Skip to main content
QUICK REVIEW

[论文解读] Prediction-Oriented Transfer Learning for Survival Analysis

Yu (Jen) Gu, Donglin Zeng|arXiv (Cornell University)|Mar 12, 2026
Statistical Methods and Inference被引用 0
一句话总结

引入预测导向的迁移学习(POTL),将源研究中的预测知识转移以提升目标研究的预测性能,同时不共享个体层数据。

ABSTRACT

Transfer learning is beneficial for survival analysis, especially when the target study has a limited number of events. However, existing transfer learning methods rely on the restrictive assumption that the target and source studies share similar parameters under Cox models, and most require access to individual-level source data. In this article, we propose a novel transfer learning framework that enhances model-based survival prediction by transferring predictive rather than distributional knowledge from source studies. Our approach employs flexible semiparametric transformation models for the target data while eliminating the need to model or share the source data. The ingeniously designed penalty enables simple and stable computation via an EM algorithm. We rigorously establish the asymptotic properties of the proposed estimator and show that it achieves a faster convergence rate than the target-only estimator when source knowledge is sufficiently accurate. We demonstrate the advantages of our methods through extensive simulation studies and an application to two major breast cancer studies.

研究动机与目标

  • 在事件有限的目标研究中提升生存预测的动机与目标。
  • 开发一种灵活的迁移学习框架,转移的是预测知识而非分布参数。
  • 在不分享个体层数据的前提下实现对多种源模型的使用。
  • 提供一个具有理论收敛性保证的EM算法及收敛速率分析。
  • 通过仿真和乳腺癌数据应用,验证其有效性。

提出的方法

  • 针对目标使用广义的转换G的半参数转换模型。
  • 构建一个预测导向的惩罚项,使目标生存预测与汇聚的源预测S_hat(t|X)对齐。
  • 将优化形式化为在目标对数似然最大化的同时对生存预测施加类似交叉熵的惩罚。
  • 引入与当前状态数据相关联的代理惩罚,实现稳定计算的EM算法。
  • 将基线累积危险度视为阶梯函数,并在M步中显式更新其跳跃。
  • 给出渐近理论,表明POTL估计量在仅目标情况下达到同等速率,且在源预测准确时达到更快速度。

实验结果

研究问题

  • RQ1在不分享个体层数据的前提下,来自多个源研究的预测知识是否能改善目标研究的生存预测?
  • RQ2应如何量化并惩罚目标与源生存预测之间的相似性?
  • RQ3除了Cox模型,其他转换模型(如比例等级、转换模型等)在Covariate shift或模型类型差异下是否也从POTL中受益?
  • RQ4POTL的理论收敛性质是什么,实际中如何选择调参?
  • RQ5在不同源-目标模型类型和协变量配置的仿真与真实数据中,POTL的表现如何?

主要发现

SC指标POTL仅目标TransCoxCoxTL汇聚
1L2D0.037 (0.017)0.079 (0.027)0.063 (0.028)0.039 (0.031)0.021 (0.007)
10.024 (0.018)0.065 (0.029)0.049 (0.029)0.031 (0.025)0.017 (0.008)
1C-index0.581 (0.004)0.577 (0.009)0.581 (0.004)0.580 (0.005)0.581 (0.003)
1IBS0.191 (0.002)0.193 (0.003)0.192 (0.002)0.191 (0.002)0.190 (0.001)
1RMST0.650 (0.007)0.652 (0.015)0.650 (0.013)0.651 (0.008)0.651 (0.005)
2L2D0.046 (0.015)0.079 (0.027)0.063 (0.028)0.045 (0.022)0.035 (0.008)
20.045 (0.023)0.065 (0.029)0.048 (0.030)0.051 (0.023)0.045 (0.014)
2C-index0.581 (0.004)0.577 (0.009)0.580 (0.004)0.580 (0.005)0.581 (0.003)
2IBS0.191 (0.002)0.193 (0.003)0.192 (0.002)0.191 (0.002)0.190 (0.001)
2RMST0.651 (0.007)0.652 (0.015)0.651 (0.013)0.652 (0.008)0.653 (0.005)
3L2D0.057 (0.017)0.079 (0.027)0.065 (0.023)0.057 (0.021)0.049 (0.009)
30.052 (0.020)0.065 (0.029)0.050 (0.023)0.055 (0.020)0.052 (0.010)
3C-index0.580 (0.005)0.577 (0.009)0.580 (0.005)0.580 (0.005)0.581 (0.003)
3IBS0.192 (0.002)0.193 (0.003)0.192 (0.002)0.192 (0.002)0.191 (0.001)
3RMST0.649 (0.008)0.652 (0.015)0.648 (0.013)0.649 (0.008)0.648 (0.005)
4L2D0.052 (0.029)0.084 (0.030)0.066 (0.023)0.049 (0.025)0.054 (0.026)
40.040 (0.028)0.070 (0.032)0.050 (0.025)0.041 (0.022)0.064 (0.036)
4C-index0.581 (0.005)0.577 (0.009)0.580 (0.005)0.580 (0.005)0.581 (0.003)
4IBS0.191 (0.002)0.193 (0.003)0.192 (0.002)0.191 (0.002)0.190 (0.001)
4RMST0.651 (0.009)0.652 (0.015)0.653 (0.013)0.649 (0.009)0.654 (0.006)
5L2D0.053 (0.027)0.084 (0.030)0.064 (0.024)0.050 (0.023)0.058 (0.028)
50.043 (0.029)0.070 (0.032)0.050 (0.023)0.043 (0.020)0.063 (0.038)
5C-index0.580 (0.005)0.577 (0.009)0.580 (0.005)0.580 (0.005)0.581 (0.004)
5IBS0.191 (0.002)0.193 (0.003)0.192 (0.002)0.191 (0.002)0.190 (0.001)
5RMST0.653 (0.009)0.652 (0.015)0.653 (0.013)0.657 (0.009)0.656 (0.006)
  • POTL在五种情景下与目标独有、TransCox、CoxTL及汇聚分析相比,具有竞争性甚至更优的预测准确度。
  • POTL在L2D、Dτ等误差上通常表现更低,且在保持数据隐私的前提下具有与C指数、RMST等指标的竞争力。
  • 该方法在协变量漂移以及源/目标模型类型差异(包括非Cox转换)情形下仍具鲁棒性。
  • 利用代理惩罚的EM算法确保稳定、单调提升并提升计算效率。
  • 仿真与对TCGA–BRCA及METABRIC乳腺癌数据的应用证明了其实用有效性。

更好的研究,从现在开始

从论文设计到论文写作,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。