TL;DR
This paper critically examines the reliability of using Electroencephalogram (EEG) phase signals for interpreting brain activity, particularly in cognitive neuroscience and Brain-Computer Interfaces (BCIs). It highlights a significant problem called the Low-Amplitude Analytic Signal (LAAS) issue, where low signal-to-noise ratios can introduce spurious phase variations, leading to potentially incorrect interpretations of cognitive states. The authors propose a method to identify and mitigate these artifacts, emphasizing the need for careful consideration of signal amplitude when analyzing EEG phase.
What Problem Does This Paper Address?
The field of neuroscience increasingly relies on EEG phase information to understand brain function, from cognitive processes to brain connectivity and BCI applications. The phase of brain signals is thought to reflect neural communication and synchronized firing. However, a critical challenge arises from the "Low-Amplitude Analytic Signal" (LAAS) problem. This occurs when the desired, narrow-band EEG signal has a low amplitude, making it susceptible to noise from the broader background EEG. This noise can introduce artificial spikes and shifts in the extracted phase, leading to misinterpretations of brain activity. The paper aims to demonstrate how these LAAS artifacts can compromise the utility and reliability of commonly used EEG phase-based metrics and to investigate the implications for cognitive neuroscience research.
Methods
The authors employ a two-pronged approach:
- Conventional Phase Estimation: They first detail the standard method for extracting EEG phase. This involves:
- Frequency Filtering: Applying a narrow-band Finite Impulse Response (FIR) filter to isolate the desired frequency band.
- Analytic Signal Formation: Using the Hilbert transform to create an analytic signal from the filtered EEG.
- Phase Extraction: Deriving the instantaneous phase (IP) and instantaneous amplitude (IA) from the analytic signal.
- LAAS-D (Low-Amplitude Analytic Signal Discarded) Phase Estimation: To address the LAAS problem, they introduce a modified procedure:
- LAAS Detection: Identifying time instants where the analytic signal amplitude falls below a predefined threshold (indicating low signal-to-noise ratio).
- Discarding Artifacts: Simply discarding the phase values that occur concurrently with these low-amplitude segments.
- Recalculation: Re-calculating phase-related quantities using this "cleaned" phase information.
They then compare the results obtained from both methods for several popular phase-related quantities, including Phase Derivative (PD)/Instantaneous Frequency (IF), Phase Shift (PS), Phase Lock (PL), and Phase-Locking Value (PLV).
Datasets
The study utilized a dataset consisting of 5 trials of 128-channel EEG recordings from 4 different subjects. These recordings were obtained during visual stimuli, specifically steady-state Visual Evoked Potentials (SSVEPs) sampled at 256Hz. The SSVEP stimulation frequencies used were 8Hz, 14Hz, and 28Hz.
Key Insights and Ideas
- The "LAAS Problem" is a Silent Killer of Reliability: The core insight is that even with standard filtering, low signal amplitudes in EEG can inject significant, non-physiological noise into phase estimations. This noise can manifest as spurious phase shifts and spikes, making interpretations unreliable.
- Amplitude is Not Just an Afterthought: The paper strongly suggests that when analyzing EEG phase, the instantaneous amplitude of the signal cannot be ignored. The amplitude provides crucial context about the signal's quality and the potential for artifacts.
- Common Phase Metrics are Vulnerable: Popular metrics like Phase Resetting (PR) and Phase Shift (PS) are shown to be particularly sensitive to LAAS. The study demonstrates how discarding phase values during low-amplitude segments can drastically alter the patterns observed for these metrics.
- Clinical Interpretations May Be Flawed: The authors raise a significant concern that many previous studies relying on EEG phase for cognitive interpretations (e.g., in epilepsy, dementia, autism, or cognitive states) might have been influenced by LAAS artifacts, potentially leading to incorrect conclusions.
- A Simple Solution Can Make a Big Difference: While not a perfect solution, the proposed LAAS-D method of simply discarding phase data during low-amplitude periods significantly improves the reliability of phase-based analyses, highlighting the importance of addressing signal quality.
Key Results
- Visual Evidence of Artifacts: Figures in the paper clearly illustrate how low-amplitude segments in the EEG signal (and its analytic form) directly correspond to spurious phase shifts and spikes in the extracted phase and instantaneous frequency.
- Altered Phase Resetting Patterns: For both single-channel and multi-channel EEG, discarding phase values concurrent with low amplitudes (LAAS-D) led to noticeable changes in phase resetting patterns, suggesting that many previously identified PR events were artifacts.
- Correlation Differences: Linear (Pearson) and non-linear (Spearman) correlation analyses between different phase estimation methods (conventional, LAAS-D, and a more advanced TFP method) showed significant differences, underscoring how the choice of estimation procedure impacts the interpretation of phase synchrony.
- Reduced Spurious Events: The LAAS-D approach significantly reduced the number of detected phase shift events that were not physiologically plausible, indicating that many such events in conventional analyses are indeed artifacts.
Limitations and Future Work
Limitations:
- Subjectivity of Threshold: The threshold used to define "low amplitude" in the LAAS-D method is acknowledged as subjective and dataset-dependent, requiring empirical tuning.
- Simplicity of LAAS-D: The LAAS-D method is a simple approach that discards data. It doesn't attempt to "clean" the signal within low-amplitude segments, which might lead to loss of potentially valid information.
- Focus on Specific Metrics: While the paper covers several key phase metrics, the full spectrum of EEG phase analysis techniques might not have been explored.
Future Work:
- More Sophisticated Solutions: The authors suggest exploring more advanced statistical approaches, such as using confidence intervals based on analytic envelopes to define "normal" phase variations, or employing robust phase estimation methods like Transfer Function Perturbation (TFP).
- Re-evaluation of Past Studies: The findings imply a need to re-examine previous research that relied heavily on EEG phase without adequately addressing the LAAS problem, potentially leading to a re-interpretation of existing results.
- Clinical Validation: Further investigation is needed to rigorously prove the integrity of phase-based interpretations for cognitive and pathological states, especially after applying robust estimation techniques.
Summary updated