【作者】 曹秀堂；

【导师】 郭祖超； 徐勇勇；

【作者基本信息】 第四军医大学 ， 医学统计学， 1993， 博士

【摘要】 高维列联表资料是医学研究中常见的一种资料类型，统计上常需拟合对数线性模型。对数线性模型的拟合首先需在分析资料性质的基础上确定选模型的准则以及模型的择优策略，然后借助统计软件拟合恰当的模型。对于所选出的模型，常需对其效应项及参数进行解释，必要时要进行残差分析。如果统计软件没有提供自动寻找最优模型的功能，则要求数据分析者对对数线性模型的拟合、寻优过程及结果解释等有着深入的了解，这对广大医学研究者来说是件很困难的事。 本研究完成的广义线性模型的医学统计专家系统(ESGLM，Expert System for fitting Generalized Linear Models)采用人工智能技术，将广义线性模型的统计背景知识整理出来，构成知识库和推理机，指导用户根据资料情况选择较优对数线性模型、logistic模型和线性回归模型，通过ESGLM的咨询功能和计算功能帮助用户选择较合适的模型并对结果进行解释。 结合Turbo Prolog语言的特点及ESGLM系统的需要，ESGLM的知识体系由知识库和推理机两部分组成。知识库用逻辑表达方式存放广义线性模型的确定性事实和统计规则，推理机负责求解用户的具体问题并向用户提供统计咨询建议。推理机用Turbo Prolog语言实现，其推理的基础是谓词逻辑。问题的解通过逻辑推理而获得。推理机包含一个模式匹配器检索现存的(已知的)信息，匹配答案与问题，以推导一个假设是否为真。ESGLM的推理策略是，系统在一个目标制导下在知识库中搜索结论可实现该目标或可形成几个子目标的规则。在规则调用过程中，动态地生成推理树，根据知识库自动匹配形成菜单。用户在ESGLM的推理过程指导下，形成对该问题的统计处理步骤，之后调用相应的C语言模块完成模型拟合及其它复杂的统计计算。在模型拟合时ESGLM采用了广义线性模型的统一算法，不仅可以拟合对数线性模型、logistic模型和线性回归模型，今后亦可推广到其它形式的广义线性模型。更多还原

【Abstract】 This paper describes a newly developed statistical expert system in medical statistics called ESGLM (Expert System for fitting Generalized Linear Models) which can run on personal computers and display with Chinese characters. There are four modules in ESGLM, those are "data management","knowledge domain", "calculation for model fitting" and "advice-giving presentation".ESGLM is based on the techniques of artificial intelligence(AI), and the design strategy is to separate "knowledge domain" of the AI system from its "inference, engine", the former consists of the facts and rules about generalized linear models (GLM), and the later is the problem-solving advices from statistical background used to guide the user for choosing the most appropriate GLM. The AI system of ESGLM is performed by Turbo Prolog language and two knowledge bases were constructed to store statistical facts and rules in logic pattern. The inference engine of ESGLM is based on the predicate logic mechanism in Turbo Prolog, and a problem-solving answer is achieved by a "hypothesis-fa(?)ts-rules matching engine" that can search the facts and rules which meet the hypothesis in the knowledge bases, show them in the window as guiding heuristics and wait for user’s decision. The reasoning starts with the root of a decision tree created by the program and processes step by step. Several windows are prepared to display the statistical concepts and other heuristic knowledge in each step until the user reaches the final answer for a given question. The module for model fitting and other complex statistical computation is carried out by the program written in C language,which can be interactively operated with statistical inference if the user fills questionable on the output. The algorithm is flexible for GLM if the link function is known. For the moment EXG can treat the problems in fitting log-linear model, logistic model and linear regression model like a human expert, especially in analysis of high-dimension contingency tables by fitting log-linear model because understanding of the estimates in the fitted model is much more difficult for the users who are not familiar with medical statistics. ESGLM also supplies with a friendly operating environment,such as the prompts with Chinese* the pull-down menus, multiple windows, data edit and transformation, graphical display and so on.更多还原