基于随机对照试验的个体处理效应预测模型的系统评价

Individual treatment effects models based on randomized controlled trials:a systematic review

目的回顾基于随机对照试验开发的个体处理效应(individual treatment effect,ITE)模型,旨在系统总结模型开发现状,评价偏倚风险。方法计算机检索PubMed和Embase数据库,筛选1990年至2024年6月14日发表的研究。使用CHARMS清单提取数据,PROBAST风险偏倚工具评估模型质量。结果纳入11篇研究,包含19个ITE模型。ITE建模方法有含交互作用项的回归模型(n=8,42.1%)、双方程模型(n=5,26.3%)和机器学习(n=6,31.6%)。ITE模型的区分度、校准度和临床有效性报告率分别为78.9%、73.2%和10.5%。14个模型评定为高风险偏倚(73.7%),尤其是统计分析领域,原因是缺失数据处理不当(n=15,78.9%),不恰当考虑模型最优拟合问题(n=5,26.3%)等。结论ITE模型开发的常用方法有构建交互项、双方程理论和机器学习,但存在模型开发数量少,建模方法较复杂和报告不规范等问题。未来应强调对ITE模型进一步探索,鼓励多元建模方法和规范报告,以提高模型的临床推广和实际应用价值。

Objective To review individual treatment effect(ITE)models developed from randomized controlled trials,with the aim of systematically summarizing the current state of model development and assessing the risk of bias.Methods PubMed and Embase databases were searched for studies published between 1990 and 14 June 2024.Data were extracted using the CHARMS inventory,and the PROBAST risk of bias tool was used to assess model quality.Results A total of 11 publications were included,containing 19 ITE models.The ITE modelling methods were regression models with interaction terms(n=8,42.1%),dual-range models(n=5,26.3%)and machine learning(n=6,31.6%).The ITE models had a reporting rate of 78.9%,73.2%and 10.5%for differentiation,calibration and clinical validity,respectively.Fourteen models were assessed as having a high risk of bias(73.7%),particularly in the area of statistical analysis,due to inappropriate handling of missing data(n=15,78.9%),inappropriate consideration of model fit issues(n=5,26.3%),etc.Conclusion Common approaches to ITE model development include constructing interaction terms,dual procedure theory,and machine learning,but suffer from a low number of model developments,more complex modeling methods,and non-standardized reporting.In the future,emphasis should be placed on further exploration of ITE models,promoting diversified modeling methods and standardized reporting to improve the clinical promotion and practical application value of the models.