Robust electroencephalogram phase estimation with applications in brain-computer interface systems

Esmaeil Seraj, Reza SameniPhysiological 2017 · DOI

TL;DR

This paper introduces a novel, robust method for extracting phase information from electroencephalogram (EEG) signals, which is crucial for understanding brain activity and developing brain-computer interfaces (BCIs). By addressing the unreliability of traditional phase estimation methods, especially in noisy or low-amplitude signal segments, their approach significantly improves the accuracy and robustness of BCI applications.

What Problem Does This Paper Address?

Understanding brain activity often relies on analyzing the synchronized firing of neuronal populations, which is reflected in EEG signals. The phase of these signals is believed to carry important information about neural communication. However, traditional methods for extracting this phase information are prone to errors, particularly during periods of low EEG amplitude or in the presence of noise. These errors can manifest as "fake" phase jumps that don't correspond to actual brain events, leading to unreliable interpretations and poor performance in applications like BCIs. This paper aims to tackle this challenge by developing a more robust and accurate method for EEG phase estimation.

Methods

The core of the proposed method involves a "Monte Carlo" approach to phase estimation. Instead of relying on a single, fixed filter for signal processing, the authors introduce a novel technique that involves:

  • Zero-Pole Perturbation: They take a standard bandpass filter prototype and introduce tiny, random perturbations to its zero and pole locations. This creates an ensemble of slightly different filters.
  • Forward-Backward Filtering: To ensure zero-phase distortion (crucial for accurate phase analysis), they apply the filters in a forward-backward manner. This is done offline, so it doesn't introduce latency.
  • Ensemble Averaging: The EEG signal is processed with each of the perturbed filters, and the resulting phase estimates are averaged. This ensemble averaging helps to smooth out the unreliable phase estimates that occur in low-amplitude or noisy segments, leading to a more stable and robust final phase estimate.

This approach is designed to be less sensitive to minor variations in filter parameters and to mitigate the impact of noise and low-amplitude signal segments.

Datasets

The study utilizes a publicly available dataset from the Physionet database, specifically the "Neuroelectric and Myoelectric Databases." This dataset includes recordings from 109 volunteers performing various motor and motor-imagery tasks, recorded using the BCI2000 System. For their BCI case study, they focused on specific electrode channels (FCz, C3, and C4) and classified three types of activities: rest, left/up side motion (real or imagined), and right/down side motion (real or imagined).

Key Insights and Ideas

  • Fake Phase Jumps are a Major Issue: The paper strongly emphasizes that many observed "phase jumps" in EEG are not genuine brain events but artifacts of the estimation process, especially in low-amplitude signal segments.
  • Robustness is Key for BCI: Traditional phase estimation methods are too sensitive to noise and signal variations, making them unreliable for practical BCI applications.
  • Ensemble Averaging Smooths Out Artifacts: By averaging phase estimates from multiple slightly perturbed filters, the method effectively cancels out spurious phase fluctuations.
  • Zero-Phase Filtering is Essential: For accurate phase analysis, filters must not introduce phase distortions. The forward-backward filtering approach ensures this.

Key Results

The proposed robust phase estimation method demonstrated significant improvements in BCI classification accuracy:

  • Random Forest Classifier: Achieved a 7% improvement in classification rates without additive noise and a 12% improvement with additive noise.
  • K-Nearest Neighbors Classifier: Showed a 4% improvement without additive noise and an 8% improvement with additive noise.
  • Statistical Significance: These improvements were statistically confirmed with p-values of 0.01 and 0.03, indicating a high likelihood that the observed gains are real.
  • Robustness to Noise and Filter Variations: The method proved to be significantly less susceptible to additive noise and variations in filter parameters compared to conventional methods.

Limitations and Future Work

  • Inherent Limitations of LAAS: The authors acknowledge that phase ambiguities due to low-amplitude analytical signal (LAAS) epochs are an intrinsic limitation of phase estimation, even with their robust method.
  • Future Applications: The proposed method is not limited to BCI and can be applied to other areas of EEG analysis, such as brain connectivity and cognitive studies.
  • Combining Phase and Envelope: Future work could explore combining phase information with analytical signal envelopes for even better performance in BCI and other applications.
  • Sleep Stage Classification and Pathological Cases: The method is also suggested for use in sleep stage classification and the analysis of pathological EEG signals (e.g., epileptic EEG).

Summary updated

Abstract

The average performance was improved between 4-7% (in absence of additive noise) and 8-12% (in presence of additive noise). The significance of these improvements was statistically confirmed by a paired sample t-test, with 0.01 and 0.03 p-values, respectively. The proposed method for EEG phase calculation is very generic and may be applied to other EEG phase-based studies.
BibTeX
@inproceedings{seraj2017robust,
  author = {Esmaeil Seraj and Reza Sameni},
  title = {{Robust electroencephalogram phase estimation with applications in brain-computer interface systems}},
  booktitle = {Physiological},
  year = {2017},
  doi = {10.1088/1361-6579/aa5bba},
}
Figure