Predicting Dysglycemia in Patients with Diabetes Using Electrocardiogram.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Continuous blood glucose monitoring is crucial for managing diabetes.
• Non-invasive methods for predicting dysglycemia are highly sought after.
• Electrocardiography (ECG) offers a potential non-invasive data source.

Purpose of the Study:

• To explore the potential of ECG for non-invasive dysglycemia prediction.
• To develop and validate an AI model for predicting blood glucose levels.
• To assess the accuracy and lead time of ECG-based dysglycemia prediction.

Main Methods:

• Collected ECG data from patients with diabetes.
• Extracted heart rate variability (HRV) features from ECG signals.
• Utilized a residual block-based 1D convolutional neural network (CNN) for prediction.
• Determined optimal feature set size through feature elimination.

Main Results:

• The AI model accurately predicted dysglycemia at various time points (measurement, 15 min prior, 30 min prior).
• Optimal prediction was achieved using 12 HRV features.
• Dysglycemia prediction 30 minutes prior showed high accuracy (90.5%), sensitivity (87.52%), specificity (92.74%), and precision (89.86%).

Conclusions:

• ECG data alone can effectively predict dysglycemia.
• The developed AI model performs comparably to methods using multiple vital signs.
• This non-invasive ECG-based approach shows significant promise for diabetes management.