Advancements in Continuous Glucose Monitoring: Integrating Deep Learning and ECG Signal

Area of Science:

• Biomedical Engineering
• Cardiology
• Medical Informatics

Background:

• Noninvasive hyperglycemia monitoring remains a significant challenge.
• Existing electrocardiogram (ECG)-based glucose detection methods face generalization and scalability issues.
• Effective generalization requires models trained and tested on unseen subjects.

Purpose of the Study:

• To develop a novel deep neural network for noninvasive hyperglycemia detection using ECG.
• To address limitations in generalization and scalability of previous ECG-based glucose monitoring approaches.
• To validate the model's performance on a dataset including unseen subjects.

Main Methods:

• A deep neural network was designed to identify spatial features and interdependencies within ECG data.
• ECG signals were segmented to isolate individual heartbeats for accelerated processing.
• The model was trained on 727 subjects, validated on 168, and tested on 224 unseen subjects (9,000 segments).

Main Results:

• The proposed algorithm achieved a high area under the curve (AUC) of 93.05%.
• The model demonstrated a sensitivity of 83.46% and a specificity of 86.04% in detecting hyperglycemia.
• The approach showed effective generalization capabilities on unseen subject data.

Conclusions:

• The developed deep neural network offers a promising noninvasive method for hyperglycemia monitoring.
• The findings suggest that ECG-based analysis can be a scalable and generalizable approach for glucose level detection.
• This novel technique has the potential to improve diabetes management and patient care.