Scholarlink

<p>This study introduces a novel federated deep learning framework designed for multi-institutional healthcare networks, ensuring robust privacy while delivering superior diagnostic accuracy. Unlike conventional centralized models that expose sensitive patient data, the proposed approach integrates adaptive model personalization, multi-tier aggregation, and secure update protocols to safeguard confidentiality under heterogeneous environments. Experiments across imaging, physiological, and synthetic EHR datasets show the framework achieving 94.1% accuracy, 93.5% precision, 92.7% recall, 93.1% F1-score, and ROC-AUC of 0.981, while also recording reduced training time of 6&#xa0;h, only 115 communication rounds, and the lowest Brier Score of 0.061. Furthermore, the system exhibits a privacy score of 5, a personalization score of 10, and enhanced fairness with minimized demographic disparities. These results confirm that the proposed solution not only strengthens diagnostic reliability but also achieves efficiency, scalability, and equity, making it highly practical for real-world deployment in collaborative healthcare AI.</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

Federated deep learning for privacy-preserving diagnosis in multi-institutional healthcare networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月2日

查看全文
  • Princy Diwan,
  • Ramgopal Kashyap,
  • Soubhagya Ranjan Mallick,
  • Mukesh Bathre,
  • Jameer Kotwal

摘要

This study introduces a novel federated deep learning framework designed for multi-institutional healthcare networks, ensuring robust privacy while delivering superior diagnostic accuracy. Unlike conventional centralized models that expose sensitive patient data, the proposed approach integrates adaptive model personalization, multi-tier aggregation, and secure update protocols to safeguard confidentiality under heterogeneous environments. Experiments across imaging, physiological, and synthetic EHR datasets show the framework achieving 94.1% accuracy, 93.5% precision, 92.7% recall, 93.1% F1-score, and ROC-AUC of 0.981, while also recording reduced training time of 6 h, only 115 communication rounds, and the lowest Brier Score of 0.061. Furthermore, the system exhibits a privacy score of 5, a personalization score of 10, and enhanced fairness with minimized demographic disparities. These results confirm that the proposed solution not only strengthens diagnostic reliability but also achieves efficiency, scalability, and equity, making it highly practical for real-world deployment in collaborative healthcare AI.