临床预测模型中处理时依性变量的策略及进展

Methodology and progress in adjusting time-dependent covariates in clinical prediction models

预测模型中考虑时依性变量可改善模型的总体表现,提高其临床应用价值。界标模型、联合模型等基于传统回归策略在处理时依性变量个数和适用情境等方面存在局限,神经网络等机器学习算法有望对其灵活处理。本文针对传统模型、机器学习算法,总结各自纳入时依性变量的建模思路,梳理各方法的适用场景,概括现有方法仍存在的问题,以期为未来预测建模处理时依性变量提供方法学启示。

Adjusting time-dependent covariates into prediction models may help improve model performance and expand clinical applications.The methodology of handling time-dependent covariates is limited in traditional regression strategies(i.e.,landmark model,joint model).For example,the number of predictors and practical situations which can be handled are restricted when using regression models.One new strategy is to use machine learning(i.e.,neural networks).This review summarizes the methodology of handling time-dependent covariates in prediction models,such as applicable scenarios,strengths,and limitations,to offer methodological enlightenment for processing time-dependent covariates.