Abstract
Electrocardiogram (ECG) signals are frequently contaminated by various noise sources in clinical environments, complicating diagnosis. This paper proposes a Fully Convolutional Denoising Autoencoder (FCN-DAE) to effectively remove noise while preserving critical clinical features. By eliminating fully connected layers and pooling operations, the model achieves superior signal reconstruction and a 32x compression ratio. Experimental results on standard datasets demonstrate that the FCN-DAE outperforms traditional Deep Neural Networks (DNNs) and standard Convolutional Neural Networks (CNNs) in terms of Signal-to-Noise Ratio (SNR) improvement and visual fidelity.
Introduction
The electrocardiogram (ECG) is a vital diagnostic tool for identifying cardiac arrhythmias. However, real-world ECG signals are often degraded by noise, including Baseline Wander (BW), Muscle Artifacts (MA), and Electrode Motion (EM). Traditional filtering techniques often struggle to differentiate between high-frequency noise and the sharp R-peaks of the ECG signal, potentially leading to diagnostic errors. There is a critical need for robust denoising methods that can maintain the morphological integrity of the signal, particularly the QRS complex.
Methods: The FCN-DAE Architecture
The researchers developed a 13-layer Fully Convolutional Denoising Autoencoder (FCN-DAE). Unlike standard CNNs, this architecture avoids fully connected (dense) layers, which often lead to signal "averaging" and loss of peak amplitude.
Key design choices include:
- Strided Convolutions: Instead of max-pooling, which discards spatial information, the model uses strided convolutions to downsample the signal. This ensures better alignment between the input and reconstructed output.
- Bottleneck Compression: The encoder compresses a 1024-sample input into a 32-sample latent representation ($z$). This forces the network to capture only the most salient features of the clean ECG signal.
- Activation and Normalization: Exponential Linear Units (ELU) and Batch Normalization are employed to stabilize training and improve convergence speed.
The model was trained using clean signals from the MIT-BIH Arrhythmia Database and noise profiles from the MIT-BIH Noise Stress Test Database (NSTDB).
Key Results and Performance
The FCN-DAE was evaluated against standard DNN and CNN architectures. Quantitative analysis focused on Signal-to-Noise Ratio improvement ($SNR_{imp}$) and Percentage-root-mean-square Difference (PRD).
As shown in the performance metrics, the FCN-DAE consistently achieved the lowest PRD across various input SNR levels. At a highly noisy input of -1 dB, the FCN improved the signal by 15.49 dB, compared to 11.75 dB for the DNN and 14.14 dB for the CNN. Qualitatively, the FCN-DAE was uniquely capable of preserving the sharp R-peaks and smooth baselines, whereas other models often introduced artifacts or attenuated the signal peaks.
Discussion and Conclusion
The success of the FCN-DAE stems from its focus on local connectivity. By discarding dense layers, the model maintains the translation invariance necessary for processing temporal signals like the ECG. Furthermore, the ability to reconstruct clean signals from a 32x compressed bottleneck suggests significant potential for efficient data transmission in wearable medical devices.
Future work will focus on real-time implementation on low-power hardware and clinical validation to ensure that the denoising process does not introduce artifacts that could be misinterpreted as pathological features.
Summary updated