Spectral Cross-Domain Neural Network with Soft-adaptive Threshold Spectral Enhancement

Che Liu, Sibo Cheng, Weiping Ding, Rossella Arcucci

TL;DR

SCDNN (Spectral Cross-Domain Neural Network) introduces a novel way to classify ECG signals by integrating time and frequency domain information directly within the neural network architecture. By using a "Soft-adaptive Threshold Spectral Enhancement" (SATSE) block with trainable thresholds, the model achieves state-of-the-art accuracy and robustness while maintaining exceptionally low inference times.

What Problem Does This Paper Address?

Electrocardiography (ECG) signals are complex, multi-variable time-series data. Traditionally, researchers have approached ECG classification in two ways: manual feature engineering (extracting specific wave patterns like the QRS complex) or deep learning (typically using CNNs or LSTMs).

However, both approaches have significant gaps. Feature engineering is labor-intensive and often fails to generalize across different cardiac diseases. Deep learning models, while powerful, usually treat the time domain and the spectral (frequency) domain as separate entities. They either ignore frequency data entirely or process it outside the main network, preventing the model from learning the intricate relationships between a signal's timing and its frequency components. Furthermore, existing spectral methods often use "hard" thresholds to filter noise, which can lead to permanent loss of critical diagnostic information.

Methods

The authors propose the Spectral Cross-domain Neural Network (SCDNN). The architecture uses a ResNet18 backbone enhanced by a new component: the Soft-adaptive Threshold Spectral Enhancement (SATSE) block.

Fig. 1: Workflow of SCDNN where λL and λH denote the coefficients corresponding to low-frequency and high-frequency information

Key technical innovations include:

  • ConvF Components: Instead of standard convolutional blocks, SCDNN uses "Convolutional Fourier" blocks. These pass the signal through a standard residual block and then branch into the SATSE block.
  • Soft-Adaptive Thresholding: Inside SATSE, the model performs a Fast Fourier Transform (FFT). Instead of using fixed cutoffs to filter frequencies, it employs dual trainable sigmoid functions. This allows the network to learn exactly which frequency ranges (high or low) are important for a specific task through backpropagation.
  • Cross-Domain Communication: A trainable weight matrix ($W$) is applied in the spectral domain to allow different frequency modes to "talk" to each other before the signal is converted back to the time domain via Inverse FFT (IFFT).
  • Residual Integration: The enhanced spectral information is added back to the original time-domain features, ensuring no information is lost during the spectral transformation.
Fig. 3: SATSE Architecture

Datasets

The model was evaluated on two major public 12-lead ECG datasets:

  1. PTB-XL: A massive dataset containing 21,837 signals classified into 5 primary diagnostic categories (Normal, Myocardial Infarction, ST/T Change, Conduction Disturbance, and Hypertrophy).
  2. CPSC2018: Collected from 11 hospitals, featuring 6,877 records with 9 distinct labels, including various types of arrhythmias like Atrial Fibrillation (AF) and Bundle Branch Blocks.

Key Insights and Ideas

  • The Power of "Soft" Filtering: By making the frequency thresholds trainable parameters rather than fixed constants, the model avoids the "information loss" bottleneck common in signal processing. The network effectively learns its own optimal filters.
  • Intra-Network Communication: The primary "aha" moment is that spectral and time-domain features shouldn't just be concatenated at the end; they should interact throughout the deep layers of the network to capture complex cardiac morphologies.
  • Efficiency vs. Performance: The authors found that a shallower backbone (ResNet18) actually outperformed deeper ones (ResNet50). This suggests that for signal processing, the way you handle domain information is more important than simply adding more layers.
  • Generalizability: The model maintained high performance across two very different datasets without needing task-specific tuning, addressing a major hurdle in medical AI.

Key Results

SCDNN consistently outperformed state-of-the-art (SOTA) models like SPN-V2 and CKNA across all metrics:

  • Accuracy: Achieved 0.835 on PTB-XL and 0.859 on CPSC2018, significantly higher than competitors.
  • Robustness: The model showed remarkably low standard deviations (around 0.003), indicating highly stable performance.
  • Speed: Despite the added complexity of FFT/IFFT operations, SCDNN exhibited the second-lowest inference time among all compared models, making it suitable for real-time monitoring in wearable devices.
Performance comparison regarding inference time and accuracy on PTB-XL

Limitations and Future Work

While the results are impressive, the authors note a few areas for expansion:

  • Domain Expansion: The current study focuses strictly on ECG signals. The authors suggest that the SATSE mechanism could be highly effective in other fields like computer vision, natural language processing, and speech recognition, but these have not yet been tested.
  • Physics Systems: There is potential to apply this cross-domain learning to complex physics-based time-series data.
  • Complexity of Initial Values: While the model is robust, the paper notes that the convergence of the threshold ratio ($\phi$) can be sensitive to initial values in certain blocks, though it generally stabilizes during training.

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