目的 比较不同算法对桂枝茯苓胶囊内容物吸湿性预测模型性能的影响，确定最优建模算法。方法 以54个物理性质参数为输入，胶囊内容物吸湿性为输出，对比偏最小二乘算法（partial least squares，PLS）、决策树算法（classification and regression tree，CART）、多元自适应回归样条算法（multivariate adaptive regression splines，MARS）和广义路径追踪算法（generalized path seeker，GPS）对建立吸湿性预测模型性能的影响。结果 MARS算法建立的预测模型性能最佳，预测能力最强，模型的校正集决定系数（R_c²）为0.843，预测集决定系数（R_p²）为0.808，校正集均方根误差（root mean square error of calibration，RMSEC）为0.391，预测集均方根误差（root mean square error of prediction，RMSEP）为0.472，平均相对预测误差为2.69%，小于5%。结论 MARS算法建立的吸湿性预测模型更适合桂枝茯苓胶囊的生产应用，该算法可嵌入在线控制系统，为生产过程的质量控制智能化提供技术支持。

Objective The purpose of this paper was to study the differences in the stability and prediction accuracy of the moisture absorption prediction models of the contents of Guizhi Fuling Capsules (桂枝茯苓胶囊, GFC) established by different algorithms and choose the optimal modeling algorithm. Methods In this paper, 54 physical property parameters of the measured powder properties were taken as input, and the hygroscopicity of the content was taken as output. The traditional modeling algorithm partial least squares algorithm, classification and regression tree, multivariate adaptive regression splines and generalized path seeker algorithm were used to establish the moisture absorption prediction models, and the effect of different models on the prediction accuracy of the verification set was compared and studied. Results The best model was established by the multivariate adaptive regression splines (MARS) algorithm among the machine learning algorithm which had the best fitting effect and the strongest predictive ability. The calibration determination coefficient of model was 0.843, the prediction determination coefficient was 0.808, and the correction root mean square error of calibration was 0.391. The root mean square error of prediction was 0.472, and the mean relative prediction error was 2.69%. Conclusion The hygroscopicity prediction model established by the MARS algorithm was more suitable for production applications, which provided a basis for the quality control of the production process of GFC.

R283.6

国家"重大新药创制"科技重大专项（2018ZX09201010-004）