Privacy Preserved Blood Glucose Level Cross-Prediction: An Asynchronous Decentralized Federated Learning Approach.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Diabetes Technology

Background:

• Newly diagnosed Type 1 Diabetes (T1D) patients face challenges with effective blood glucose (BG) prediction due to insufficient continuous glucose monitoring (CGM) data, creating a
• cold start
• problem.
• Training population models for BG prediction is hindered by privacy concerns and the decentralized nature of patient data stored on personal devices.

Purpose of the Study:

• To introduce GluADFL (blood Glucose prediction by Asynchronous Decentralized Federated Learning) as a privacy-preserving solution for the T1D
• cold start
• problem.
• To evaluate GluADFL's effectiveness in cross-patient BG prediction using CGM data.

Main Methods:

• Implemented an asynchronous decentralized federated learning framework (GluADFL) for BG prediction.
• Compared GluADFL against eight baseline methods using four T1D datasets from 298 participants.
• Assessed model performance across various network topologies (random, cluster, ring) and participant inactivity levels (up to 70% inactive).

Main Results:

• GluADFL demonstrated superior performance in accurate BG level prediction for cross-patient analysis compared to baseline methods.
• The model showed stability even with up to 70% of participants being inactive.
• Effective data sharing across diverse communication networks was achieved.

Conclusions:

• GluADFL provides a practical and privacy-preserving approach to BG prediction for T1D management.
• The method significantly enhances the quality of diabetes care by addressing data scarcity and privacy concerns.
• The asynchronous and decentralized nature of GluADFL allows for flexible and robust patient participation.