Scholarlink

Sepsis is an urgent condition to treat in the intensive care unit, where early intervention can improve chances of survival. This paper proposes a federated learning framework to predict sepsis without compromising privacy based on three large, publicly available electronic health record (EHR) data sets: MIMIC-IV, PhysioNet Sepsis Challenge 2019, and eICU. Multiple datasets provide variety to improve model robustness and generalizability. In the federated learning context, each institution's patient data stays with the institution, but encrypted model parameters are sent to a central server where a global model is created. The model features are selected based on a combination of mutual information and recursive feature elimination procedures to reduce the features to important clinically relevant features, including heart rate, blood pressure, respiratory rate, white blood cell count, and lactate. XGBoost was chosen as the model for its speed with heterogeneous and high-dimensional data. Cross-dataset validation confirmed strong generalization of the model to previously unseen hospital systems, and the model achieved 96% accuracy, outperforming random forest, logistic regression, and support vector machine baselines. The developed framework is a scalable, secure, and high-performance solution for predicting sepsis in acutely ill patients in the ICU that is suitable for real-time integration in clinical decision support.

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 Learning-Based XGBoost Framework for Privacy-Preserving Sepsis Prediction Using EHRs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jayalaxmi Dash,
  • Pratik Kumar Swain,
  • Rabinarayan Satpathy,
  • Suneeta Satpathy

摘要

Sepsis is an urgent condition to treat in the intensive care unit, where early intervention can improve chances of survival. This paper proposes a federated learning framework to predict sepsis without compromising privacy based on three large, publicly available electronic health record (EHR) data sets: MIMIC-IV, PhysioNet Sepsis Challenge 2019, and eICU. Multiple datasets provide variety to improve model robustness and generalizability. In the federated learning context, each institution's patient data stays with the institution, but encrypted model parameters are sent to a central server where a global model is created. The model features are selected based on a combination of mutual information and recursive feature elimination procedures to reduce the features to important clinically relevant features, including heart rate, blood pressure, respiratory rate, white blood cell count, and lactate. XGBoost was chosen as the model for its speed with heterogeneous and high-dimensional data. Cross-dataset validation confirmed strong generalization of the model to previously unseen hospital systems, and the model achieved 96% accuracy, outperforming random forest, logistic regression, and support vector machine baselines. The developed framework is a scalable, secure, and high-performance solution for predicting sepsis in acutely ill patients in the ICU that is suitable for real-time integration in clinical decision support.