Existing Techniques for Removing EOG Artifacts from EEG Signals
Current methodologies for eliminating electrooculogram (EOG) artifacts from electroencephalogram (EEG) signals encompass various advanced techniques. One such method is Empirical Mode Decomposition (EMD), which is data-driven and breaks down signals into intrinsic mode functions (IMFs) based on their natural oscillatory characteristics. Although EMD can differentiate EOG artifacts from EEG signals, it may encounter issues with mode mixing, potentially leading to the omission of significant signal components. Another approach is Singular Spectrum Analysis (SSA), a subspace technique that uses singular value decomposition to break down signals into distinct components. While SSA is adept at isolating low-frequency noise, including EOG artifacts, it necessitates precise threshold adjustments for optimal results. Circulant SSA is an enhanced version of SSA that employs circulant matrices, facilitating more efficient separation of oscillatory components from the overall signal. The Discrete Wavelet Transform (DWT) is another notable method that dissects EEG signals into various frequency components using wavelets, enabling targeted identification and removal of EOG artifacts within specific frequency bands. Although DWT is effective in addressing low-frequency components linked to artifacts, it may struggle with accurately reconstructing the original signal post-artifact removal. Independent Component Analysis (ICA) is a statistical method that decomposes mixed signals into independent components, widely utilized in multi-channel EEG to isolate and eliminate artifacts, including EOGs. However, ICA’s effectiveness in single-channel EEG is limited, as it requires multiple channels for optimal separation. Variational Mode Decomposition (VMD) decomposes signals into a set of modes with distinct bandwidth properties, effectively isolating and eliminating artifacts while necessitating careful adjustment of decomposition parameters. A newer method, the Fixed Frequency Empirical Wavelet Transform (FF-EWT), specifically targets fixed frequency ranges linked to EOG artifacts, offering precise and focused removal while maintaining the integrity of non-artifact content in EEG signals. By integrating FF-EWT with other filtering techniques, such as the Generalized Matrix Total Variation (GMETV) filter, the overall efficacy of artifact removal is enhanced, resulting in improved signal quality compared to more traditional methods.
Proposed Methods
The proposed approach for eliminating eyeblink artifacts is illustrated in a detailed flowchart. The procedure unfolds in three primary stages: initially, FF-EWT decomposes a contaminated EEG signal into six IMFs. Following this, EOG-related IMFs are identified based on a specific feature threshold value. Finally, a four-stage cascaded Savitzky-Golay (SG) filter is employed to mitigate the eyeblink events, leveraging EOG data as its foundational element. The upcoming sections will delve deeper into the intricacies of each stage involved in this process.
EWT
The Empirical Wavelet Transform (EWT) serves as a robust technique for analyzing nonstationary signals, first introduced in previous research. This method generates adaptive wavelet-based filters that are dynamically aligned with the characteristics of the specific signal being analyzed. These filters are versatile, tailored to capture relevant information embedded within the signal’s spectrum. Upon applying EWT, the resulting sub-band signals (SBSs) are focused on particular frequencies, showcasing compact frequency support. The EWT operates through several steps: first, the Fast Fourier Transform (FFT) is employed to obtain the signal’s spectral data within the range of \(0 – \pi\) radians. Next, limit identification processes segment the Fourier spectrum accordingly. Finally, empirical wavelets are defined as bandpass filters on each segment, utilizing the principles of Littlewood-Paley and Meyer wavelets. The wavelet function and scaling functions are mathematically represented to ensure tight frame properties, allowing for effective signal approximation and detail coefficient extraction.
Feature Extraction for EOG-Related IMFs Selection
Once the sub-band signals (SBSs) are acquired, the objective of this phase is to automatically identify EOG-related SBSs for subsequent analysis. The current study employs kurtosis, power spectral density (PSD), and permutation entropy (DisEn) features to characterize components associated with EOG artifacts. Kurtosis serves as a statistical measure that reflects the peak’s sharpness and assesses the steepness of a distribution, indicating that SBSs linked to EOG typically exhibit elevated kurtosis values. The kurtosis statistic is computed using a defined formula involving central moments of the distribution. DisEn, on the other hand, is a nonlinear dynamical metric widely used for analyzing the dynamics of time series, where greater instability correlates with higher values of DisEn. The calculation of DisEn involves assessing the probability of various ordinal patterns within the sequence. PSD quantifies the signal’s power across different frequencies, typically calculated from the squared magnitude of the Fourier transform. Once specific feature thresholds are established, any IMFs exceeding these thresholds are classified as EOG-related through a defined algorithm. The identified EOG components are then aggregated and processed through the cascaded SG filter for further refinement.
Smoothing of Signal Using GMETV Filter
The Generalized Matrix Total Variation (GMETV) approach, unlike conventional methods, incorporates a matrix attribute, enhancing its capability in processing EOG artifacts. The VME technique is employed to extract the EOG artifact segment, which is subsequently evaluated using the GMETV denoising cost function. This function is structured to minimize the difference between the EOG segment and the filtered output, while also incorporating a regularization term to manage the overall complexity of the signal. The equation governing this cost function illustrates the interplay between the EOG segment and the filtered signal, adjusting dynamically based on the matrix parameter. To optimize the cost function, specific conditions must be met, ensuring that the output signal is effectively smoothed while accurately detecting EOG components amid the noisy EEG data.
