Sparse ECG Denoising with Generalized Minimax Concave Penalty

Zhongyi Jin, Anming Dong, Minglei Shu, Yinglong WangSensors 2019 · DOI

TL;DR

This paper introduces a novel ECG denoising framework that combines low-pass filtering with sparse signal recovery. They propose using the Generalized Minimax Concave (GMC) penalty for sparsity recovery, which effectively addresses the underestimation of important signal components (like QRS peaks) that plagues traditional $\ell_1$-norm methods, leading to improved denoising performance.

What Problem Does This Paper Address?

Electrocardiogram (ECG) signals are crucial for diagnosing cardiovascular diseases, but they are highly susceptible to noise. This noise can obscure important waveform features, leading to misdiagnoses. While various denoising techniques exist (linear filters, EMD, wavelets), they often struggle with spectral overlap between signal and noise, or can distort critical signal details. This paper focuses on addressing the limitations of existing sparse recovery methods, particularly their tendency to underestimate the amplitude of key ECG components, which can be detrimental for accurate clinical interpretation.

Methods

The core of the proposed method is a two-stage denoising framework:

  1. Low-Pass Filtering (LPF): The noisy ECG signal is first passed through a zero-phase LPF. This stage smooths out the low-frequency components of the signal.
  2. Sparsity Recovery: The residual signal (original noisy signal minus the LPF output) is then processed to recover the sparse, high-frequency components. This is where the novelty lies. The authors propose two approaches for this stage:
  3. $\ell_1$-norm Penalty: This is a standard approach in sparse recovery, aiming to find a sparse representation of the signal. However, it suffers from underestimation of signal amplitudes.
  4. Generalized Minimax Concave (GMC) Penalty: This is the proposed novel approach. The GMC penalty is a non-convex penalty function that, when carefully formulated, can maintain the convexity of the overall cost function. Its key advantage is that it penalizes large values less severely than the $\ell_1$-norm, thereby mitigating the underestimation problem and preserving signal details, especially the sharp peaks characteristic of ECGs.

The GMC-based sparsity recovery problem is formulated as a saddle-point optimization problem and solved using the Forward-Backward Splitting (FBS) method.

Datasets

The proposed methods were evaluated using ECG signals from the MIT-BIH Arrhythmia database. This is a widely recognized benchmark dataset for cardiac signal analysis. The authors used records from this database to test their algorithms and compare them against existing methods. No new benchmark datasets were introduced.

Key Insights and Ideas

  • GMC Penalty for ECG: The paper successfully extends the application of the GMC penalty, previously used in areas like fault detection, to ECG denoising. This demonstrates its versatility for signals with sharp, high-amplitude transients.
  • Overcoming $\ell_1$-Norm Underestimation: The core insight is that the GMC penalty's unique penalization characteristic (less severe punishment for larger values) directly addresses a critical flaw in $\ell_1$-norm-based methods, which tend to shrink important signal peaks.
  • LPF + Sparse Recovery Synergy: The framework's effectiveness stems from the complementary nature of LPF and sparse recovery. LPF handles the smooth components, while sparse recovery isolates and reconstructs the detailed, high-frequency sparse components.
  • Convexity Maintenance: Despite the non-convex nature of the GMC penalty itself, the authors show how to formulate the problem such that the overall objective function remains convex, ensuring a globally optimal solution can be found.

Key Results

  • Improved Peak Preservation: The GMC-based method significantly outperforms the $\ell_1$-norm method and other classical approaches (like SASS) in preserving the amplitude of QRS peaks, a critical feature in ECG analysis.
  • Quantitative Performance Gains: The GMC method showed significant improvements in Signal-to-Noise Ratio Improvement (SNRimp), Root Mean Square Error (RMSE), and Percent Root Mean Square Difference (PRD) across various Signal-to-Noise Ratios (SNRs) compared to referenced algorithms.
  • Time-Frequency Analysis: Spectrogram analysis confirmed that the GMC method preserves more high-frequency components of the ECG signal compared to the $\ell_1$-norm method.
  • Robustness to Noise: The algorithm effectively removed added Gaussian noise, even at low SNR levels (-10 dB).
  • Performance on Arrhythmia: The GMC-based algorithm also demonstrated effectiveness in denoising abnormal ECG signals, suggesting its broad applicability.

Limitations and Future Work

  • Parameter Tuning: The GMC-based algorithm requires careful selection of parameters ($\lambda$ and $\gamma$). While the paper investigates the impact of $\lambda$, further research into optimal parameter selection strategies for different ECG signal types could be beneficial.
  • Computational Complexity: The paper notes that the proposed algorithms converge slower than some existing methods like SASS. Future work could focus on optimizing the computational efficiency of the GMC-based approach.
  • Real-time Application: While the framework is promising, further investigation into its suitability for real-time ECG monitoring systems, which often have strict latency requirements, would be valuable.

Summary updated

Abstract

The electrocardiogram (ECG) is an important diagnostic tool for cardiovascular diseases. However, ECG signals are susceptible to noise, which may degenerate waveform and cause misdiagnosis. In this paper, the ECG noise reduction techniques based on sparse recovery are investigated. A novel sparse ECG denoising framework combining low-pass filtering and sparsity recovery is proposed. Two sparsity recovery algorithms are developed based on the traditional ℓ 1 -norm penalty and the novel generalized minimax concave (GMC) penalty, respectively. Compared with the ℓ 1 -norm penalty, the non-differentiable non-convex GMC penalty has the potential to strongly promote sparsity while maintaining the convexity of the cost function. Moreover, the GMC punishes large values less severely than ℓ 1 -norm, which is utilized to overcome the drawback of underestimating the high-amplitude components for the ℓ 1 -norm penalty. The proposed methods are evaluated on ECG signals from the MIT-BIH Arrhythmia database. The results show that underestimating problem is overcome by the proposed GMC-based method. The GMC-based method shows significant improvement with respect to the average of output signal-to-noise ratio improvement ( S N R i m p ), the average of root mean square error (RMSE) and the percent root mean square difference (PRD) over almost any given SNR compared with the classical methods, thus providing promising approaches for ECG denoising.
BibTeX
@inproceedings{jin2019sparse,
  author = {Zhongyi Jin and Anming Dong and Minglei Shu and Yinglong Wang},
  title = {{Sparse ECG Denoising with Generalized Minimax Concave Penalty}},
  booktitle = {Sensors},
  year = {2019},
  doi = {10.3390/s19071718},
}
Figure