基于机器学习的中药升降浮沉药性判别作者：油雨忻山东中医药大学智能与信息工程学院2019级生物医学工程专业指导教师：魏国辉，陈素玲摘要升降浮沉药性表示中药作用在人体内的四种走向和趋势，是中药药性理论的主要组成部分，在临床指导用药以及药方配伍中具有重要的指导作用。为了探究中药的升降浮沉药性与中药成分是否有关，基于前期课题组研究得到的结论“中药寒热药性物质基础是中药成分”，提出假说“中药升降浮沉药性的物质基础也是中药成分”。本文以升降浮沉药性为切入点，引入机器学习、人工智能等现代科技，探索中药升降浮沉药性与中药成分的相关关系。为了验证假说，本文利用国家“973”项目所得数据集中的54味植物类中药在石油醚溶剂下的紫外指纹图谱数据表征中药成分。首先对原始数据进行预处理，包括PCA降维、过采样、数据标准化等操作。使用预处理后的数据训练模型，采用网格优化算法寻找模型最优参数、交叉验证用于评估模型性能。构建中药升降浮沉药性判别模型时，分别采用决策树、随机森林、AdaBoost、支持向量机、朴素贝叶斯算法和K近邻算法等6种算法对中药按照升浮类和降沉类进行分类。最终采取统一的评价指标对识别效果评价，评价指标包括准确率(Accuracy)、精确度(Precision)、召回率(Recall)、AUC值和混淆矩阵。实验结果表明，对比6种机器学习算法的模型交叉验证结果准确率，支持向量机、随机森林及AdaBoost算法的模型训练集识别准确率均可达85%以上。比较AUC值发现，朴素贝叶斯算法和支持向量机模型的AUC值在六种模型中最高，模型稳定性较好。本文构建的机器学习模型表现出较高的准确性，训练集准确率均在80%以上，对中药升降浮沉药性判别具有较好效果，证明了中药紫外指纹图谱可以表征中药成分。此外，结果也验证了中药的升降浮沉药性与其成分之间的密切相关性。关键词：中药：升降浮沉：机器学习；药性判别The Classification of Traditional Chinese MedicineBased on the Machine LearningABSTRACTIt is a major component of the pharmacological theory of traditional Chinese medicine,and plays an important role in clinical guidance of drug use and prescription compounding.Inorder to investigate whether the rising,falling and sinking properties of Chinese medicine arerelated to the constituents of Chinese medicine,we propose the hypothesis that the materialbasis of the rising,falling and sinking properties of Chinese medicine is also the constituentsof Chinese medicine,based on the conclusion obtained by the previous research group that"the material basis of the cold and hot properties of Chinese medicine is the constituents ofChinese medicine".In this paper,we take the rising,falling and sinking properties of Chinesemedicine as the starting point and introduce modern technology such as machine learning andartificial intelligence to explore the correlation between the rising,falling and sinkingproperties of Chinese medicine and the ingredients of Chinese medicine.In this paper,the ultraviolet fingerprints of 54 botanicals in petroleum ether solvent fromthe dataset of the National "973"Project,"Research on the basic problems related to thepharmacological theory of Chinese medicine",were used to characterize the constituents ofChinese medicines.The UV fingerprint data were used to characterize the components of theChinese medicines.The raw data were first pre-processed,including PCA dimensionalityreduction,oversampling and data normalization.The model was trained using thepre-processed data,and a grid optimization algorithm was used to find the optimal parametersand cross-validation to evaluate the model performance.Six algorithms,including decisiontree,random forest,AdaBoost,support vector machine,plain Bayesian algorithm andK-nearest neighbour algorithm,were used to classify the Chinese medicines according to theascending and descending classes.A unified evaluation index was adopted to evaluate therecognition effect,which included Accuracy,Precision,Recall,AUC value and confusionmatrix.The experimental results show that when comparing the accuracy of modelcross-validation results of the six machine learning algorithms,the recognition accuracy ofthe training set of the models of Support Vector Machine,Random Forest and AdaBoostalgorithms can all reach over 85%.Comparing the AUC values,we found that the AUCvalues of the plain Bayesian algorithm and the support vector machine model were the highestamong the six models,and the model stability was better.The machine learning modelsconstructed in this paper showed high accuracy,with the accuracy of the training set beingabove 80%,and had good results in the discrimination of the lifting and floating properties ofChinese medicines,proving that the UV fingerprinting of Chinese medicines can characterizethe components of Chinese medicines.In addition,the results validated the close correlationbetween the lifting and sinking properties of TCM and its components.Key words:Chinese Medicine;Lifting and Sinking;Machine learning;Pharmaceuticalproperties identify