TL;DR
This paper introduces a novel ECG denoising method by combining Extreme-Point Symmetric Mode Decomposition (ESMD) with Nonlocal Means (NLM). This hybrid approach effectively removes noise from ECG signals while preserving crucial diagnostic details like QRS complexes, outperforming existing methods in various performance metrics.
What Problem Does This Paper Address?
Electrocardiogram (ECG) signals are vital for diagnosing heart conditions, but they are often contaminated by various types of noise (e.g., baseline wander, high-frequency noise from electromyography, motion artifacts). This noise can obscure important diagnostic information, leading to inaccurate diagnoses. Existing denoising techniques, such as wavelet shrinkage, Empirical Mode Decomposition (EMD), and standard Nonlocal Means (NLM), struggle to achieve both sufficient noise reduction and preservation of subtle ECG features, especially under high noise levels. This paper aims to bridge this gap by proposing a more robust and effective ECG denoising solution.
Methods
The core of the proposed method is a hybrid approach combining two techniques:
- Extreme-Point Symmetric Mode Decomposition (ESMD): This is a novel signal decomposition technique that breaks down the noisy ECG signal into several intrinsic mode functions (IMFs) and an adaptive global mean (AGM) curve. Unlike traditional EMD, ESMD automatically determines the optimal number of decomposition steps and can handle a certain number of extreme points in the residue. The key idea here is that noise is often concentrated in the initial IMFs, while lower-frequency components and signal details reside in later IMFs and the AGM.
- Nonlocal Means (NLM) Filtering: This image denoising technique is adapted for ECG signals. NLM works by replacing a noisy pixel (or signal point) with a weighted average of similar pixels (or points) in its neighborhood or across the entire image. The weights are determined by the similarity of small image patches (or signal segments) around the points.
The proposed ESMD-NLM method integrates these two:
- The noisy ECG signal is decomposed into IMFs and AGM using ESMD.
- Crucially, QRS complexes are detected within the first few IMFs (where noise and signal components are mixed).
- The first few IMFs (containing significant noise) are then filtered using NLM. The NLM's decay parameter is adaptively adjusted based on the QRS detection results, using smaller parameters for QRS complexes to preserve their sharpness and larger parameters for smoother segments to enhance noise reduction.
- Subsequent IMFs and the AGM, which are assumed to contain fewer noise components and more signal details, are left unprocessed.
- Finally, the denoised IMFs and the unprocessed IMFs/AGM are recombined to form the final denoised ECG signal.
Datasets
The paper evaluates its method using:
- Simulated ECG signals: These were generated using the Open-Source Electrophysiological Toolbox (OSET) and corrupted with Gaussian white noise and muscle artifact noise at various variance levels.
- Real ECG signals: Data was sourced from three publicly available databases:
- MIT-BIH Arrhythmia Database (e.g., 100m.dat, 111m.dat, 221m.dat)
- QT database (e.g., sel100m.dat, sel117m.dat, sel114m.dat)
- MIH-BIH ECG Compression Test Database (e.g., 13420_12m.dat, 13649_04m.dat, 12713_04m.dat)
Key Insights and Ideas
- Decomposition as a Preprocessing Step: The core insight is that decomposing the ECG signal into its constituent IMFs using ESMD allows for targeted denoising. Noise is often concentrated in the higher-frequency IMFs, enabling selective filtering.
- Adaptive NLM for Detail Preservation: Standard NLM can smooth out important signal features. By adaptively adjusting the NLM decay parameter based on QRS complex detection, the method achieves both noise suppression in smooth regions and sharp preservation of critical waveform components.
- Leveraging QRS Complexes: The QRS complex is identified as a dominant feature. Its detection is not just for analysis but also serves as a crucial guide for adapting the denoising parameters, ensuring these vital parts of the ECG are not degraded.
- Selective Denoising: Not all IMFs need to be denoised. Leaving later IMFs and the AGM untouched preserves signal integrity and reduces computational load.
Key Results
The proposed ESMD-NLM method demonstrated superior performance compared to several state-of-the-art methods (NLM, VMD, EEMD, MED filter) across various metrics:
- Simulated ECG Signals:
- Consistently achieved higher Signal-to-Noise Ratio (SNR).
- Produced lower Root Mean Squared Error (RMSE) and Percent Root Mean Square Difference (PRD).
- Obtained significantly lower Mean Opinion Score (MOS) error values, indicating better perceived signal quality. For instance, MOSerror values for ESMD-NLM were often below 15, while others reached 30 or higher.
- Qualitative visual comparisons showed better preservation of smooth segments and ECG amplitudes.
- Real ECG Signals:
- Visual comparisons indicated that ESMD-NLM preserved the amplitude values of clinical ECG signals more closely than other methods.
- It effectively suppressed noise in smooth segments while maintaining the shape of QRS complexes, P waves, and T waves.
- Blind MOS tests conducted by cardiologists confirmed that ESMD-NLM provided the highest visual quality for denoised signals, suggesting it could facilitate more accurate diagnoses.
Limitations and Future Work
- Computational Efficiency: While effective, the ESMD-NLM method is computationally more intensive than standard NLM and VMD, primarily due to the ESMD decomposition step. The authors suggest that parallel computing strategies could improve efficiency.
- Future Work: The authors propose extending the ESMD-NLM method to denoise other physiological signals, such as electrooculogram (EOG), electroencephalogram (EEG), and electromyogram (EMG).
Summary updated