Decoding English Alphabet Letters Using EEG Phase Information

YiYan Wang, Pingxiao Wang, Yuguo YuFrontiers in Neuroscience 2018 · DOI

TL;DR

This study demonstrates that the phase of low-frequency brain oscillations (theta and alpha bands) in EEG signals can be used to decode English alphabet letters with higher accuracy than using signal power. This finding opens new avenues for brain-computer interface (BCI) development, particularly for communication aids.

What Problem Does This Paper Address?

Brain-computer interfaces (BCIs) are crucial for enabling severely impaired individuals to interact with the world, often through tasks like typing. While existing BCI methods like P300 and SSVEP responses have shown promise, there's growing evidence that the dynamic patterns of neural oscillations, specifically their phase and power, carry significant information about sensory stimuli. This paper investigates whether these oscillatory patterns, particularly the phase information, can be directly decoded from EEG to identify individual English alphabet letters, a fundamental building block for communication BCIs. The researchers aim to determine which frequency bands and time windows are most informative for this decoding task and explore the spatial distribution of this information across the scalp.

Methods

The researchers employed a combination of EEG recording, signal processing, and machine learning techniques.

  • EEG Recording: EEG data was collected from 14 healthy participants performing a visual letter recognition task. They were presented with five lowercase letters ("a," "e," "i," "o," "t") on a screen and instructed to silently read them.
  • Signal Preprocessing: Raw EEG data was bandpass filtered (0.5-220 Hz), epoched around letter presentation, and baseline corrected. Artifacts from eye movements and blinks were removed using Independent Component Analysis (ICA).
  • Feature Extraction: The artifact-cleaned EEG signals were then transformed using the Hilbert transform to extract amplitude and phase sequences. These sequences were further filtered into specific frequency bands (delta, theta, alpha, beta, low-gamma, high-gamma).
  • Classification: A Support Vector Machine (SVM) classifier was used for multi-class classification to distinguish between the five letters. The SVM's parameters were optimized using a gradient ascent approach.
  • Validation: To ensure the robustness of the results, a 30-fold cross-validation approach was used. Additionally, a "shuffled-label training set" method was employed to statistically validate the classification accuracy against chance levels.

Datasets

The study utilized EEG data collected from 14 healthy participants performing a visual letter recognition task. This dataset was specifically generated for this research and serves as the benchmark for their decoding approach.

Key Insights and Ideas

  • Phase is King (for Letter Decoding): The most striking finding is that EEG phase information provides significantly higher decoding accuracy for letters compared to EEG power information. This suggests that the timing of neural oscillations, rather than just their strength, is a critical carrier of information for visual stimuli like letters.
  • Theta and Alpha Bands Shine: Low-frequency oscillations, specifically theta (4-8 Hz) and alpha (8-14 Hz) bands, were found to be the most informative for decoding letters based on their phase patterns. This aligns with previous research highlighting the role of these bands in cognitive processes.
  • The Right Time and Place: The optimal time window for decoding letters was found to be between 180 and 380 ms after stimulus presentation. Furthermore, the occipital and occipital-temporal scalp regions (specifically PO7 and PO8 electrodes) showed the highest classification accuracy, suggesting the involvement of visual processing areas.
  • Beyond Traditional BCI: This research offers a novel approach to BCI by moving beyond traditional event-related potentials (ERPs) and focusing on the continuous phase information within ongoing neural oscillations.

Key Results

  • Phase vs. Power: EEG phase decoding achieved significantly higher accuracy than EEG power decoding across multiple electrodes and subjects (p < 10⁻⁹).
  • Frequency Band Performance: Theta and alpha bands showed the highest classification accuracy for phase decoding, with no significant difference between them. For power decoding, theta band was significantly better than alpha.
  • Optimal Time Window: The best decoding performance was observed for analysis windows starting around 180 ms and extending to 380 ms after letter presentation.
  • Spatial Localization: Electrodes in the lateral occipital and posterior temporal regions (PO7 and PO8) showed the highest classification accuracies.
  • Accuracy Levels: The highest accuracy achieved for wide-band EEG phase sequence was 46.61% (chance level was 20%). For the optimal time window and frequency bands, accuracies in the theta and alpha phases reached around 35-36%.

Limitations and Future Work

  • Subject Dependence: The study acknowledges that performance can vary significantly between subjects, similar to other BCI approaches. Future work could explore subject training to improve classification performance.
  • Real-time vs. Offline: The current method involves offline analysis, collecting sufficient data before decoding. Future research aims to achieve real-time decoding by leveraging faster computing and improved algorithms.
  • SVM as a "Black Box": The use of SVM, while effective, is described as a "black box." Further research is needed to fully understand how stimulus characteristics are encoded into EEG oscillation phase patterns and the limitations of this approach.
  • Auditory vs. Visual Input: While participants were instructed to read letters silently, the possibility of imaginary pronunciation influencing the EEG signals (particularly theta oscillations) is raised. Future experiments should aim to disentangle purely visual processing from potential auditory-motor influences.
  • Cross-Frequency Coupling: The paper mentions the potential importance of cross-frequency coupling (interactions between different frequency bands) for decoding, which warrants further investigation.
  • Spatial Resolution: While occipital-temporal regions showed promise, combining EEG with functional MRI could provide better spatial resolution for identifying precise cortical locations involved in visual stimulus encoding.

Summary updated

Abstract

Increasing evidence indicates that the phase pattern and power of the low frequency oscillations of brain electroencephalograms (EEG) contain significant information during the human cognition of sensory signals such as auditory and visual stimuli. Here, we investigate whether and how the letters of the alphabet can be directly decoded from EEG phase and power data. In addition, we investigate how different band oscillations contribute to the classification and determine the critical time periods. An English letter recognition task was assigned, and statistical analyses were conducted to decode the EEG signal corresponding to each letter visualized on a computer screen. We applied support vector machine (SVM) with a self-adaptation method to learn the potential features for classification. It was observed that the EEG phase signals have a higher decoding accuracy than the oscillation power information. Low-frequency theta and alpha oscillations have phase information with higher accuracy than do other bands. The decoding performance was best when the analysis period began from 180 ms to 380 ms after stimulus presentation, especially in the lateral occipital and posterior temporal scalp regions (PO7 and PO8). These results may provide a new approach for brain-computer interface techniques (BCI) and may deepen our understanding of EEG oscillations in cognition.
BibTeX
@inproceedings{wang2018decoding,
  author = {YiYan Wang and Pingxiao Wang and Yuguo Yu},
  title = {{Decoding English Alphabet Letters Using EEG Phase Information}},
  booktitle = {Frontiers in Neuroscience},
  year = {2018},
  doi = {10.3389/fnins.2018.00062},
}
Figure