Spectral Cross-Domain Neural Network With Soft-Adaptive Threshold Spectral Enhancement

Area of Science:

• Biomedical Signal Processing
• Machine Learning
• Deep Learning

Background:

• Current electrocardiography (ECG) classification methods often process spectral and time domains separately.
• This separation limits the ability to identify complex ECG patterns due to a lack of spectral-time domain communication within models.

Purpose of the Study:

• To propose a novel deep learning model, the spectral cross-domain neural network (SCDNN), for enhanced ECG signal classification.
• To develop a new block, soft-adaptive threshold spectral enhancement (SATSE), for simultaneously extracting information from spectral and time domains.

Main Methods:

• Developed SCDNN, a deep learning model utilizing a CNN backbone to capture cross-domain information.
• Integrated SATSE block employing Fast Fourier Transformation (FFT) with soft trainable thresholds for knowledge extraction.
• Employed a self-adaptive mechanism to merge information from time and spectral domains.

Main Results:

• SCDNN demonstrated superior performance across multiple classification tasks on the PTB-XL and CPSC2018 ECG databases.
• The model achieved state-of-the-art results with significantly lower computational cost compared to existing approaches.
• Numerical investigation confirmed the convergence of trainable thresholds within the spectral domain.

Conclusions:

• The proposed SCDNN model offers a new perspective for exploiting cross-domain knowledge in deep learning for time series analysis.
• SCDNN effectively integrates spectral and time domain information, leading to robust ECG classification.
• The findings suggest a promising direction for improving the accuracy and efficiency of ECG analysis.