhanced prescription recommendation with graph attention network.Further considering that in the stage of “treatment determination”,the formulation of herbs follows the “compatibility of seven emotions” principle.Thus,different herbs have different functions and cooperate with others to relieve all symptoms.Therefore,in a prescription,each herb should have its own correspond-ing syndrome embedding to reflect the diversified mappings between symptoms and herbs.Based on this discovery,this thesis proposes a knowledge-enhanced graph attention network model.In the representation learning stage,this thesis ap-plies the graph attention networks to distinguish the importance of different neigh-bors for each node on the symptom-herb graph,symptom-symptom graph,and herb-herb graph.In addition,to enrich the semantic information,this thesis in-corporates the entity features extracted from a TCM knowledge graph,and learns the entity embeddings containing topological structure and node feature informa-tion.In the prediction stage,given a set of symptoms,the attention mechanism is adopted to learn the corresponding syndrome embedding for each herb,and then the appropriate herbs form the recommended prescription.This research depicts the fine-grained syndrome induction and treatment process,helping to mine the complex relations in prescriptions.The experimental results verify that introduc-ing knowledge graph features and the attention mechanism can help improve the recommendation accuracy.3.Meta-path guided prescription recommendation with graph attention net-work.Only depending on TCM theories to recommend prescriptions is not con-ducive for the TCM modernization.Therefore,this thesis constructs an integrated TCM-Modern Medicine knowledge graph,proposes a meta-path-guided graph at-tention network,and builds a multi-view prescription recommendation system.Specifically,this research first builds a knowledge graph containing both TCM and modern medicine knowledge.Subsequently,this thesis defines several meta-paths covering TCM and modern pharmacological theories to guide the graph at-tention network propagation and learn the multi-view embeddings for symptoms and herbs.Finally,this thesis adopts the syndrome induction method similar to the previous two researches to recommend prescriptions.This research employs a multidisciplinary approach to combine the TCM and modern medicine theories,which is conducive to exploring the action mechanism of herbs.The experimental results show that integrating TCM and modern medicine help improve prescription recommendation accuracy,and the model can provide explainable recommenda-tions to some extent.