Common Interferences Removal from Dense Multichannel EEG Using Independent Component Decomposition

Weifeng Li, Yuxiaotong Shen, Jie Zhang, Xiaolin Huang, Ying Chen, Yun GeComput. Math. Methods Medicine 2018 · DOI

TL;DR

This paper introduces a novel method to remove common interferences from dense multichannel EEG data by leveraging Independent Component Analysis (ICA). By identifying and removing these widespread artifacts, the method aims to improve the accuracy of brain connectivity and network analysis, which are often distorted by such noise.

What Problem Does This Paper Address?

Electroencephalogram (EEG) is a powerful tool for measuring brain activity with high temporal resolution. However, its spatial resolution is limited, leading researchers to increasingly use dense multichannel EEG systems (with 32 or more channels). While this increases data richness, it also introduces significant redundancy and common interferences. These widespread artifacts can obscure subtle neural signals from specific brain regions and, more critically, create spurious correlations between different brain areas. This distortion severely impacts the reliability of brain connectivity and network analyses, which are crucial for understanding brain function and dysfunction. The paper aims to address this challenge by providing a robust method to isolate and remove these common interferences, thereby revealing more accurate representations of underlying neural activity.

Methods

The proposed method builds upon Independent Component Analysis (ICA) and involves three main steps:

  1. Independent Component Decomposition: The initial step uses the FastICA algorithm to decompose the multichannel EEG signals into statistically independent components. This process aims to separate the mixed signals into their underlying source components.
  2. Common Interference Identification: The core innovation lies in how common interferences are identified. Instead of relying on frequency or morphology, the method uses the vector angles between the mixing vectors of each independent component and a unitary vector. Components with mixing vectors that span the minimum angles are considered to represent common interferences, as they suggest a more uniform impact across all channels. An additional restriction is applied: elements in the mapping vector must have the same sign.
  3. Removal and Inverse Transformation: Once identified, the common interference components are removed by setting their corresponding source signals to zero. The demixing matrix (derived from the baseline data) is then applied to the task signals to obtain the processed data, effectively removing the identified interferences. The authors assume that both the common interference and its mixing vectors remain consistent throughout the experiment, allowing for the use of components extracted from baseline data to inform the removal process for task data.

Datasets

The paper utilizes the following datasets:

  • BCI Competition Dataset 1: Used for validating the proposed method on real EEG data, specifically focusing on motor imagery tasks.
  • BCI Competition IV (Dataset 1): This dataset, provided by the Berlin BCI group, was used for applying the method to both artificial and real EEG data. It includes artificial data and recordings from healthy subjects performing motor imagery experiments.

Key Insights and Ideas

  • Angle-based interference identification: Using vector angles to identify common interferences is a novel approach that moves beyond traditional frequency or morphology-based artifact detection. This method effectively captures the "spread" of a component's influence across channels.
  • Baseline-informed interference removal: The assumption that common interferences and their mixing vectors are stable throughout an experiment allows for the use of baseline data to identify and remove artifacts from task-related data. This is particularly useful in stimulus-locking paradigms.
  • Unmasking subtle brain activity: By removing widespread common interferences, the method successfully reveals underlying neural activities that were previously masked, leading to more accurate brain connectivity and network analyses. This is demonstrated by the emergence of prominent mu rhythms during motor imagery tasks.

Key Results

  • Simulation: The proposed method significantly improved the correlation coefficient between the processed signals and the pure signals in simulations, nearly doubling it. This indicates a substantial improvement in signal quality, even in low signal-to-noise ratio scenarios.
  • EEG Data Analysis:
    • Eye Movement Artifacts: The method effectively removed eye movement artifacts, which are common in frontal EEG leads, and alleviated the high correlation between these leads.
  • Relative Power: After processing, the relative power maps showed more distinct distribution characteristics, suggesting the uncovering of intrinsic neural activities.
  • Coherence Heatmaps: The processing revealed more nuanced coherence patterns, with previously obscured coherences between distant lead pairs becoming prominent.
  • Global Coherence: For human subjects, the processing suppressed spurious high coherences in low and high-frequency bands, while a distinct coherence peak in mu rhythms (associated with motor imagery) became prominent. This effect was not observed in artificial signals, as expected.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that it is difficult to provide an accurate signal quality evaluation because the "real" signal is unknown.
  • The method relies on the assumption that common interferences and their mixing vectors remain constant throughout the experiment. This might not hold true in all scenarios.
  • While the method aims to remove common interferences, it does not explicitly filter specific frequencies. However, the results show suppression of spurious high coherences in certain frequency bands.
  • Future Work: The authors suggest that the proposed method has wide applications due to its ability to reveal true correlations between local sources, even with low signal-to-noise ratios, and its applicability in stimulus-locking paradigms. Further exploration of its use in various brain connectivity and network analysis tasks is implied.

Summary updated

BibTeX
@article{li2018common,
  author = {Weifeng Li and Yuxiaotong Shen and Jie Zhang and Xiaolin Huang and Ying Chen and Yun Ge},
  title = {{Common Interferences Removal from Dense Multichannel EEG Using Independent Component Decomposition}},
  journal = {Comput. Math. Methods Medicine},
  year = {2018},
  doi = {10.1155/2018/1482874},
}
Figure