Key result
CNN micro-architectures, with a slight advantage for SqueezeNet, achieved perfect binary arrhythmia/non-arrhythmia classification using 2D multiresolution scalogram images from MIT-BIH databases.
Why the study?
Detecting and classifying cardiac arrhythmias is critical, requiring ECG beat classification CNN architectures that meet modern healthcare requirements for accuracy, latency, and energy efficiency.
Which micro-CNN architecture provides the best accuracy and efficiency for ECG beat classification using scalograms?
Which micro-CNN architecture provides the best accuracy and efficiency for ECG beat classification using scalograms?
Micro-CNN architectures, particularly SqueezeNet, combined with time-frequency scalogram analysis, can achieve perfect binary classification of arrhythmias from ECG signals.
Answers are grounded in this paper and cite their sources.
Scalogram-based micro-CNNs may enable efficient on-device ECG arrhythmia classification; leaves open optimal architecture selection for clinical deployment.
Detecting and classifying cardiac arrhythmias is critical to the diagnosis of patients with cardiac abnormalities. In this work, we compare different convolutional neural network (CNN) architectures for electrocardiogram (ECG) beat classification that meet the accuracy, latency, and energy efficiency required for modern healthcare systems. In this paper, we compare six micro-CNN architectures for the classification of ECG signals using the scalogram. Calculated from the absolute value of the continuous wavelet transform (CWT) coefficients from a signal, a scalogram provides time–frequency analysis, which has been proven an efficient way for analyzing noisy signals. We classify the resulting 2D multiresolution scalogram images with pre-trained models by transfer learning, such as MobileNetV2, SqueezeNet, ShuffleNet, GoogleNet, EfficientNet, and ResNet-18, and compare their performances. These models have a number of parameters ranging from 1.24 million to 11.7 million to consider the performance/efficiency compromise. The effectiveness of our model is illustrated using three widely used MIT-BIH databases, and the accuracies are quite close to each other with a slight advantage for the SqueezeNet, which has obtained a perfect score for all metrics, and all models have obtained a perfect binary arrhythmia/non-arrhythmia classification. Results also demonstrate the advantage of time–frequency ECG analysis. This AI-based approach could help mass screening people suffering from heart problems, save time for clinicians, and help them with an accurate diagnosis.
No takes yet. Share an insight, caveat, or question.
Benmalek et al. (2021) studied Cardiac arrhythmias. CNN micro-architectures (e.g., SqueezeNet) using scalograms vs. Other CNN architectures was evaluated on ECG beat classification accuracy. CNN micro-architectures, with a slight advantage for SqueezeNet, achieved perfect binary arrhythmia/non-arrhythmia classification using 2D multiresolution scalogram images from MIT-BIH databases.
Synapse has enriched 10 related papers. Consider them for comparative context: