中医CPG共识专家多视角数据半监督特征选择方法研究

Consensus method is the main method to develop the clinical practice guidelines (CPG) of traditional Chinese medicine and form the recommendation opinions. The selection of experts is the key to implement consensus method. At present, the experts of TCM CPG consensus are selected subjectively by the guidance group based on relatively broad selection criteria and lack of objective evaluation indicators. The commonly used mathematical statistical methods need to assume the relationship between variables, which is not suitable for the analysis of the relationship between complex data variables. This study proposes to start from big data and use the feature selection method of machine learning to optimize the feature subset of TCM CPG consensus experts, which may provide new ideas and methods for the selection of TCM CPG consensus experts. Accordingly, this study intends to take existing Chinese medicine CPG consensus experts as learning samples, take the introduction of experts from relevant networks at home and abroad as the target objects, and select and construct the feature set of Chinese medicine CPG consensus experts. In this study, relevant information is searched from public databases and network platforms, and each expert's feature is marked, and the optimal expert feature is extracted based on SSMVFS algorithm. Finally, this study takes the traditional Chinese medicine CPG of hypertension as an example, implements the consensus method based on the group opinions, analyzes the reliability of the opinions given by each expert, and verifies the applicability of this research method to the selection of traditional Chinese medicine CPG consensus experts.

共识法是制定中医临床实践指南（CPG），形成推荐意见的主要方法，遴选专家是实施共识法的关键，目前主要基于比较宽泛的遴选标准由指南制定小组主观遴选共识专家，缺少客观评价指标。既往筛选指标常用的数理统计方法需先假设变量间关系，不适合分析复杂数据，难以解决多属性、多特征专家的客观评价指标遴选问题。本研究提出从大数据着手，利用机器学习的特征选择方法，优选中医CPG共识专家的特征子集，可能为遴选中医CPG共识专家提供新思路、新方法。据此，本研究拟以现有的中医CPG共识专家作为学习样本，以国内外相关网络的专家介绍为目标对象，筛选构建中医CPG共识专家特征集，从公开数据库、网络平台中寻找相关信息，进行每一位专家的特征标记，基于SSMVFS算法提取最优专家特征子集，最后，以高血压病中医CPG为例，实施共识法，以群体意见为标准，分析每一位专家给出意见的可靠度，验证本研究方法对遴选中医CPG共识专家的适用性。

共识法是制定中医临床实践指南（CPG），形成推荐意见的主要方法，遴选专家是实施共识法的关键，目前主要基于比较宽泛的遴选标准由指南制定小组主观遴选共识专家，缺少客观评价指标。既往筛选指标常用的数理统计方法需先假设变量间关系，不适合分析复杂数据，难以解决多属性、多特征专家的客观评价指标遴选问题。本研究提出从大数据着手，利用机器学习的特征选择方法，优选中医CPG共识专家的特征子集，可能为遴选中医CPG共识专家提供新思路、新方法。据此，本研究基于 2014 年公共卫生专项基金制修订 254 项指南的任务，选取最终以指南发布、实施共识法的123项指南，委托中华中医药学会，与指南开发项目组沟通，收集指南共识法实施的临床专家2087人，作为学习样本。以国内外相关网络的专家介绍，尤其是以专家所在单位官网介绍为主要信息来源，通过构建数据采集引擎，对关键信息段进行提取进行匿名标识化，将同组数据进行合并，构架文本分词及名词提取，去重后进行持久化，对已分词好的数据进行匹配，归类保存。收集2087位专家的特征信息，共包含102个信息变量，选择其中40个规范化核心变量纳入分析。基于SSMVFS算法提取4个视角的最优专家特征子集。最后，以眩晕中医CPG为例，实施共识法，以群体意见为标准，分析每一位专家给出意见的可靠度，验证了本研究方法对遴选中医CPG共识专家的适用性。本研究从大数据着手，利用机器学习的特征选择方法，优选中医 CPG 共识专家的特征子集，为中医 CPG 共识专家遴选提供客观指标，使今后指南制定者在遴选中医 CPG 共识专家时能够有的放矢，从而提高中医 CPG 共识过程严谨性和共识结果可靠性，为制定中医 CPG 时共识法的应用提供方法学支撑。