目的 识别禁忌表述尚不明确中药中的妊娠期禁忌药，并将其划分为禁用药、忌用药和慎用药，为妊娠期妇女安全用药提供依据。方法 以《中国药典临床用药须知》2015年版收载的666味中药为研究对象，选取其中“禁忌慎”分类明确的药物，基于神经网络、支持向量机、朴素贝叶斯和随机森林4种机器学习算法建模。根据交叉验证的受试者工作特征（receiver operating characteristic，ROC）曲线下面积（area under ROC curve，AUC）与F1分数（F1 score，F1）评价模型优劣，筛选相对最优算法建立妊娠期中药“禁忌慎”判别模型，并应用该模型预测禁忌表述尚不明确的药物。结果 纳入224种药物建模，模型由基于神经网络的中药禁忌与否判别模型（AUC＝0.952，F1＝0.885）、基于支持向量机的禁忌药禁用与否判别模型（AUC＝0.912，F1＝0.779）、基于朴素贝叶斯的禁忌药忌用与否判别模型（AUC＝0.843，F1＝0.333）和基于神经网络的禁忌药慎用与否判别模型（AUC＝0.932，F1＝0.877）构成。应用模型成功划分442种禁忌表述尚不明确的中药，预测提示妊娠期禁忌药163种，禁用药8种，忌用药1种，慎用药134种。结论 构建的判别模型具有良好的稳健性与预测能力，适用于妊娠期中药“禁忌慎”判别评估，指导临床用药决策与实践，提高临床合理用药水平。

Objective To provide evidence for the safe medication of pregnant women, the identification of contraindications in traditional Chinese medicine (TCM) during pregnancy was not clear, and divided them into prohibited drugs, avoided drugs and cautious drugs. Methods A total of 666 TCMs collected in the 2015 edition of Instructions for Clinical Medications were selected as the research objects. The drugs with clear classification of "prohibiting, avoiding or using caution" were selected from the research objects, and four machine learning algorithms, neural network, support vector machine, naive Bayes and random forest, were used for modeling. According to the cross-validated area under receiver operating characteristic (ROC) curve (AUC) and F1 score (F1) to evaluate the pros and cons of the model, and to screen the relatively optimal machine learning algorithm to establish the "prohibiting, avoiding or using caution" discriminant mode of TCM during pregnancy, and the model was used to predict the drugs whose contraindication was not clear. Results A total of 224 kinds of drugs were included to establish the model. The models consisted of a neural network-based discrimination model for the contraindications of TCM (AUC=0.952, F1=0.885), a support vector machine-based discrimination model for the prohibited drugs of contraindications (AUC=0.912, F1=0.779), a naive Bayes-based discrimination model for the avoided drugs of contraindications (AUC=0.843, F1=0.333), and a neural network-based discrimination model for the cautious drugs of contraindications (AUC=0.932, F1=0.877). The application model successfully divided 442 kinds of TCM whose contraindications were not clear. The prediction suggested that there were 163 kinds of contraindicated drugs during pregnancy, eight kinds of prohibited drugs, one kind of avoided drugs and 134 kinds of cautious drugs. Conclusion The discriminant models have good robustness and predictive ability. They are suitable for the discriminant evaluation of the "prohibiting, avoiding or using caution" in TCMs during pregnancy. They could guide the decision-making and practice of clinical medication, and improve the level of clinical rational medication.

R285.64

国家中医药管理局-国家中医药领军人才支持计划“岐黄学者”项目（10400633210004）；第三批国家高层次人才特殊支持计划（万人计划）教学名师项目（2020063320001）