Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

MohammadReza Hosseinzadehketilateh, Banafsheh Adami, Nima Karimian

TL;DR

This work tackles noninvasive hyperglycemia monitoring using ECG by addressing generalization to unseen subjects. It introduces a CNN enhanced with a convolutional block attention module (CBAM) to extract spatial and channel-wise dependencies from short ECG segments around R-peaks, trained on a large dataset and evaluated under strict unseen-subject splits. The study contributes a sizable database of 1,119 subjects, a segment-based preprocessing pipeline, and a CBAM-enabled architecture that achieves 91.60% AUC, 81.05% sensitivity, and 85.54% specificity on unseen data, demonstrating robust performance in realistic conditions. These results support the feasibility of wearable ECG-based hyperglycemia monitoring with strong generalization, enabling potential real-time, noninvasive management for individuals with diabetes.

Abstract

This paper presents a novel approach to noninvasive hyperglycemia monitoring utilizing electrocardiograms (ECG) from an extensive database comprising 1119 subjects. Previous research on hyperglycemia or glucose detection using ECG has been constrained by challenges related to generalization and scalability, primarily due to using all subjects' ECG in training without considering unseen subjects as a critical factor for developing methods with effective generalization. We designed a deep neural network model capable of identifying significant features across various spatial locations and examining the interdependencies among different features within each convolutional layer. To expedite processing speed, we segment the ECG of each user to isolate one heartbeat or one cycle of the ECG. Our model was trained using data from 727 subjects, while 168 were used for validation. The testing phase involved 224 unseen subjects, with a dataset consisting of 9,000 segments. The result indicates that the proposed algorithm effectively detects hyperglycemia with a 91.60% area under the curve (AUC), 81.05% sensitivity, and 85.54% specificity.

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

TL;DR

This work tackles noninvasive hyperglycemia monitoring using ECG by addressing generalization to unseen subjects. It introduces a CNN enhanced with a convolutional block attention module (CBAM) to extract spatial and channel-wise dependencies from short ECG segments around R-peaks, trained on a large dataset and evaluated under strict unseen-subject splits. The study contributes a sizable database of 1,119 subjects, a segment-based preprocessing pipeline, and a CBAM-enabled architecture that achieves 91.60% AUC, 81.05% sensitivity, and 85.54% specificity on unseen data, demonstrating robust performance in realistic conditions. These results support the feasibility of wearable ECG-based hyperglycemia monitoring with strong generalization, enabling potential real-time, noninvasive management for individuals with diabetes.

Abstract

This paper presents a novel approach to noninvasive hyperglycemia monitoring utilizing electrocardiograms (ECG) from an extensive database comprising 1119 subjects. Previous research on hyperglycemia or glucose detection using ECG has been constrained by challenges related to generalization and scalability, primarily due to using all subjects' ECG in training without considering unseen subjects as a critical factor for developing methods with effective generalization. We designed a deep neural network model capable of identifying significant features across various spatial locations and examining the interdependencies among different features within each convolutional layer. To expedite processing speed, we segment the ECG of each user to isolate one heartbeat or one cycle of the ECG. Our model was trained using data from 727 subjects, while 168 were used for validation. The testing phase involved 224 unseen subjects, with a dataset consisting of 9,000 segments. The result indicates that the proposed algorithm effectively detects hyperglycemia with a 91.60% area under the curve (AUC), 81.05% sensitivity, and 85.54% specificity.
Paper Structure (9 sections, 3 equations, 3 figures, 1 table)

This paper contains 9 sections, 3 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Experiment Setup
  3. Data Acquisition
  4. Data Processing
  5. Evaluation Metrics
  6. Model
  7. Experimental Results
  8. Discussion and comparison
  9. Conclusion

Figures (3)

  • Figure 1: (a) Demonstrate a recording of ECG waveform from a participant. (b) Shows one segment of ECG signal.)
  • Figure 2: proposed hyperglacemia detection using CNN with block attention module. a) convolutional block attention module (CBAM). b) channel attention module (CAM), c) spatial attention module (SAM), d) convolutional neural network with attention mechanism.
  • Figure 3: (a) ROC for testing dataset with AUC of 93%. (b) Depicts model accuracy for both training and testing under different epoch.)