ECG based human activity-specific cardiac pattern detection using machine-learning and deep-learning models.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Cardiac monitoring is vital for detecting stress and abnormalities.
• Analyzing electrocardiogram (ECG) signals during varied activities simulates real-world stressors.
• Non-linear dynamics in cardiac behavior are key indicators of cardiac health.

Purpose of the Study:

• To develop and compare machine learning models for classifying cardiac patterns under stress.
• To evaluate the efficacy of deep learning and feature-based approaches for ECG analysis.
• To identify significant cardiac features indicative of autonomic responses to stressors.

Main Methods:

• ECG signals were collected during rest, cognitive tasks, and physical activities.
• A feature-based Random Forest (RF) model used time, frequency, and statistical features.
• An image-based Convolutional Neural Network (CNN) model utilized Continuous Wavelet Transform (CWT) scalograms and bispectrum-based contours.

Main Results:

• The RF model achieved 96.80% accuracy and 92.22% F1-score.
• The CNN model with CWT scalograms reached 98.44% accuracy and 96.11% F1-score.
• The CNN model with bispectrum-based contours achieved the highest accuracy (99.16%) and F1-score (97.89%).

Conclusions:

• The CNN model employing bispectrum-based contours demonstrates superior performance in classifying cardiac patterns.
• This deep learning approach shows significant potential for reliable cardiac function monitoring across diverse activities.
• Key features like stress index and SNS-to-PNS ratio effectively highlight autonomic responses to cognitive and physical stressors.