基于软阈值领域聚合图网络的《伤寒论》方剂推荐方法  

Prescription recommendation method of Treatise on Febrile Diseases based on soft threshold domain aggregation graph network

作　　者：苗春龙 沈融[2] 兰杰 刘思彤 陈启亮 罗静静 MIAO Chunlong;SHEN Rong;LAN Jie;LIU Sitong;CHEN Qiliang;LUO Jingjing(Engineering Research Center of Intelligent Robotics,Jihua Laboratory,Foshan 528200,China;Yueyang Hospital of Integrated Traditional Chinese and Western Medicine,Shanghai University of Traditional Chinese Medicine,Shanghai 200437,China;School of Basic Medicine,Guangzhou University of Chinese Medicine,Guangzhou 510006,China;Academy for Engineering&Technology,Fudan University,Shanghai 200433,China)

机构地区：[1]季华实验室智能机器人工程研究中心,广东佛山528200 [2]上海中医药大学附属岳阳中西医结合医院,上海200437 [3]广州中医药大学基础医学院,广东广州510006 [4]复旦大学工程与应用技术研究院,上海200433

出　　处：《上海中医药大学学报》2024年第6期8-18,共11页Academic Journal of Shanghai University of Traditional Chinese Medicine

基　　金：广东省重点领域科研计划(季华实验室)基金项目(X190051TB190);上海市市级重大专项基金项目(2017SHZDZX01);广东省普通高校重点领域专项(新一代信息技术)基金项目(2020ZDZX3003)。

摘　　要：目的:针对《伤寒论》辅助诊断和决策支持缺乏等问题,以《伤寒论》名家医案为基础,提出一种基于深度学习的《伤寒论》医案数据建模方法和《伤寒论》方剂推荐模型。方法:本研究采用基于Levenshtein距离的同义词识别算法对医案资料中的症状描述进行聚类,减小症状数据集维度;以“患者”-“症状”为本体构建《伤寒论》异质图数据集,并采用Sentence-BERT模型获得具有深层语义的节点嵌入向量;提出融合自适应软阈值滤波的领域聚合STGraphSAGE图卷积深度学习模型,并将其应用于《伤寒论》方剂推荐任务。结果:最终将非结构化的《伤寒论》医案建模为图结构数据集,并在该数据集上进行了实验,在F1-Score等4项评价指标下,提出的ST-GraphSAGE取得了优于机器学习及其他主流图深度学习模型的成绩,方剂推荐准确率可达76%,且相较于基础模型,ST-GraphSAGE在处理含噪数据时准确率提高0.51%~3.92%。结论:提出《伤寒论》医案建模方案及ST-GraphSAGE图深度学习模型,其方剂推荐准确率较高且泛化能力强,能充分利用《伤寒论》医案的个体及关系信息,为中医临床的决策支持提供新的研究思路。Objective:In view of the problems such as the lack of auxiliary diagnosis and decision support in the field of Treatise on Febrile Diseases(TFD),this paper proposes a data modeling scheme for TFD medical records and a TFD prescription recommendation model based on deep learning per the medical records of famous experts in TFD.Methods:In this study,the synonym recognition algorithm based on Levenshtein distance is used to cluster the symptom descriptions in TFD medical records to reduce the dimension of the symptom data set.The"patient"-"symptom"is used as the ontology to construct a heterogeneous graph data set of TFD and Sentence-BERT model was adopted to obtain node embedding vectors with deep semantics.A domain aggregation ST-GraphSAGE deep learning model fused with adaptive soft threshold filtering was proposed and applied to the TFD prescription recommendation task.Results:Unstructured TFD medical records were modeled as a graph-structured data set,and experiments were conducted on this data set.Through four evaluation indicators such as F1-Score,the performance of ST-GraphSAGE proposed in this paper achieved better results than machine learning and other mainstream graph deep learning models.The accuracy rate of prescription recommendation reached 76%,and compared to the basic model,the accuracy rate was increased by 0.51%~3.90%when ST-GraphSAGE was used to deal with noisy data.Conclusion:The TFD medical records modeling scheme and ST-GraphSAGE graph deep learning model proposed in this paper have higher accuracy rate in prescription recommendation and strong generalization ability,which can make full use of the individual and relational information of TFD medical records to provide new research ideas for decision support in clinical practice of traditional Chinese medicine.

关 键 词：中医方剂推荐 《伤寒论》医案 深层语义嵌入 领域聚合图网络 软阈值滤波 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术] R289[自动化与计算机技术—计算机科学与技术]