Scholarlink

<p>The sensitive nature of electronic health records (EHR) and wearable data presents challenges in sharing biomedical resources while minimizing re-identification risks. This article introduces an end-to-end, titratable pipeline that generates privacy-preserving “digital twin” datasets from complex EHR and wearable-device records (Apple Watch data from 3029 participants) using DataSifter and Synthetic Data Vault (SDV) methods. Various obfuscation levels were applied (DataSifter: small, medium, large; SDV: CTGAN, Gaussian Copula) and benchmarked using utility (statistical fidelity, machine learning performance) and privacy (re-identification risk, detection likelihood) metrics. The highest-obfuscation DataSifter twin delivered the strongest privacy protection (0.83) while preserving key statistical and predictive signals (83.1% confidence interval overlap in regression models), outperforming SDV, particularly for longitudinal data. Despite declining performance in machine learning tasks with higher obfuscation, utility was generally preserved. The study underscores the importance of digital twin datasets and highlights DataSifter’s adaptability in privacy-utility trade-offs, advocating its utility for secure data sharing.</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

Medical data sharing and synthetic clinical data generation – maximizing biomedical resource utilization and minimizing participant re-identification risks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月16日

查看全文
  • Simeone Marino,
  • Ruth Cassidy,
  • Joseph Nanni,
  • Yuxuan Wang,
  • Yipeng Liu,
  • Mingyi Tang,
  • Yuan Yuan,
  • Toby Chen,
  • Anik Sinha,
  • Balaji Pandian,
  • Ivo D. Dinov,
  • Michael L. Burns

摘要

The sensitive nature of electronic health records (EHR) and wearable data presents challenges in sharing biomedical resources while minimizing re-identification risks. This article introduces an end-to-end, titratable pipeline that generates privacy-preserving “digital twin” datasets from complex EHR and wearable-device records (Apple Watch data from 3029 participants) using DataSifter and Synthetic Data Vault (SDV) methods. Various obfuscation levels were applied (DataSifter: small, medium, large; SDV: CTGAN, Gaussian Copula) and benchmarked using utility (statistical fidelity, machine learning performance) and privacy (re-identification risk, detection likelihood) metrics. The highest-obfuscation DataSifter twin delivered the strongest privacy protection (0.83) while preserving key statistical and predictive signals (83.1% confidence interval overlap in regression models), outperforming SDV, particularly for longitudinal data. Despite declining performance in machine learning tasks with higher obfuscation, utility was generally preserved. The study underscores the importance of digital twin datasets and highlights DataSifter’s adaptability in privacy-utility trade-offs, advocating its utility for secure data sharing.