Neural gradient boosting in federated learning for hemodynamic instability prediction: towards a distributed and...

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Federated learning (FL) addresses data privacy and security challenges in healthcare by enabling collaborative model training without data sharing.
• Current FL architectures primarily support neural networks, lacking scalability and efficient integration of decision tree-based models, especially for sensitive clinical data.
• Decision tree models are state-of-the-art for tabular clinical data but face performance and security issues when implemented in FL due to encryption costs.

Purpose of the Study:

• To develop and evaluate a federated learning pipeline that enables distributed gradient boosting for predicting hemodynamic instability in intensive care unit (ICU) patients.
• To enhance FL architectures to support decision tree models, overcoming limitations of existing neural network-centric approaches.
• To assess the performance of the proposed FL approach against a centralized setup using a multi-hospital clinical dataset.

Main Methods:

• Implemented a federated learning pipeline incorporating distributed gradient boosting for hemodynamic instability prediction.
• Designed the pipeline to support both neural-based boosting models and conventional neural networks, enabling decision tree integration within FL.
• Utilized a clinical dataset from 25 hospitals, derived from the Philips eICU database, for model training and validation.

Main Results:

• Achieved comparable performance in accuracy, precision, recall, and F1 score for hemodynamic instability detection using FL versus a centralized setup.
• Demonstrated the feasibility of a scalable FL approach for analyzing ICU data while preserving patient privacy.
• Validated the effectiveness of integrating decision tree models into FL for clinical classification tasks.

Conclusions:

• Federated learning can be effectively scaled for detecting hemodynamic instability in ICU patients, maintaining data privacy.
• The developed FL pipeline successfully integrates decision tree models, offering advantages over traditional neural network-based FL architectures for tabular clinical data.
• This work paves the way for broader adoption of advanced machine learning techniques in privacy-preserving healthcare analytics.