Frequency Recognition in SSVEP-Based BCI using Multiset Canonical Correlation Analysis

Yu Zhang, Guoxu Zhou, Jing Jin, Xingyu Wang, Andrzej CichockiInt. J. Neural Syst 2014 · OpenReview

TL;DR

This paper introduces MsetCCA, a novel method for improving the accuracy of brain-computer interfaces (BCIs) that rely on steady-state visual evoked potentials (SSVEPs). By learning optimal reference signals directly from training data, MsetCCA significantly outperforms existing methods, especially in challenging scenarios with limited data.

What Problem Does This Paper Address?

Brain-computer interfaces (BCIs) offer a way for individuals with severe motor impairments to communicate and interact with their environment by translating brain activity into commands. One popular type of BCI utilizes steady-state visual evoked potentials (SSVEPs), which are brain responses to flickering visual stimuli. The core challenge in SSVEP-based BCIs is accurately identifying the frequency of the visual stimulus that the user is focusing on. Traditional methods, like Canonical Correlation Analysis (CCA), often use pre-defined sine-cosine waves as reference signals. However, these artificial signals may not fully capture the complex characteristics of real brain data, leading to suboptimal recognition accuracy, particularly when dealing with limited training data or short observation windows. This paper aims to address this limitation by developing a more adaptive and data-driven approach to reference signal generation.

Methods

The paper proposes a new method called Multiset Canonical Correlation Analysis (MsetCCA) for SSVEP frequency recognition. MsetCCA builds upon the principles of CCA but extends it to handle multiple sets of data. The key innovation lies in how it generates reference signals. Instead of relying on fixed sine-cosine waves, MsetCCA learns optimal reference signals directly from the user's training EEG data.

Here's a breakdown of the approach:

  • Reference Signal Optimization: MsetCCA treats multiple training trials for a specific stimulus frequency as different "sets" of data. It then applies a multiset canonical correlation analysis technique (specifically, the MAXVAR objective function) to find multiple linear transforms (spatial filters). These transforms are designed to maximize the overall correlation among "canonical variates" derived from these training sets. These canonical variates represent common features extracted from the EEG data, which are hypothesized to be more representative of the actual SSVEP response than artificial sine-cosine waves.
  • Optimized Reference Signals: The learned common features from the training data are then combined to form the optimized reference signals for each stimulus frequency.
  • SSVEP Frequency Recognition: Once these optimized reference signals are generated, the standard CCA method is used to compare a new, unseen EEG test signal against each of these optimized reference signals. The stimulus frequency that yields the highest correlation is identified as the user's intended command.

The paper also compares MsetCCA against three other methods:

  • CCA: The standard method using pre-defined sine-cosine waves.
  • MwayCCA (Multiway CCA): An extension of CCA that considers multiple dimensions of EEG data but still relies on pre-constructed reference signals.
  • PCCA (Phase Constrained CCA): Another method that attempts to improve CCA by incorporating phase information from training data into the reference signals.

Datasets

The study used EEG data collected from ten healthy subjects. The experiments involved presenting visual stimuli at four different frequencies (6 Hz, 8 Hz, 9 Hz, and 10 Hz). The EEG data was recorded using 30 channels at a sampling rate of 250 Hz. For analysis, specific subsets of channels were selected (e.g., 8 channels: P7, P3, Pz, P4, P8, O1, Oz, and O2). The paper does not introduce a new benchmark dataset; rather, it uses data from a controlled experimental setup.

Key Insights and Ideas

  • Data-Driven Reference Signals are Superior: The core insight is that learning reference signals directly from the user's own EEG training data (as MsetCCA does) is significantly more effective than using generic, pre-defined sine-cosine waves. This is because real brain signals are complex and contain nuances not captured by artificial signals.
  • Common Features Capture SSVEP Characteristics: MsetCCA's ability to extract "common features" across multiple training trials for a given frequency is key. These features are more representative of the actual SSVEP response, including its harmonic components, leading to better recognition.
  • No Need for Pre-defined Harmonics: A significant advantage of MsetCCA is that it automatically estimates the optimal features, meaning it doesn't require the user or researcher to pre-specify the number of harmonics (H) to consider, unlike CCA, PCCA, and MwayCCA. This simplifies the process and potentially leads to better performance as the optimal number of harmonics can vary.
  • Robustness in Challenging Conditions: MsetCCA shows particular strength when dealing with limited data (fewer channels) and short time windows, which are common challenges in real-world BCI applications.

Key Results

  • Significantly Higher Accuracy: MsetCCA consistently outperformed CCA, PCCA, and MwayCCA across various time window lengths (TWs) and numbers of channels (C). For instance, with C=4 channels and a TW of 1s, MsetCCA showed statistically significant improvements over CCA (p < 0.001), PCCA (p < 0.005), and MwayCCA (p < 0.05).
  • Performance with Varying Harmonics: The accuracy of MsetCCA was the same for H=1, 2, and 3, highlighting its ability to adapt without pre-defined harmonic assumptions. Other methods showed improvements with higher H, but MsetCCA's performance was already optimal.
  • Efficiency: While MsetCCA requires a training phase for reference signal optimization, the computational time for this process was found to be efficient (e.g., 0.357s for a TW of 4s). Once optimized, the SSVEP recognition using CCA was very fast (0.0056s).
  • Impact of Training Data: The study indicated that accuracy improved with the number of training trials, but the gains plateaued after about ten trials, suggesting that a moderate amount of training data is sufficient.

Limitations and Future Work

  • Training Data Requirement: A key limitation acknowledged is that MsetCCA requires training data for reference signal optimization. This makes it less suitable for "zero-training" BCI applications where immediate use without prior calibration is desired. In such scenarios, CCA might still be preferred.
  • Optimal Number of Reference Signals: The paper suggests that further research is needed to determine the optimal number of reference signals (which relates to the number of training trials) for MsetCCA to balance accuracy and computational efficiency.
  • Wide Range of Stimulus Frequencies: The authors propose investigating the effectiveness of MsetCCA on a wider range of stimulus frequencies in future work.

Summary updated

BibTeX
@article{zhang2014frequency,
  author = {Yu Zhang and Guoxu Zhou and Jing Jin and Xingyu Wang and Andrzej Cichocki},
  title = {{Frequency Recognition in SSVEP-Based BCI using Multiset Canonical Correlation Analysis}},
  journal = {Int. J. Neural Syst.},
  year = {2014},
  url = {https://openreview.net/forum?id=FAZkpcIubu},
}
Figure