不同纵向信息利用策略下肝细胞癌患者动态生存率预测

Dynamic Survival Probability Prediction of Hepatocellular Carcinoma Patients under Different Longitudinal Information Utilization Strategies

摘要: 动态生存预测能够随患者随访信息更新个体风险评估，但不同纵向信息利用策略模型的预测表现仍有待系统比较。本文基于肝细胞癌患者的多次随访数据，分别构建仅使用基线协变量的Cox比例风险模型、在随访时点更新协变量的时间依赖Cox模型以及联合建模纵向轨迹与生存结局的联合模型，并在联合模型中比较当前值、变化率与累积效应三种关联结构。围绕动态预测任务，设计多预测起点与多预测窗口的评价框架，采用时间依赖C-index与Brier score综合评估模型区分能力与概率预测误差。研究表明，纵向随访信息有助于提升动态风险刻画能力，所提出的动态预测任务评价框架可为动态生存模型的系统比较提供参考。

Abstract: Dynamic survival prediction allows individual risk estimates to be continuously updated as follow-up information accumulates, yet the predictive performance of different longitudinal information utilization strategies has not been systematically compared. Based on follow-up data from patients with hepatocellular carcinoma, this study constructed a Cox proportional hazards model using baseline covariates only, a time-dependent Cox model with covariates updated at follow-up time points, and joint models that explicitly link longitudinal trajectories with survival outcomes, and further compared three association structures within the joint modeling framework, including the current value, slope, and cumulative effect. To match the dynamic prediction task, a multi-prediction-time and multi-prediction-window evaluation framework was designed, and model performance was assessed using the time-dependent C-index and Brier score to quantify discrimination ability and probability prediction error. These findings demonstrate the value of longitudinal follow-up information for dynamic risk assessment and suggest that the proposed evaluation framework provides a practical reference for the systematic comparison of dynamic survival models.