Deep Multi-Scale Fusion Neural Network for Multi-Class Arrhythmia Detection.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Automated electrocardiogram (ECG) analysis is crucial for early cardiovascular disease diagnosis.
• Current methods struggle with feature extraction from noisy ECG signals with variable rhythms.
• Existing research often overlooks complementary information from different signal scales.

Purpose of the Study:

• To develop a novel end-to-end deep learning architecture for multi-class arrhythmia detection.
• To enhance feature extraction from raw ECG signals by incorporating multi-scale information.
• To improve the accuracy and generalization of automated arrhythmia classification.

Main Methods:

• Proposed a Deep Multi-Scale Fusion convolutional neural network (DMSFNet) architecture.
• Implemented multi-scale feature extraction using convolution kernels with different receptive fields.
• Employed a joint optimization strategy with multiple losses for cumulative multi-scale feature learning.

Main Results:

• DMSFNet achieved state-of-the-art performance on two public datasets (CPSC_2018 and PhysioNet/CinC_2017).
• Demonstrated superior F1 scores on both 12-lead and single-lead ECG datasets compared to previous methods.
• Showcased effective noise suppression and capture of abnormal cardiac patterns.

Conclusions:

• The proposed DMSFNet excels in extracting discriminative features for a wide range of arrhythmias.
• The architecture exhibits strong generalization ability for ECG signals across different leads.
• This deep multi-scale fusion approach offers a promising direction for automated ECG analysis and arrhythmia detection.