《伤寒论》的“方—证要素”对应体系及其神经网络数学模型的构建

【作者】 陈擎文；

【导师】 李宇航；

【作者基本信息】 北京中医药大学 ， 中医临床基础， 2011， 博士

【摘要】 本研究乃是根据“证候要素、方剂要素”最新研究进展、与计算机领域的人工智能技术相结合来研究伤寒论的处方,建立一套能处理伤寒论“方-证要素对应、主证-药物对应”的数学模型,分述如下。目的(1)将伤寒论处方结构分析,并建立“证候要素、方剂要素、方-证要素对应、主证-药物对应”体系。(2)以人工神经网络来建立整个伤寒论“主证-药物对应”的数学模型。(3)以人工神经网络来建立整个伤寒论“方-证要素对应”的数学模型。系统数学模型：Yj=输出变数向量,即是伤寒论的方剂要素(或药物)Xi=输入变数向量,即是伤寒论的证候要素(或主证)f=人工神经网络神经元模型的转换函数wij=模仿第i个与第j个生物神经元间的突触强度(即连结加权值)θj=模仿第j个生物神经元的阀值(即门限值)(4)将大量的“证候要素、方剂要素、方-证要素对应、主证-药物对应”等数据汇入神经网络模型内,让系统不断学习后,即可获取每个参数的权值与阀值,如此的伤寒论“方-证要素对应、主证-药物对应”数学模型,即可有推算预测处方药物的功能,对于辅助教学与临床诊治具有一定参考价值,可供后续萃取其中的知识与信息,为更好地继承与发扬仲景学术,提供方法学借鉴。方法(1)系统所需计算机程序皆是自行编写开发,采用Microsoft Visual Studio C#. NET 2008结合SQL Server 2008与Access 2007数据库来开发系统所需要的功能。(2)并采用神经网络技术其中的一种数学算法倒传递网络(Back Propagation Network, BPN)来建立整个数学模型。(3)建立“证候要素、方剂要素、方-证要素对应、主证-药物对应”体系,并将这些对应资料汇入神经网络模型内。(4)经由系统不断的训练学习后,即可获取每个神经元节点的权值与阀值,从而构建伤寒论“方-证要素对应、主证-药物对应”体系的数学模型。结果与结论(1)本研究以《伤寒论讲义》“十一五”规划教材为蓝本来进行《伤寒论》处方的结构化分析,药性归类依据乃是以中国药典2005年版的内容为准。从伤寒论处方结构化分析后的大规模数据中可以得到多种互为对应的讯息,例如：方-证要素对应关系、主证-药物对应关系、方-药对应关系、方-证对应关系.等等,本研究将只针对其中的“方-证要素对应关系、主证-药物对应关系”分别建立伤寒论的的数学模式,建立数学模式的目的除了可以探讨不同证候与药物等变数之间的定性关联,以及彼此间定量的加重权值外,并可在一旦确定数学模型的权值等参数后,进而将之应用于临床辨证论证的辅助上。(2)以人工神经网络技术建立伤寒论“主证-药物对应关系”的数学模型：方法：先将伤寒论处方的结构化资料撷取其中的主证-药物对应关系,转化成输入x向量(主证,最多有354个神经元节点)与输出y向量(药物,最多有98个节点),经由倒传递神经网络来训练学习后,再求出权值ω与阀值0向量最佳解。最佳解的判别准则乃是以在50000次的计算cycle次数内,若误差均方根函数数值低于0.06,则判定本次学习过程求取到权值Wij与阀值θj向量的最佳解。整个平台乃是在在IBM P4个人电脑与微软作业系统Windows XP的环境下运作,每次的求解计算最少要耗时60分钟以上才能计算完毕。结果：测试后发现在一般维度下可正常运作及计算求解,将求得解的数据加以应用于测试实际的辅助临床推测方药,结果符合预期,初步验证可行。但当神经元节点数过大时会发生无法收敛的现象,后经两种方法共同运用,可改良求解过程的收敛效果,包括：调整学习率参数与使用非线性共轭梯度法,结果可将计算出最佳解的系统维度由(219 x 68)提高到(258 x 74),或输入样本可允许由75个提高到88个。(3)以人工神经网络技术建立伤寒论“方-证要素对应关系”的数学模型：方法：先将伤寒论处方的结构化资料撷取其中的证候要素-方剂要素对应关系,转化成输入x向量(证候要素,最多有160个神经元节点)与输出y向量(方剂要素,最多有83个节点),再经由神经网络来训练学习后求出权值Wij与阀值θj向量的最佳解。结果：测试后发现在一般维度下可正常运作及计算求解,将求得解的数据加以应用于测试实际的辅助临床推测方药,结果符合预期,初步验证可行。但当神经元节点数过大时会发生无法收敛的现象,后经三种方法共同运用,可改良求解过程的收敛效果,分别是调整学习率参数、使用非线性共轭梯度法与采用前次最佳解当作起始值,结果可解出最佳解的系统维度由(129 x 72)提高到(134 x 74),或输入样本可允许由读入196个提高到204个。(4)研究结果显示以神经网络所建立的数学模型,在主证-药物对应系统或方-证要素对应系统皆有推算预测处方药物的功能,对于辅助教学与临床诊治具有一定参考价值,也具有辅助教学的功能,从中可知伤寒论处方之证候与药物等变数间的定性关联,甚至彼此间的定量关系。(5)目前根据伤寒论所分析出的主证-药物对应系统或方-证要素对应系统,出现很多同义或相近词汇的主证(或证候要素),故未来的进一步研究还须要再经过词汇统一或简化的探讨,因为词汇的定义与统一将紧密影响着神经网络学习知识的结果,这些影响都将直接表现在相关联神经元节点的权值与阀值上。这些词义相近的证候包括有：(1)身痛类：周身疼痛、身疼痛、身痛(或重)身体痛。(2)呕逆类：呕恶、呕逆、呕恶、干呕、欲呕吐、呕、心烦喜呕、微呕、吐、呕吐涎沫。(3)口渴咽干类：口渴舌燥、舌上燥而口渴甚、口微渴、咽干口干舌躁、口燥咽干。(4)汗出类：汗出、汗漏不止。(5)烦躁类：心烦、心烦、烦躁、微烦、昼日烦躁不得眠,夜而安静、烦躁、心烦不得安、虚烦不得眠、郁郁微烦、心烦不得卧、烦躁欲死、心烦不得眠、心烦懊憹、心烦失眠、大烦渴不解、烦渴。(6)小便利类：小便自利、小便自利、小便利。(7)发热类：身热不去、发热、翕翕发热。(8)喘息类：喘息、喘、喘咳。(9)下利类：下利不止、下利不止、下利日数十行、自利而渴、下利清谷、下利、下利不止、利、下利不止、下利便溏、泄利不止。(10)腹胀满痛类：少腹拘急硬满、少腹硬满、腹痛、腹胀满、腹大满、腹胀满痛、脘腹冷痛、腹胀满、腹满时痛、腹痛拒按、腹痛绵绵。(11)心悸类：心中悸而烦、心悸、心动悸。(12)心下硬痛类：心下硬痛拒按、心下硬满,按之疼痛、心下痞硬满、胸中痞硬。(13)项强类：颈项强、项背拘急不舒。(14)心下痞类：心下痞、心下痞、心下痞满、心下痞硬、心下痞硬而满、心下痞硬、心下痞满。(15)胸胁苦满类：胸胁苦满、胸胁苦满、胸胁苦满.胸胁满而呕、胸胁满微结、胸胁满闷。(16)手足厥冷类：手足厥冷、肢厥、手足厥寒、手足厥寒、四肢厥逆、手足厥逆、手足厥逆、厥逆无脉、四肢厥逆、手足逆冷、手足寒、四肢厥冷、四肢厥冷。(17)风寒束表类：风寒束表、风寒外束、风寒之邪束表。(6)方-证要素对应系统的收敛效率比主证-药物对应系统的收敛效率好,造成这个现象的第一个原因乃是：方-证要素对应系统是一个小规模的对应关系,而且是“一对多”的对应,但是主证-药物对应系统里面的对应关系,是“多对多”,而且经常是“6对8,或8对7,甚至是6对12”的复杂关系,这么紧密繁杂的对应较难计算到最佳解。第二原因乃是两个系统之人工神经元节点数量不同所致,在方-证要素对应中系统维度为(160 x 60),但在主证-药物对应系统中统维度为(354 x 98),因此后者较难求出最佳解。更多还原

【Abstract】 This study coupled both new system of syndrome-essential-factor and formula-essential-factor with the artificial intelligence technique to analyze all the formula of the Shang Han Lun to build a mathematical model of formula-syndrome essential-factor correspondence and formula-syndrome correspondence. These are introduced as follow.Purpose1.To establish the structural analyses of Shang Han Lun and establish the systems of formula-syndrome essential-factor correspondence and formula-syndrome correspondence.2. To establish the mathematical model of formula-syndrome correspondence with artificial neural network technique.3. To establish the mathematical model of formula-syndrome essential-factor correspondence with artificial neural network technique. mathematical model: Yj=output vector, the formula essential-factor (or herb) Xi=input vector, the syndrome essential-factor (or syndrome) f=the transfer function of artificial neural network model Wij= weight Valuesθj= Bias ValuesMethods1. It is developed by self-coding with the software of Microsoft Visual Studio C#. NET 2008, SQL Server 2008 and Access 2007 to design all the functions that we need.2. One of the method called Back Propagation Network (BPN) of the artificial neural network technique was employed to build the mathmatiocal model.3.All the data in the system of formula-syndrome essential-factor correspondence and formula-syndrome correspondence were established and then imported into the mathematical model of artificial neural network 4. The weight and bias values were calculated by reiteration training to produce the new mathematical model of systems of formula-syndrome essential-factor correspondence and formula-syndrome correspondence.Results and Discussion1.This study employed the book of伤寒论讲义by Professor Wang Qingguo (2007) for the materials to undertake the structural analyses of Shang Han Lun. There are many information could be extracted from these structural analyses, i. e. formula essentialfactor, syndrome essential factor, formula-syndrome essential factor correspondence and formula syndrome correspondence.2. To establish the mathematical model, Yj=f(∑wijxi_θj), of formula syndrome correspondence with artificial neural network technique and apply it to the clinical therapy of prescription.Method:The data of structural analyses of Shang Han Lun was imported into the x-vecter and y-vector and then transferred into artificial neural network system to calculate the weight and bias values. If the error of Root-Mean-Square Value is less 0.06 among the iteration of 5000, the optimum solution of the weight and bias values are done. All the system was done at the IBM PC with Windows XP. Every calculation at least nedd 60 minutes to finish the itereation.Result:These results reveal that most of time it could get the optimum solution of the weight and bias values. The mathematical model then be applied to practical prediction of prescription. Those demonstration shows that it is feasible and practical to forecast the precise formulas for the patients. However, some divergence appear while the amount of neuron node is too large. Two methods were employed to imporved the efficiency of convergence, i.e. the adjustment of learning rate and the conjugate gradient method. The systematic dimension could also improved from (219×68) to (258×74). The input samples could also improved from 75 to 88.3. To establish the mathematical model, Yj=f(∑wijxi_θj),of formula-syndrome essential-factor correspondence with artificial neural network technique and apply it to the clinical therapy of prescription.Method:The data of structural analyses of Shang Han Lun was imported into the x-vecter and y-vector and then transferred into artificial neural network system to calculate the weight and bias values. If the error of Root-Mean-Square Value is less 0.06 among the iteration of 5000, the optimum solution of the weight and bias values are done. All the system was done at the IBM PC with Windows XP. Every calculation at least nedd 60 minutes to finish the itereation.Result:These results reveal that most of time it could get the optimum solution of the weight and bias values. The mathematical model then be applied to practical prediction of prescription. Those demonstration shows that it is feasible and practical to forecast the precise formulas for the patients. However, some divergence appear while the amount of neuron’node is too large. Three methods were employed to imporved the efficiency of convergence, i.e. the adjustment of learning rate, the conjugate gradient method and importing the previous optimum solution as current default original vector. The systematic dimension could also improved from (129×72) to (134×74). The input samples could also improved from 196 to 204.4. There are many synonym with the similar terms of syndrome appear while we establish the structural analyses of Shang Han Lun for the systems of formula-syndrome essential-factor correspondence and formulasyndrome correspondence. Because the meanings of the syndrome term will has the significant influence on the caluculation of weight and bias values.It is necessary to deal with the unity and simplification for these syndrome terms.5. The efficiency of convergence for the system of formula-syndrome essential-factor correspondence is better than the one of formula syndrome correspondence. The first reason is because it is a smaller scope of correspondence relation for the system of ormula-syndrome essential-factor correspondence. But it is a more complex correspondence relation for the system of formula syndrome correspondence. The second reason is because it is different for the amount of neuron nodes. The systemic dimension of formula-syndrome essential-factor correspondence is 160x160, but that for formula syndrome correspondence is 354x98. It is harder to calculate the optimum solution of the weight and bias values.更多还原