基于原型范畴理论名老中医诊治肺癌认知模型优化研究

A cognitive model that can reflect Piao Bingkui’s clinical experience and thinking process in treating lung cancer had been preliminarily established through our previous study. But it reported an obviously low classification accuracy of some drug models which possessed functions of strengthening vital Qi, clearing heat and detoxifying the body. This study is a model optimization method research proposed to solve problems of clinical information processing mechanism, generalization and division of cognitive elements, interpretation of thinking mechanism, and application of model simulation, etc. Based on 2000 large sample data, the “prototype category theory” will be applied to establish the categorization attribute of syndrome elements of lung cancer through interviews, including disease location, pathogenesis, syndrome identification and potential syndrome. It is to subjectively summarize and classify clinical information of old famous traditional Chinese medicine experts, which can express prior experience knowledge and thinking mechanism. We will use RBF kernel, cross-validation and grid-search to obtain the optimization support vector machine c and g parameters, to achieve algorithm optimization. We will also implement case correct classification rate Kappa, F-Measure , MCC and ROC Area to evaluate its reliability and effectiveness. Confusion matrix will be adopted to visualize and precisely evaluate the performance of algorithm. The study will provide basis for the development of clinical decision system on cancer prevention and treatment by old famous traditional Chinese medicine experts.

在对前期研究建立的朴炳奎诊治肺癌认知模型评价中分析发现，由于名老中医存在基于核心病机和现代药理学先验知识，故仅基于客观临床信息的扶正培本、清热解毒药物认知模型分类准确率较低。本研究是为解决临床信息处理机制、认知要素概括和划分、思维机制阐释、仿真应用等问题而提出的模型优化方法学研究。拟引入原型范畴理论，通过访谈研究构建肺癌病位、病性、病机、病证、病势范畴化属性，用于表述名老中医对临床信息主观概括和类属划分，以表达专家先验知识和思维机制，并在此基础上探讨基于2000例大样本数据，应用RBF kernel，cross-validation和grid-search 得到最优化支持向量机c、g参数实现算法优化的模型优化方法，通过混淆矩阵，正确分类率，Kappa，F-Measure，MCC，ROC Area对算法性能、精度及可靠性进行评价。为开发肿瘤防治名老中医临床决策系统提供基础。

认知模型是基于认知心理学理论，反映人类思维机制和认知过程的计算机模型，能够从隐性经验中获得可学习的知识。研究以朴炳奎诊治肺癌大样本医案数据作为训练集，借助WEKA 机器学习软件，基于原型范畴理论和访谈研究构建医案数据范畴化证素、治则属性，通过监督学习（贝叶斯网络、支持向量机）和无监督学习（关联规则），对前期研究的支持向量机监督学习方法进行参数选择优化，输出朴炳奎诊治肺癌“用药经验”优化认知模型，并引入主观评价与效能评价，提高模型精度。.研究共纳入朴炳奎诊治肿瘤病例3441例，建立了基于贝叶斯网络的病-证、病-法、病-药、证-药、病-法等条件关系模型，以及基于支持向量机的证素、治法等权重模型。其中"脾肺气虚"证作为分类属性对2107个实例进行贝叶斯网络分析，分类准确率达73.80%，Kappa值0.4043，发现病-证-症-法-方-药之间的结构关系338条，梳理形成病证-症关系、病证-法关系、病证症-方药关系、法-方药关系、方-药关系图谱。其中与朴炳奎名老中医中医诊疗相关的结构关系292条，并得出父项（节点）与子项（节点）之间相互关联并形成一定规律，解决了模型效能较低、证素赋值差异化等问题。.经过优化后的模型在数量与分类准确率上显著提升，对2107个实例中340项属性进行基于支持向量机的SMO分析，得出分类准确率70%以上药物模型41个。其中原有模型中与临床辨病论治决策相关而分类准确率较低的药物得到提升，如生薏苡仁、半枝莲、法半夏、土茯苓分类准确率达到76.93%、76.93%、90.98%和87.04%，分别较原数据模型提升17.29%、11.43%、11.17%和2.56%。.经过基于原型范畴理论的名老中医诊治肺癌认知模型作为应用基础研究，其成果可以更加全面、客观、可靠、高效、系统的阐释名老中医药专家思维认知规律。有望为开发名老中医防治肿瘤中医药决策辅助系统提供基础。