基于证据链推理的鲁棒性分类及对心脏病诊断决策支持

发布时间：2018-05-28 16:24

  本文选题：证据链推理 + 实体异构性　； 参考：《天津大学》2015年博士论文

【摘要】：数据驱动决策广泛存在于工程实践与管理中,对于数据融合中的知识推理理论和方法提出了新的挑战。这些决策数据源自多传感器、关系数据库、不同经验水平专家知识等,致使传统的决策方法难以有效处理。因此,本文基于证据融合的决策框架,利用多源数据感知、信息传递与共享、CBR/RBR分类及智能决策等理论,研究了基于证据链推理的鲁棒性分类决策。主要工作及创新性如下:(1)研究了层次关联证据链推理方法。综合评述了群决策的证据推理相关研究,并定义了证据链的知识结构、可信度序和指数型相似度。从属性量、特征量和标识类别三层次,分析了用于分类决策的证据链性质。(2)研究了多源证据链推理模型,揭示了异构实体下数据驱动决策的推理机制。首先在单数据表证据链关联基础上,推导出查询案例推论的融合可信度,改进了相似度频率加权近邻算法sf-NN,分析了类别错误标识对决策结构的影响。之后研究了证据链融合的正交合成规则,拓展了多数据表证据链推理模型mr FUER,利用多源证据链关联算法x D-NN,提供解释能力强的鲁棒性决策。(3)为揭示时态系统在不同时间尺度下决策机制,将证据链从单一尺度推理拓展到多尺度推理,提出了多尺度特征的证据链推理模型ms FUER。使用相似度矩阵和辨识准则,构建了二次优化的分类辨识框架,获取了特征量的鲁棒性权值参数。之后利用多尺度互信息,提出了二级混合整数优化的多尺度特征优选策略,用以解决特征组合增长问题,使得推理信息价值最大化。提出时态相似度的最近邻算法ts-NN,其推理机制优于传统的单一尺度决策的推理机制。(4)为揭示系统在过程感知下的动态决策机制,将证据链从全域一次性推理拓展到序贯推理,发掘了感知模糊性下决策状态转移及可信度更新规律。放松了之前查询的感知数据拥有完全信息的假设,在部分信息下构建了过程感知证据链推理模型md FUER,以单个特征量的特异度和灵敏度估算似然概率,提出基于狄利克雷函数的可信度更新算法df-BU,有效地实现了过程感知的鲁棒性决策。(5)仿真实验研究了基于FUER模型集的医疗决策支持。使用弗雷明汉心脏研究中4240个病历、源自三个医疗机构的920例异构性病例和重症监护室的多参数智能监护(MIMICII)的时态数据,研发了临床决策支持系统原型及其共享的知识库。对于h类和l类高低水平专家,从均衡准确度和证据链长度上评价了决策质量和效率改善效果,结果表明增强了分类决策鲁棒性。
[Abstract]:Data-driven decision making is widely used in engineering practice and management, which poses a new challenge to the theory and method of knowledge reasoning in data fusion. These decision data are derived from multi-sensor, relational database, expert knowledge of different levels of experience and so on, which makes the traditional decision making method difficult to deal with effectively. Therefore, based on the decision framework of evidence fusion, using the theories of multi-source data perception, CBR / RBR classification and intelligent decision, this paper studies the robust classification decision based on evidence-chain reasoning. The main work and innovation are as follows: 1) the hierarchical correlation evidence chain reasoning method is studied. This paper reviews the research on evidence reasoning in group decision making, and defines the knowledge structure, credibility order and exponential similarity of evidence chain. From the three levels of attribute quantity, feature quantity and identification category, this paper analyzes the nature of evidence chain used in classification decision. It studies the reasoning model of multi-source evidence-chain, and reveals the reasoning mechanism of data-driven decision under heterogeneous entity. Firstly, based on the evidence chain association of single data table, the fusion credibility of query case inference is derived, the similarity frequency weighted nearest neighbor algorithm sf-NN is improved, and the influence of category error identification on decision structure is analyzed. Then the orthogonal synthesis rules of evidence chain fusion are studied. In this paper, the multi-datasheet evidence-chain reasoning model, mr fer, is extended, and the multi-source evidence-chain association algorithm x D-NN is used to provide robust decision making with strong explanatory power. It reveals the decision mechanism of temporal systems at different time scales. The evidence chain is extended from single scale reasoning to multi scale reasoning, and a multi scale feature based evidence chain reasoning model Ms fer is proposed. By using similarity matrix and identification criterion, a quadratic optimization classification and identification framework is constructed, and the robust weight parameters of the feature quantity are obtained. Then, using multi-scale mutual information, a multi-scale feature optimization strategy for two-level mixed integer optimization is proposed to solve the problem of feature combination growth and maximize the value of reasoning information. This paper presents a temporal similarity nearest neighbor algorithm ts-NN, whose reasoning mechanism is superior to that of the traditional single-scale decision making mechanism. In order to reveal the dynamic decision-making mechanism of the system under process awareness, the evidence chain is extended from global one-off reasoning to sequential reasoning. The rules of decision state transition and credibility update under perceived fuzziness are explored. After relaxing the assumption that the perceptual data of the previous query has complete information, a process perception evidence chain reasoning model, md fer, is constructed under partial information. The likelihood probability is estimated by the specificity and sensitivity of a single feature quantity. A credibility update algorithm df-BUbased on Delikley function is proposed. The robust decision making of process awareness is realized effectively. The simulation experiment is carried out to study the medical decision support based on FUER model set. The prototype of clinical decision support system and its shared knowledge base were developed by using the temporal data of 920 heterogeneous cases in three medical institutions and MIMICII in intensive care unit using 4240 medical records in the Framingham heart study. For high and low level experts of class h and class l, the effect of improving decision quality and efficiency is evaluated in terms of equilibrium accuracy and length of evidence chain. The results show that the robustness of classification decision is enhanced.
【学位授予单位】：天津大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：R541
，

本文编号：1947393