Efficient Low-Frequency SSVEP Detection with Wearable EEG Using Normalized Canonical Correlation Analysis

Victor Javier Kartsch, Velu Prabhakar Kumaravel, Simone Benatti, Giorgio Vallortigara, Luca Benini, Elisabetta Farella, Marco Buiatti2022 · DOI

TL;DR

This paper introduces Normalized Canonical Correlation Analysis (NCCA) as a highly efficient method for detecting Steady-State Visual Evoked Potentials (SSVEPs) at very low frequencies (1-6 Hz) using wearable EEG. NCCA significantly outperforms traditional spectral analysis, requiring much shorter recording times (as little as 4 seconds) and enabling new possibilities for out-of-the-lab neurocognitive testing, especially in vulnerable populations.

What Problem Does This Paper Address?

Extracting reliable brain activity from EEG signals typically requires long recording sessions due to noise and artifacts. Steady-State Visual Evoked Potentials (SSVEPs), generated by periodic visual stimulation, offer a way to overcome this by "tagging" brain responses to specific frequencies. While effective at higher frequencies (8-20 Hz), using SSVEPs at lower frequencies (delta and theta bands, 0.5-7 Hz) is crucial for studying higher-level cognitive processes (like face perception in infants) or for populations with limited attention spans. However, low-frequency SSVEP detection is challenging due to increased background EEG noise and the need for longer stimulation periods to capture the response. Furthermore, wearable EEG systems, while offering portability, often have limited channels and are susceptible to noise, exacerbating these challenges. This paper aims to find an efficient and robust method for low-frequency SSVEP detection using wearable EEG, minimizing data requirements.

Methods

The core of this paper is the introduction and evaluation of Normalized Canonical Correlation Analysis (NCCA). NCCA is a frequency-normalized extension of Canonical Correlation Analysis (CCA). CCA is a multivariate statistical technique that finds linear combinations of two sets of variables that maximize their correlation. In this context, one set is the EEG signals from multiple channels, and the other is a set of reference signals representing the expected SSVEP response at a specific frequency and its harmonics. NCCA then computes the ratio of the CCA correlation at the target frequency to the average correlation at adjacent frequencies, effectively creating a "peak" indicator.

The authors compared NCCA against Frequency Tagging Analysis (FTA), a standard method based on Fourier Transform and power spectrum analysis. They also employed a Frequency-Tagged Response (FTR) metric, which is a ratio of the power spectrum at the tag frequency to a fitted background power spectrum.

Key methodological aspects include:

  • Wearable EEG System: A BioWolf platform with 8 dry electrodes was used, highlighting the use of a low-density, portable system.
  • Low-Frequency Stimulation: Visual stimuli were presented at four frequencies: 1 Hz, 3.125 Hz, 7.8125 Hz, and 10.6125 Hz, covering both typical BCI and lower cognitive research ranges.
  • Data Segmentation: The performance of NCCA and FTA was evaluated across progressively shorter data window lengths to determine the minimum recording time required.
  • Statistical Significance: Wilcoxon signed-rank tests were used to assess the statistical significance of the detected peaks.

Datasets

The study utilized data collected from 15 healthy adult subjects using the described 8-channel wearable EEG system. For statistical power analysis and sample size estimation, an independent dataset from 11 adult subjects recorded with a 64-channel EEG system was also used. No new benchmark datasets were introduced; rather, the study focused on evaluating existing methods on a novel combination of low-frequency SSVEP and wearable EEG.

Key Insights and Ideas

  • NCCA's Superiority at Low Frequencies: NCCA demonstrates significantly better performance than traditional power spectrum analysis (FTA) at low stimulation frequencies (especially 1 Hz). This is attributed to CCA's unbounded frequency resolution and its ability to inherently handle multivariate data without explicit channel selection or artifact removal preprocessing.
  • Reduced Data Requirements: NCCA enables reliable SSVEP detection with remarkably short data segments. For the challenging 1 Hz stimulation, as little as 4 seconds of data were sufficient for statistically significant detection, a substantial improvement over methods requiring longer recordings.
  • Wearable EEG Viability for Low-Frequency SSVEP: The study validates that even low-density, portable EEG systems can achieve high-quality SSVEP detection at low frequencies, opening doors for out-of-the-lab research.
  • Artifact Robustness: NCCA's performance was not significantly improved by preliminary artifact removal, suggesting it is inherently more robust to artifacts compared to FTA, which often requires such preprocessing.

Key Results

  • NCCA Detects SSVEP Rapidly: NCCA could detect statistically significant SSVEP responses at 1 Hz with just 4 seconds of data, reaching maximum significance with 8 seconds. For higher frequencies (3.125 Hz and above), 2 seconds of data were often sufficient.
  • FTA Requires Longer Data: FTA required longer data segments (e.g., 6 seconds for 1 Hz) to reach statistical significance and showed slower improvement with increasing window length compared to NCCA.
  • NCCA's Higher Specificity: NCCA showed better specificity, with its measures for the target stimulation frequency clearly separating from those for other frequencies or rest conditions, especially at lower frequencies.
  • Statistical Significance: NCCA consistently achieved higher statistical significance (p < 0.01) with shorter data windows across all tested frequencies compared to FTA.

Limitations and Future Work

  • Artifact Robustness Exploration: While NCCA showed good performance without explicit artifact removal, further research is suggested to explore its robustness to artifacts in more challenging scenarios or with different stimulation modalities.
  • Generalizability to Other Populations: The authors suggest testing NCCA's efficiency in developing or aging populations, as well as exploring its application to other fields like speech processing or pain processing.
  • Open-Source Software: The authors plan to publicly share a user-friendly, open-source version of NCCA in the near future, which will be a valuable resource for the research community.

Summary updated

BibTeX
@misc{kartsch2022efficient,
  author = {Victor Javier Kartsch and Velu Prabhakar Kumaravel and Simone Benatti and Giorgio Vallortigara and Luca Benini and Elisabetta Farella and Marco Buiatti},
  title = {{Efficient Low-Frequency SSVEP Detection with Wearable EEG Using Normalized Canonical Correlation Analysis}},
  year = {2022},
  doi = {10.3390/s22249803},
}
Figure