Intelligent Extraction of Salient Feature From Electroencephalogram Using Redundant Discrete Wavelet Transform

Xianyu Wang, Cong Li, Rui Zhang, Liang Wang, Jinlin Tan, Hai WangFrontiers in Neuroscience 2022 · DOI

TL;DR

This paper highlights a critical issue called "translation variability" in the standard Discrete Wavelet Transform (DWT) when analyzing EEG signals. The authors demonstrate that a modified approach, the Redundant Discrete Wavelet Transform (RDWT), offers superior stability and accuracy by eliminating this variability, making it a more reliable tool for brain-computer interface (BCI) research and clinical applications.

What Problem Does This Paper Address?

Electroencephalogram (EEG) signals are crucial for understanding brain activity, diagnosing diseases, and developing brain-computer interfaces (BCIs). However, raw EEG data is often noisy and contains interference from other physiological processes. To extract meaningful information, signal processing techniques are essential. Wavelet transforms (WT) are powerful tools for analyzing the time-frequency characteristics of dynamic signals like EEG.

The Discrete Wavelet Transform (DWT) is widely used for this purpose. However, the authors identify a significant problem with DWT: translation variability (TV). This issue arises from the downsampling operations inherent in the DWT's decomposition process. TV means that even a slight shift in the input signal can lead to significant distortions in the resulting wavelet subspaces, compromising the accuracy of feature extraction. This problem, though impactful, has been largely overlooked. The paper aims to demonstrate the detrimental effects of TV in DWT for EEG analysis and propose Redundant Discrete Wavelet Transform (RDWT) as a more robust alternative.

Methods

The core of the paper's methodology involves comparing the behavior of DWT and RDWT under signal translations.

  • Wavelet Transform Basics: The paper starts by defining the wavelet transform and its discrete counterpart (DWT), explaining how it decomposes a signal into different frequency bands (subspaces) using scaling and wavelet functions.
  • The Problem of Downsampling: It details how the DWT's filter bank structure, which includes downsampling and upsampling operations, leads to a reduction in the number of coefficients compared to the original signal. This reduction is the root cause of translation variability.
  • Introducing RDWT: The Redundant Discrete Wavelet Transform (RDWT) is presented as a solution. RDWT achieves translation invariance by simply removing the downsampling and upsampling operations from the DWT's filter bank. This results in an overcomplete representation, where the number of coefficients is greater than or equal to the original signal's length.
  • Testing with Delta Impulse Functions: To isolate and demonstrate the effect of translation variability, the researchers used a discrete delta impulse function. By shifting this function and applying both DWT and RDWT, they could directly observe how each transform responded to input signal translations.
  • Analysis of EEG Signals: The theoretical findings were then validated using actual EEG data from clinical patients. The same translation tests were performed on these real-world signals to show the practical implications of TV.
  • Comparison Metrics: The comparison focused on the time-domain waveforms and the frequency spectra (using FFT) of the resulting wavelet subspaces.

Datasets

  • Simulated Data: Discrete delta impulse functions were synthesized to create controlled test signals.
  • Real EEG Data: The paper utilized EEG data obtained from the "EEG dataset delivered publicly on the Internet by Neurology and Sleep Centre, Hauz Khas, New Delhi." These signals were sampled at 200 Hz and were from ictal stages (periods of seizure activity).

Key Insights and Ideas

  • TV is a Real Problem for DWT: The paper convincingly shows that DWT is highly sensitive to shifts in the input signal, leading to distorted representations in its wavelet subspaces. This is not just a theoretical curiosity but has practical implications for feature extraction.
  • RDWT is Translation Invariant: In stark contrast to DWT, RDWT consistently produces identical decomposition results regardless of how the input signal is shifted. This "translation invariance" is a significant advantage.
  • DWT as a Linear Time-Variant (LTV) System: The authors frame DWT as an LTV system, meaning its output depends not only on the input but also on the time at which the input occurs. This explains the observed distortions.
  • RDWT as a Linear Time-Invariant (LTI) System: RDWT, on the other hand, behaves as an LTI system, where the output is simply a time-shifted version of the output for a time-shifted input. This predictable behavior is crucial for reliable analysis.
  • Implications for BCI and Clinical Use: The findings strongly suggest that RDWT is a more suitable tool for applications like BCIs and clinical EEG analysis where stable and accurate feature extraction is paramount.

Key Results

  • Delta Impulse Tests: Numerical simulations using shifted delta impulse functions clearly demonstrated that DWT's wavelet subspace outputs changed significantly with input shifts, while RDWT's outputs remained identical.
  • EEG Signal Analysis: When applied to actual EEG data, RDWT produced congruent wavelet subspace results for shifted signals, whereas DWT showed noticeable differences in waveform shape and spectral content across its subspaces.
  • Frequency Band Specificity: The impact of TV was more pronounced in higher frequency subspaces of the DWT, leading to more significant distortions. Lower frequency bands showed less pronounced, but still present, differences.
  • RDWT's Stability: The paper concludes that RDWT offers a more stable performance, avoiding the additional noise and distortions introduced by TV in DWT.

Limitations and Future Work

  • Computational Cost of RDWT: While RDWT offers translation invariance, the paper acknowledges that it significantly increases computational complexity due to the lack of downsampling.
  • Alternative Solutions: The authors briefly mention other approaches to enhance time invariance, such as the dual-tree complex wavelet transform and frequency-domain wavelet transforms, suggesting these as avenues for future research.
  • Generalizability of DB8 Basis: The study used the Daubechies-8 (DB8) wavelet basis as an example. While the authors state that the TV issue is inherent to DWT regardless of the basis, further exploration with other bases could strengthen the findings.
  • Future Directions: The paper implicitly suggests that further research could focus on developing computationally efficient methods that retain the translation invariance of RDWT or exploring the application of RDWT to a wider range of EEG analysis tasks.

Summary updated

Abstract

At present, electroencephalogram (EEG) signals play an irreplaceable role in the diagnosis and treatment of human diseases and medical research. EEG signals need to be processed in order to reduce the adverse effects of irrelevant physiological process interference and measurement noise. Wavelet transform (WT) can provide a time-frequency representation of a dynamic process, and it has been widely utilized in salient feature analysis of EEG. In this paper, we investigate the problem of translation variability (TV) in discrete wavelet transform (DWT), which causes degradation of time-frequency localization. It will be verified through numerical simulations that TV is caused by downsampling operations in decomposition process of DWT. The presence of TV may cause severe distortions of features in wavelet subspaces. However, this phenomenon has not attracted much attention in the scientific community. Redundant discrete wavelet transform (RDWT) is derived by eliminating the downsampling operation. RDWT enjoys the attractive merit of translation invariance. RDWT shares the same time-frequency pattern with that of DWT. The discrete delta impulse function is used to test the time-frequency response of DWT and RDWT in wavelet subspaces. The results show that DWT is very sensitive to the translation of delta impulse function, while RDWT keeps the decomposition results unchanged. This conclusion has also been verified again in decomposition of actual EEG signals. In conclusion, to avoid possible distortions of features caused by translation sensitivity in DWT, we recommend the use of RDWT with more stable performance in BCI research and clinical applications.
BibTeX
@inproceedings{wang2022intelligent,
  author = {Xianyu Wang and Cong Li and Rui Zhang and Liang Wang and Jinlin Tan and Hai Wang},
  title = {{Intelligent Extraction of Salient Feature From Electroencephalogram Using Redundant Discrete Wavelet Transform}},
  booktitle = {Frontiers in Neuroscience},
  year = {2022},
  doi = {10.3389/fnins.2022.921642},
}
Figure