目的 评价采用机器学习（ML）算法构建的药物相互作用（DDI）风险预测模型性能，为临床应用与科研提供参考。方法 检索PubMed、Embase、中国学术期刊全文数据库（CNKI）、万方数据知识服务平台、维普数据库（VIP），收集截至2023年1月31日的相关文献，并使用PROBAST工具评估模型质量。结果 共纳入54个DDI预测模型，主要使用Drugbank、Twosides数据库构建模型，共涉及21种ML算法，以图神经网络和深度神经网络使用最频繁。药物结构特征是最常用的预测因子，药时曲线下面积（AUC）为0.83～0.99。所有模型存在较高的偏倚风险，主要源于信息偏倚和黑盒效应，但整体适用性风险低。结论 基于ML构建DDIs风险预测模型对临床用药有一定参考价值，但模型质量亟待提高，未来应开发更具可解释性的模型并验证其临床实用性。

Objective To evaluate the performance of drug-drug interactions (DDI) risk prediction models constructed by machine learning (ML) algorithm, and to provide reference for clinical application and scientific research. Methods PubMed, Embase, WanFang Data, CNKI and VIP databases were electronically searched to retrieve all ML studies on predicting DDI from inception to January 31st, 2023. PROBAST tool was used to evaluate model quality. Results A total of 54 DDI prediction models were included. The models were mainly constructed using Drugbank and Twosides databases, involving 21 ML algorithms. Figure neural network and deep neural network were used most frequently. Drug structure characteristics were the most commonly used predictors, the AUC range of 0.83 to 0.99. All models have a high risk of bias, mainly due to information bias and black box effect, but the overall suitability risk was low. Conclusions Building DDI risk prediction models based on ML has certain reference value for clinical drug use, but the quality of the model needs to be improved. In the future, more interpretable models should be developed and their clinical practicability should be verified.

R969.2

中国药品监督管理研究会研究课题-基于多元证据体探索中成药安全性评价方法的研究(2024-Y-Y-006)；临床研究和成果转化能力提升试点项目–中药制剂研发——治疗胃轻瘫中药复方佛香散(DZMG-ZJXY-23002)