EEG signals often contain unwanted artifacts that can obscure brain activity. Common culprits include eye movements, muscle activity, and electrical interference. These artifacts can significantly impact EEG analysis and interpretation, making their removal crucial for accurate results.

Various techniques exist to combat EEG artifacts. Linear filtering can remove specific frequency components, while more advanced methods like Independent Component Analysis (ICA) separate mixed signals. Evaluating artifact removal effectiveness involves visual inspection and quantitative metrics to ensure signal quality improvement.

Types of Artifacts and Linear Filtering

Common EEG artifact types

• Eye blinks and movements generate large, transient spikes in frontal electrodes (Fp1, Fp2, F7, F8) due to the electrical potential difference between the cornea and retina
• Muscle activity (EMG) introduces high-frequency components from sources like jaw clenching, neck movements, and facial expressions
• Electrical interference from 50/60 Hz power line noise and nearby devices introduces constant frequency components
• Electrode pop or movement causes abrupt, high-amplitude spikes due to sudden changes in electrode impedance
• Cardiac activity (ECG) can introduce rhythmic, low-frequency components, especially in channels near blood vessels

Linear filtering for artifact removal

• High-pass filters attenuate slow drift, baseline wander, and low-frequency artifacts by removing components below a cutoff frequency (typically 0.1-1 Hz for EEG)
• Low-pass filters attenuate high-frequency noise and muscle artifacts by removing components above a cutoff frequency (typically 30-70 Hz for EEG)
• Notch filters attenuate power line noise (50/60 Hz) and its harmonics using a narrow band-stop filter centered around the target frequency
• Filter design considerations include filter type (Butterworth, Chebyshev, elliptic), order, cutoff frequencies, stopband attenuation, and phase response (linear phase preferred)
• Linear filtering limitations: cannot remove artifacts overlapping with frequencies of interest, may introduce distortions if designed improperly, and cannot separate artifacts from EEG signals with similar frequency characteristics

Advanced Artifact Removal Techniques

ICA in EEG signal separation

• ICA separates multi-channel EEG into independent subcomponents (ICs) by finding a linear transformation (unmixing matrix) that maximizes their statistical independence, assuming a linear mixture of independent non-Gaussian sources
• For artifact removal, identify artifact-related ICs based on spatial distribution, temporal characteristics, and frequency content:
  1. Eye blinks: frontal distribution, large amplitude, low frequency
  2. Eye movements: frontal distribution, gradual changes, low frequency
  3. Muscle activity: localized distribution, high frequency, bursts of activity
  4. Cardiac activity: localized distribution near blood vessels, rhythmic
• Remove artifact-related ICs by zeroing their coefficients in the mixing matrix and reconstruct the artifact-free EEG by remixing remaining ICs
• ICA limitations: requires sufficient EEG channels (64+), assumes artifact and EEG statistical independence and non-Gaussianity, and manual IC identification can be subjective and time-consuming (automatic classification methods can help)

Effectiveness of artifact removal methods

• Visual inspection compares raw and processed EEG side-by-side for artifact reduction and preservation of relevant features
• Quantitative metrics:
  1. Signal-to-noise ratio (SNR): $SNR = 10 \log_{10} \frac{P_{signal}}{P_{noise}}$, higher SNR indicates better artifact removal
  2. Root mean square (RMS) amplitude: $RMS = \sqrt{\frac{1}{N} \sum_{i=1}^{N} x_i^2}$, decrease suggests artifact power reduction
  3. Spectral analysis compares power spectral density (PSD) for artifact-related frequency reduction and relevant EEG band preservation
• Cross-validation compares results, consistency, and reliability across methods (linear filtering, ICA, regression) to identify strengths and limitations for specific artifacts and applications
• Evaluate impact on downstream analysis steps (event-related potentials, time-frequency, connectivity) for improved quality and interpretability without introducing new biases or artifacts