基于变分证据推理多尺度决策的病例类别推荐研究

Due to the characteristics of temporal, fragmentation and uneven distribution of the clinical data in evidential reasoning, the medical data fusion process has the interesting and challenging issue of reducing its misdiagnosis and improving the system discrimination ability. However, the previous data fusion approaches on multi-source heterogenous data cannot deal with the issues for providing resilient decision support. To solve these issues, we first introduce the concept of multiscale on three dimensions of organizational structure, decision time and its locations. The multi-granularity multi-attribute data space was constructed with the variational theory in probability graph model, then the temporal medical data were analyzed with pattern recognition, segmentation and associative rules. Second, the variational evidential reasoning framework was constructed with four sub-models. The model of feature data acquisition was constructed in the homomorphic evidence context, optimizing the mechanism of medical data selectively retrieval. The model of belief fusion was constructed based on the hierarchical Dirichlet process, achieving the confidence fusion of fragmented evidence. The convex relaxation model of evidential chains-based reasoning was proposed, which expanded the multi-scale decision-making model and its reasoning mechanism. Furthermore, the integration of clinical decision Knowledge repositories and their approximate inference engine inference was cast as a random field in evidence fusion, in which the improvment fuction and deterioration function were formulated with the variables of maintenance time and effort in the decision-making scenarios. Then the decision cost control model was constructed for maintaining the knowledge base, balancing the system discrimination ability against the maintenance cost of its clinical knowledge base. Through the parameter learning of multiscale decision-making model, the recommendation of indentifying the lables of medical instances were achieved with the control trajectory of the reasoning model. Moreover, the evidence matching algorithms including k-d tree based neighbor searching have been proposed to provide relative decision-making tools. Finally, digital simulation study of human health and clinical tests verified the effectiveness and resilience of the variational evidence-based reasoning models and their multi-scale decision-making mechanism, improving the efficiency and quality of medical services, and providing scientific basis and theoretical guidance for addressing the contradictions between the growing medical needs and the lack of medical resources from the perspective of intelligent decision support.

临床病例证据推理因数据时空异构、碎片化、价值分布不均的特征，使得医疗决策支持面临降低医疗误诊率与提升系统分辨能力的科学问题。本研究拟采用概率图中的变分理论指导构建多粒度多属性数据空间，对时空异构的医疗数据进行模式辨识、切分与多尺度关联。拟从多尺度决策角度提出变分证据推理框架，构建同态下的特征数据获取模型，优化医疗数据选择性获取机制；提出基于分层Dirichlet过程的可信度聚合模型，实现碎片化证据置信度融合；构建医疗决策证据链推理凸松弛模型，拓展多尺度决策模式及其推理机制；并在随机推理场中分析其增强与衰减函数，构建知识库维护的决策成本控制模型，解决系统分辨能力提升与临床知识库维护成本控制之间矛盾。利用模型参数学习和决策控制迹，实现病例信息推荐的决策措施。通过医疗人体数字化仿真和临床比较试验，验证决策模型及推理机制的有效性，为提高医疗服务效率和品质提供智能决策支持的科学依据及理论指导。

面对健康医疗决策支持证据推理因数据时空异构、碎片化、价值分布不均的特征，而存在的医疗误诊率高与系统分辨能力低的难题，本课题构建了变分证据推理多尺度决策的病案类别推荐模型。该模型采用了概率图中的变分理论指导构建多粒度多属性数据空间，对时空异构的医疗数据进行了模式辨识、切分与多尺度关联。从多尺度决策角度提出了变分证据推理框架，构建了同态下的特征数据获取模型，解决了决策信息价值不均的医疗诊断数据组合呈指数增长的问题，优化了医疗数据选择性获取机制，降低了医疗误诊率。提出了基于分层Dirichlet过程的可信度聚合模型，实现了碎片化证据置信度融合，提升了系统分辨能力。提出了多类型干扰因素下的随机混合模型，揭示了不同概率分布下的干扰因素对医疗决策证据链类别误标、决策可信度和信息传导及更新机制的作用规律，解决了医疗决策指标的显著性与决策能量开销的矛盾。拓展了多尺度决策模式及其推理机制，解决了健康医疗数据机器学习的代价敏感性问题。应用医疗人体数字化仿真系统，模拟多模态数据生理变化趋势，与临床诊断结果比较研究，为多层次医疗决策者提供了智能决策支持。以妊娠期糖尿病诊断和互联网健康医疗社区中的问诊数据等为例，进行临床决策支持和健康医疗社区激励机制的效应评估分析，验证了决策模型及推理机制的有效性。部分理论成果用于参与国办发〔2016〕47号）文件解读工作，出版著作《大数据医疗》。与相关合作机构一起构建了妊娠期糖尿病等健康医疗数据库，涵盖近20000患者的5年间的建档及健康检查数据，用于模型验证及机器学习。研发了基于电子病历和机器学习的妊娠期糖尿病预测系统，为健康医疗决策团队和患者提供个体化健康管理应用，包括基于健康大数据监测平台的移动医疗APP等。这些应用有效地降低了健康医疗数据过载和数据分析复杂性，为诊断决策团队和患者提供更及时的决策支持。发表的系列高水平论文在同行中形成了一定的影响力并对本学科相关领域发展做出了一定贡献。