Prediction methods for time-to-event outcomes often utilize survival models that rely on strong assumptions about noninformative censoring or on how individual-level covariates and survival functions are related. When the main interest is in predicting individual-level restricted mean survival times (RMST), reliance on such assumptions can lead to poor predictive performance if these assumptions are not satisfied. We propose a generalized Bayes framework that avoids full probability modeling of all survival outcomes by using an RMST-targeted loss function that depends on a collection of inverse probability of censoring weights (IPCW). In our generalized Bayes formulation, we utilize a flexible additive tree regression model for the RMST function, and the posterior distribution of interest is obtained through model-averaging IPCW-conditional loss function-based pseudo-Bayesian posteriors. Because informative censoring can be captured by the IPCW-dependent loss function, our approach only requires one to specify a model for the censoring distribution, thereby obviating the need for complex joint modeling to handle informative censoring. We evaluate the performance of our method through a series of simulations that compare it with several well-known survival machine learning methods, and we illustrate the application of our method using a multi-site cohort of breast cancer patients with clinical and genomic covariates.

翻译：针对时间-事件结局的预测方法常采用依赖强假设的生存模型，例如非信息删失假设或个体协变量与生存函数的关系假设。当核心关注点在于预测个体水平的受限平均生存时间（RMST）时，若这些假设未被满足，将导致预测性能显著下降。我们提出一种广义贝叶斯框架，通过构建依赖于逆概率删失加权（IPCW）集合的RMST目标损失函数，避免对全部生存结局进行完整概率建模。在该广义贝叶斯公式中，我们采用灵活的加法树回归模型刻画RMST函数，并通过模型平均IPCW条件损失函数驱动的伪贝叶斯后验分布获取目标后验分布。由于信息删失可通过IPCW依赖损失函数捕获，本方法仅需指定删失分布模型，从而免除了处理信息删失所需的复杂联合建模。我们通过系列模拟实验将该方法与多种经典生存机器学习方法进行性能对比，并基于包含临床与基因组协变量的多中心乳腺癌患者队列展示实际应用效果。