Scholarlink

<p>This paper focuses on the inference of individualized treatment rules(ITRs) in personalized medicine, particularly addressing the challenges of covariate distribution differences and external validity encountered in real-world applications. To overcome these issues, we propose two frameworks, Transfer-Dragonnet(TDN) and Adaptive Transfer-Dragonnet(ATDN), which integrate neural networks with transfer learning. TDN corrects covariate distribution differences by estimating participation probabilities and incorporating them into a weighted learning framework. We then provide a detailed discussion of the asymptotic properties of the relevant estimator, ensuring the theoretical validity of TDN. Building upon this, ATDN further improves stability by directly optimizing transfer weights, reducing the impact of extreme propensity scores and better handling covariate shifts. Through simulations, we use mean squared error(MSE) to assess estimation accuracy and find that ATDN performs best, especially under significant covariate shifts. Furthermore, in the real-world application of the Tennessee Student/Teacher Achievement Ratio(STAR) project, ATDN derived an optimized small-class assignment strategy based on math scores, demonstrating its practical effectiveness.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Individualized treatment rules based on adaptive transfer-dragonnet

  • 期刊论文

  • 发表日期：

    2025年8月7日

查看全文
  • Zhonghe Huang,
  • Yujia Song,
  • Qi Zhang

摘要

This paper focuses on the inference of individualized treatment rules(ITRs) in personalized medicine, particularly addressing the challenges of covariate distribution differences and external validity encountered in real-world applications. To overcome these issues, we propose two frameworks, Transfer-Dragonnet(TDN) and Adaptive Transfer-Dragonnet(ATDN), which integrate neural networks with transfer learning. TDN corrects covariate distribution differences by estimating participation probabilities and incorporating them into a weighted learning framework. We then provide a detailed discussion of the asymptotic properties of the relevant estimator, ensuring the theoretical validity of TDN. Building upon this, ATDN further improves stability by directly optimizing transfer weights, reducing the impact of extreme propensity scores and better handling covariate shifts. Through simulations, we use mean squared error(MSE) to assess estimation accuracy and find that ATDN performs best, especially under significant covariate shifts. Furthermore, in the real-world application of the Tennessee Student/Teacher Achievement Ratio(STAR) project, ATDN derived an optimized small-class assignment strategy based on math scores, demonstrating its practical effectiveness.