Diabetes: Non-Invasive Blood Glucose Monitoring Using Federated Learning with Biosensor Signals

Area of Science:

• Biomedical Engineering
• Machine Learning
• Data Science

Background:

• Diabetes management requires accurate blood glucose level (BGL) monitoring to prevent severe complications.
• Traditional BGL monitoring methods (e.g., finger-prick tests) suffer from low patient adherence due to invasiveness.
• Non-invasive BGL monitoring techniques are crucial for improving patient compliance and long-term health outcomes.

Purpose of the Study:

• To develop a privacy-preserving, non-invasive method for blood glucose level (BGL) monitoring using photoplethysmography (PPG) signals.
• To leverage federated learning (FL) for collaborative training of deep neural networks (DNNs) across multiple institutions without data sharing.
• To enhance the accuracy and clinical reliability of non-invasive BGL prediction through advanced signal processing and feature selection.

Main Methods:

• Photoplethysmography (PPG) signals were preprocessed using continuous wavelet transform (CWT) for noise reduction and adaptive cycle-based segmentation (ACBS).
• Particle swarm optimization (PSO) was employed for feature selection to enhance classification accuracy.
• A federated learning (FL) framework enabled collaborative training of deep neural networks (DNNs) on distributed datasets from multiple healthcare organizations.

Main Results:

• The proposed FL-based system achieved a root mean square error (RMSE) of 19.1 mg/dL on diverse datasets (VitalDB, MUST).
• Clarke error grid analysis (CEGA) demonstrated high clinical reliability, with 99.31% of predictions within acceptable ranges.
• The federated learning approach significantly outperformed conventional centralized deep learning models in predictive accuracy.

Conclusions:

• Federated learning offers a robust solution for privacy-preserving, collaborative training of non-invasive glucose monitoring models.
• The integration of CWT, ACBS, and PSO with FL enhances the accuracy and clinical utility of PPG-based BGL prediction.
• This approach represents a significant advancement towards widespread adoption of non-invasive glucose monitoring for diabetes management.