目的 提出一种融合组合赋权、聚类、决策、评价等多种机器学习算法用于保健食品配方设计及评价的方法。方法 以五味子Schisandre Chinensis Fructus为例，构建含五味子处方数据库，筛选与五味子配伍高频药味并挖掘其关联规则。同时结合传统中医药理论和现代科学研究建立评价指标体系，采用兼顾主观性和客观性的层次分析（analytic hierarchy process，AHP）-CRITIC（criteria importance though intercrieria correlation）组合赋权法和优选的聚类算法对高频药味进行加权和聚类，结合中医药理论设计配方并进行逼近理想解排序（technique for order preference by similarity to an ideal solution，TOPSIS）综合评价。结果 频次统计得到黄芪、茯苓、人参等与五味子配伍高频药味31个，关联规则分析显示高频药味间更易产生强关联。构建含3个一级指标、7个二级指标的评价指标体系，AHP-CRITIC组合赋权法计算出的指标组合权重从大到小依次为药味传统功效、现代文献研究、在数据库中的出现频次，符合主观认识和客观数据。根据与专业知识的匹配度、算法运行效率及对数据的包容度优选模糊C均值（fuzzy C-means，FCM）聚类，将高频药味分为5类。结合中医药理论及上述结果设计可能的新配方共11个，TOPSIS综合评价排序的结果显示，五味子-黄芪-白术-党参是五味子保胃护肝保健食品可能的最优新配方。结论 该模型在中医药理论的指导下，提供了既能体现传统中医药配伍理论又有足够现代科学研究成果支撑的中药类保健食品配方设计与研发的创新思路与方法。

Objective A formula design and evaluation model was proposed, which integrated multiple machine learning algorithms such as combination of empowerment, clustering, decision making and evaluation. Methods Taking Wuweizi (Schisandre Chinensis Fructus) as an example, constructing the prescription database containing Schisandre Chinensis Fructus, screening high-frequency function ingredients could be used in health food and mining its association rules. Meanwhile, a comprehensive and objective evaluation index system was established, which combined theory of traditional Chinese medicine (TCM) and modern scientific research results. The analytic hierarchy process (AHP)-criteria importance though intercrieria correlation (CRITIC) combination empowerment was adopted, which considered both subjectivity and objectivity. K-means, self-organizing map (SOM) and fuzzy C-means (FCM) were optimized, high-frequency function ingredients were weighted and clustered by the approaches above. Then combined with TCM theory, the formulas were designed and conducted by technique for order preference by similarity to solution (TOPSIS) comprehensive evaluation. Results 31 Kinds of high-frequency function ingredients containing Huangqi (Astragali Radix), Fuling (Poria) and Renshen (Ginseng Radix et Rhizoma) were obtained by frequency statistics. The results of association rule analysis showed that strong association rules were more likely to be generated between high-frequency function ingredients. An evaluation index system containing three first-level indicators and seven second-level indicators was constructed. The combination weight of the index calculated by AHP-CRITIC combination empowerment was the traditional efficacy of medicine, modern literature research and the frequency of occurrence in the database from large to small, it is consistent with subjective knowledge and objective data. Fuzzy C-means (FCM) clustering was finally selected as the clustering algorithm in this paper according to the degree of matching with professional knowledge, the efficiency of algorithm operation and the degree of data tolerance, high-frequency function ingredients were divided into five categories. Combined with the TCM theory and results above, a total of 11 possible new formulas were designed. The results of TOPSIS method showed that Schisandre Chinensis Fructus-Astragali Radix-Baizhu (Atractylodis Macrocephalae Rhizoma)-Dangshen (Codonopsis Radix) was the best possible new formula of Schisandre Chinensis Fructus health food for protecting gastric mucosa and liver. Conclusion Under the guidance of the theory of TCM, this model provided innovative ideas and methods for the formula design, research and development of TCM health food, which can embody the compatibility theory of TCM and modern scientific research results.

R283.21;TS218;TP312

河南省重大科技专项(211110310100)