Robustness of Deep Learning models in electrocardiogram noise detection and classification

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• Electrocardiogram (ECG) signal analysis is crucial for heart disease detection but is hindered by noise.
• Traditional filtering methods can distort vital ECG biomarkers.
• Existing Deep Learning (DL) methods for ECG noise lack comparative analysis of Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN) architectures.

Purpose of the Study:

• To introduce a knowledge-based ECG filtering system using DL for noise classification.
• To compare the performance and complexity of popular computer vision model architectures (CNNs and RNNs) within an Internet of Medical Things (IoMT) framework.
• To evaluate DL models for selective ECG filtering, minimizing signal distortion.

Main Methods:

• Developed a DL-based system to classify ECG noise types.
• Implemented and compared various CNN architectures (AlexNet, VGG, ResNet) and RNNs.
• Evaluated models on six datasets within a practical IoMT framework.

Main Results:

• CNN-based ECG noise classifiers outperformed RNN-based models in performance and training time.
• AlexNet, VGG, and ResNet achieved over 70% accuracy, specificity, sensitivity, and F1 score.
• VGG and ResNet showed comparable performance, with ResNet being less complex than VGG.

Conclusions:

• A DL-based ECG noise classifier enhances knowledge-driven ECG filtering systems by enabling selective filtering.
• VGG and ResNet demonstrate superior performance among evaluated CNN and RNN models.
• ResNet offers a compelling alternative to VGG, providing comparable results with reduced model complexity for ECG noise classification.