SVR-EEMD: An Improved EEMD Method Based on Support Vector Regression Extension in PPG Signal Denoising

Guangda Liu, Xinlei Hu, Enhui Wang, Ge Zhou, Jing Cai, Shang ZhangComput. Math. Methods Medicine 2019 · DOI

TL;DR

This paper introduces SVR-EEMD, a novel signal denoising method for photoplethysmography (PPG) signals. By extending the signal using Support Vector Regression (SVR) before applying Ensemble Empirical Mode Decomposition (EEMD), SVR-EEMD effectively mitigates the "end effect" problem inherent in EEMD, leading to improved signal reconstruction and more accurate analysis of PPG data.

What Problem Does This Paper Address?

Photoplethysmography (PPG) is a widely used non-invasive technique for measuring blood volume and flow. However, PPG signals are susceptible to various noise sources, including motion artifacts and respiratory activities, which introduce nonlinear and nonstationary characteristics. These characteristics make traditional signal processing methods struggle with accurate denoising. While Ensemble Empirical Mode Decomposition (EEMD) has shown promise in handling such signals by decomposing them into intrinsic mode functions (IMFs), it suffers from an "end effect." This artifact causes signal divergence at the boundaries of the time series, leading to distortions and inaccuracies in the reconstructed signal. This paper aims to address this "end effect" to enhance the denoising performance of PPG signals.

Methods

The core contribution of this paper is the SVR-EEMD method, which combines Support Vector Regression (SVR) with Ensemble Empirical Mode Decomposition (EEMD). The process involves two main steps:

  1. Signal Extension using SVR: Before applying EEMD, the original PPG signal is extended at both its left and right ends. This is achieved by training an SVR model on the existing signal data. SVR, a robust regression technique, is used to predict and extrapolate the signal's trend beyond its original boundaries. This extension aims to provide a more complete signal for the subsequent decomposition.
  2. EEMD Decomposition and Reconstruction: The extended signal is then subjected to the EEMD algorithm. EEMD is an ensemble-based method that adds white noise to the signal multiple times and decomposes each noisy version into IMFs. By averaging the IMFs across these ensembles, EEMD effectively addresses the "mode mixing" problem of the original Empirical Mode Decomposition (EMD). After decomposition, appropriate IMFs are selected for reconstruction, effectively filtering out noise while preserving the essential signal components. The SVR extension is truncated during reconstruction, leaving only the denoised signal.

The paper also details the specific implementation of SVR, including the use of an epsilon-insensitive loss function and a linear kernel for efficiency. The EEMD process is described as a standard sifting process involving the identification of local extrema, envelope generation, and mean subtraction.

Datasets

The study utilizes two types of datasets:

  • Simulated Data: A synthetic signal was generated consisting of a 5 Hz sinusoidal wave and a 20 Hz cosine wave, with added Gaussian white noise to achieve a Signal-to-Noise Ratio (SNR) of 15 dB. This dataset was used to evaluate the method's performance in a controlled environment.
  • Real-World PPG Data: The authors used data from the BIDMC PPG and Respiration Dataset of PhysioBank. This dataset contains physiological recordings from critically ill patients, including PPG signals. Ten sets of PPG data from this publicly available dataset were selected for the experiments.

Key Insights and Ideas

  • SVR as an "End Effect" Solution: The paper demonstrates that SVR can effectively predict and extend the signal's trend at its boundaries, significantly mitigating the "end effect" that plagues traditional EEMD. This is a crucial advancement for time-series decomposition methods.
  • Improved Decomposition Performance: By addressing the "end effect," SVR-EEMD leads to cleaner and more accurate IMF components, which are better ordered by frequency and show less divergence at the signal's edges.
  • Enhanced Denoising for PPG: The SVR-EEMD method shows superior performance in denoising PPG signals compared to standard FIR filtering and EEMD, particularly in preserving signal integrity and reducing artifacts.
  • Robustness to Noise: The study investigates the impact of different noise intensities on the SVR-EEMD method, showing that it maintains good performance across a range of noise levels.

Key Results

  • Simulated Data: The SVR-EEMD method achieved a significantly higher SNR (44.54 dB) and correlation coefficient (0.9978) compared to FIR filtering (37.83 dB SNR, 0.9886 Corr) and EEMD (18.35 dB SNR, 0.9203 Corr).
  • Real-World PPG Data:
    • The SVR-EEMD method demonstrated higher precision and recall rates for pulse wave peak detection compared to FIR and EEMD.
  • It achieved a higher correlation coefficient and lower mean delay time (MDT) in reconstructing PPG signals, indicating better signal fidelity and reduced phase shift.
  • Qualitatively, SVR-EEMD successfully reconstructed PPG signals, respiratory activity, and motion artifacts, with the reconstructed PPG signal showing good consistency with the original, even when submerged in noise.

Limitations and Future Work

  • Noise Intensity Selection: While the paper shows the impact of noise intensity on performance, the optimal range for noise intensity (Np) for real-world PPG data (0.75-1.25) is presented as a point for further study. The paper suggests that determining how much noise should be applied needs further investigation.
  • Computational Complexity: Although not explicitly detailed as a limitation, SVR-EEMD involves an additional SVR extension step, which might increase computational time compared to standard EEMD or FIR filtering.
  • Future Work: The authors suggest that further research is needed to explore the optimal noise intensity for application. They also imply that the method's effectiveness in other nonlinear and nonstationary signal analysis applications could be investigated.

Summary updated

BibTeX
@article{liu2019svr,
  author = {Guangda Liu and Xinlei Hu and Enhui Wang and Ge Zhou and Jing Cai and Shang Zhang},
  title = {{SVR-EEMD: An Improved EEMD Method Based on Support Vector Regression Extension in PPG Signal Denoising}},
  journal = {Comput. Math. Methods Medicine},
  year = {2019},
  doi = {10.1155/2019/5363712},
}
Figure