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

Che Liu, Sibo Cheng, Weiping Ding, Rossella Arcucci

Abstract

Electrocardiography (ECG) classification is a critical task in cardiac health monitoring. Traditional deep learning models often treat time and spectral domains independently or rely on fixed frequency filters, potentially losing vital diagnostic information. This paper presents the Spectral Cross-domain Neural Network (SCDNN), which incorporates a Soft-adaptive Threshold Spectral Enhancement (SATSE) block to enable learnable, cross-domain feature fusion. By replacing rigid filters with trainable soft-thresholding, SCDNN achieves state-of-the-art performance on major benchmarks like PTB-XL and CPSC2018 while maintaining high computational efficiency.

Introduction

ECG signals are multi-variable time series that contain rich information in both time and frequency domains. While CNNs and LSTMs have improved automated classification, they often fail to fully exploit the relationship between these domains. Conventional spectral analysis uses "hard" thresholds (e.g., low-pass filters) that can destroy subtle high-frequency patterns necessary for specific diagnoses. SCDNN addresses this by introducing an internal communication mechanism where the network itself determines the importance of various frequency bands through learnable parameters.

Methods

The SCDNN architecture is built upon a ResNet18 backbone, modified with specialized SATSE blocks within each residual stage.

Architecture of the SCDNN model showing the integration of SATSE blocks.

The SATSE block functions by transforming time-domain features into the spectral domain via Fast Fourier Transform (FFT). It then applies a soft-adaptive thresholding mechanism using a modified Sigmoid function with trainable parameters. This allows the model to adaptively filter frequency bands based on the specific diagnostic task. Filtered features are processed through a trainable weight matrix to capture inter-frequency dependencies and then converted back to the time domain via Inverse FFT (IFFT) for residual addition. This design allows the model to refine its understanding of the heartbeat's shape using insights about its underlying rhythms.

Results

The SCDNN was evaluated on the PTB-XL (21,837 records) and CPSC2018 (6,877 records) datasets.

Comparison of SCDNN against state-of-the-art methods across multiple metrics on the PTB-XL dataset.

SCDNN consistently outperformed existing models, reaching an accuracy of 0.835 on PTB-XL and 0.859 on CPSC2018. The model also achieved high F1-scores (0.792 on PTB-XL), indicating a robust balance between precision and recall. Crucially, SCDNN demonstrated exceptional efficiency, recording the second-lowest inference time among compared models. This high performance combined with low latency highlights its potential for deployment in real-time wearable monitoring devices.

Conclusion

The Spectral Cross-domain Neural Network successfully bridges the gap between time and frequency domain analysis in ECG classification. By utilizing trainable soft filters, the model preserves critical signal information that fixed filters might discard. Future research will focus on generalizing this cross-domain architecture to other time-series domains and exploring alternative spectral transformations beyond the standard Fourier Transform.

Summary updated

Abstract

Electrocardiography (ECG) signals can be considered as multi-variable time-series. The state-of-the-art ECG data classification approaches, based on either feature engineering or deep learning techniques, treat separately spectral and time domains in machine learning systems. No spectral-time domain communication mechanism inside the classifier model can be found in current approaches, leading to difficulties in identifying complex ECG forms. In this paper, we proposed a novel deep learning model named Spectral Cross-domain neural network (SCDNN) with a new block called Soft-adaptive threshold spectral enhancement (SATSE), to simultaneously reveal the key information embedded in spectral and time domains inside the neural network. More precisely, the domain-cross information is captured by a general Convolutional neural network (CNN) backbone, and different information sources are merged by a self-adaptive mechanism to mine the connection between time and spectral domains. In SATSE, the knowledge from time and spectral domains is extracted via the Fast Fourier Transformation (FFT) with soft trainable thresholds in modified Sigmoid functions. The proposed SCDNN is tested with several classification tasks implemented on the public ECG databases \textit{PTB-XL} and \textit{MIT-BIH}. SCDNN outperforms the state-of-the-art approaches with a low computational cost regarding a variety of metrics in all classification tasks on both databases, by finding appropriate domains from the infinite spectral mapping. The convergence of the trainable thresholds in the spectral domain is also numerically investigated in this paper. The robust performance of SCDNN provides a new perspective to exploit knowledge across deep learning models from time and spectral domains. The repository can be found: https://github.com/DL-WG/SCDNN-TS
BibTeX
@article{liu2023spectral,
  author = {Che Liu and Sibo Cheng and Weiping Ding and Rossella Arcucci},
  title = {{Spectral Cross-Domain Neural Network With Soft-Adaptive Threshold Spectral Enhancement}},
  journal = {IEEE Trans. Neural Networks Learn. Syst.},
  year = {2023},
}
Figure