A personalized federated learning-based glucose prediction algorithm for high-risk glycemic excursion regions in type...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Data Privacy

Background:

• Continuous glucose monitoring (CGM) enables real-time glucose readings for better glycemic control.
• Predicting rare glycemic excursions (hypoglycemia and hyperglycemia) remains a significant challenge.
• Limited access to sensitive patient data hinders the development of robust machine learning models.

Purpose of the Study:

• To develop accurate glucose predictions in excursion regions while addressing data privacy concerns.
• To introduce a novel Hypo-Hyper (HH) loss function for improved prediction accuracy at glucose extremes.
• To propose FedGlu, a federated learning (FL) model for collaborative learning without compromising patient data.

Main Methods:

• Proposed a novel Hypo-Hyper (HH) loss function to penalize prediction errors, with higher penalties at glycemic extremes.
• Developed FedGlu, a machine learning model utilizing a federated learning (FL) framework for privacy-preserving collaborative training.
• Integrated the HH loss function within the FedGlu FL framework to simultaneously address prediction accuracy and data privacy.

Main Results:

• The HH loss function showed a 46% improvement over mean-squared error (MSE) loss across 125 patients.
• FedGlu improved glycemic excursion detection by 35% compared to local models.
• Enhanced prediction of hypoglycemia and hyperglycemia for 105 out of 125 patients.

Conclusions:

• The proposed HH loss function effectively enhances the predictive capabilities of glucose monitoring.
• Implementing models within a federated learning framework ensures improved predictive performance while safeguarding sensitive patient data.
• The combined HH loss and FedGlu approach offers a promising solution for accurate and private glucose excursion prediction.