Scholarlink

Achieving a balance between the advantages of machine learning and the protection of patient confidentiality is a major obstacle in clinical reinforcement learning. We evaluate three reinforcement learning algorithms: PPO, SAC, and A2C: under five differential privacy mechanisms using anonymized MIMIC-III intensive care unit data for ventilation and sedation treatments. The models were trained with a specified privacy budget of \(\epsilon \)  = 1.0 and \(\delta \)  =  \(1 \times 10^{-5}\) . SAC achieved an accuracy of up to 97’ in ventilation control, and A2C with Exponential or Rényi DP performed exceptionally well in sedation. Our findings underscore the significance of choosing suitable combinations of RL,DP pairs for clinical decision support systems.

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

Differentially Private RL for ICU Decision Support: A Comparative Analysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Rakesh Batchala,
  • Priyank Jain,
  • Manasi Gyanchandani,
  • Sanyam Shukla

摘要

Achieving a balance between the advantages of machine learning and the protection of patient confidentiality is a major obstacle in clinical reinforcement learning. We evaluate three reinforcement learning algorithms: PPO, SAC, and A2C: under five differential privacy mechanisms using anonymized MIMIC-III intensive care unit data for ventilation and sedation treatments. The models were trained with a specified privacy budget of \(\epsilon \)  = 1.0 and \(\delta \)  =  \(1 \times 10^{-5}\) . SAC achieved an accuracy of up to 97’ in ventilation control, and A2C with Exponential or Rényi DP performed exceptionally well in sedation. Our findings underscore the significance of choosing suitable combinations of RL,DP pairs for clinical decision support systems.