Source-linked AI summary
Frequency Recognition in SSVEP-based BCI using Multiset Canonical Correlation Analysis
Yu Zhang, Guoxu Zhou, Jing Jin, Xingyu Wang, Andrzej Cichocki
TL;DR
Pre-constructed sine-cosine references may lack real EEG features needed for optimal SSVEP frequency recognition. MsetCCA learns EEG-based common features to optimize CCA reference signals, improving recognition over CCA and competing methods, especially with few channels and short windows.
Problem
Pre-constructed sine-cosine reference signals may lack features from real EEG data, limiting optimal SSVEP frequency recognition accuracy.
Method
MsetCCA learns multiple linear transforms through joint spatial filtering of EEG trials at the same stimulus frequency, combining extracted common features into training-data-based CCA reference signals.
Results
MsetCCA outperformed CCA, PCCA, and MwayCCA, with especially strong advantages for few channels and short time windows; with four channels and a 1-second window, all pairwise comparisons were significant.
Takeaways & Limitations
MsetCCA is a promising candidate for SSVEP frequency recognition in SSVEP-based BCIs and avoids requiring a pre-defined number of harmonics.
Takeaways & Limitations
Insufficient training trials may produce ineffective references, whereas too many introduce redundancy and increase memory, computation, and recognition time; the optimal number remains unresolved.
Abstract
from arXiv · showhide
Canonical correlation analysis (CCA) has been one of the most popular methods for frequency recognition in steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs). Despite its efficiency, a potential problem is that using pre-constructed sine-cosine waves as the required reference signals in the CCA method often does not result in the optimal recognition accuracy due to their lack of features from the real EEG data. To address this problem, this study proposes a novel method based on multiset canonical correlation analysis (MsetCCA) to optimize the reference signals used in the CCA method for SSVEP frequency recognition. The MsetCCA method learns multiple linear transforms that implement joint spatial filtering to maximize the overall correlation among canonical variates, and hence extracts SSVEP common features from multiple sets of EEG data recorded at the same stimulus frequency. The optimized reference signals are formed by combination of the common features and completely based on training data. Experimental study with EEG data from ten healthy subjects demonstrates that the MsetCCA method improves the recognition accuracy of SSVEP frequency in comparison with the CCA method and other two competing methods (multiway CCA (MwayCCA) and phase constrained CCA (PCCA)), especially for a small number of channels and a short time window length. The superiority indicates that the proposed MsetCCA method is a new promising candidate for frequency recognition in SSVEP-based BCIs.
1. Introduction
SSVEP-based BCIs recognize user commands from stimulus-frequency responses in EEG, but noise contamination makes accurate recognition with short time windows challenging. CCA is widely used, yet its pre-constructed sine-cosine references may lack features of real EEG data.
- SSVEP-based BCI: SSVEP is elicited at a visual flicker frequency and its harmonics, enabling BCI command recognition through EEG frequency detection.Responses occur over occipital scalp areas but can be contaminated by ongoing EEG activity and background noise.
- SSVEP-based BCI: High recognition accuracy with a short time window is important for developing high-performance SSVEP-based BCIs.
- CCA-based recognition: CCA recognizes SSVEP frequency by maximizing correlation between multichannel EEG signals and sine-cosine reference signals for each stimulus frequency.The stimulus frequency associated with the maximal correlation coefficient is selected.
- CCA-based recognition: Pre-constructed sine-cosine references may limit CCA recognition accuracy because they lack features from real EEG data.This motivates optimizing reference signals using information derived from EEG recordings.
- Proposed approach: MsetCCA optimizes reference signals by extracting common features from multiple EEG datasets recorded at the same stimulus frequency.The method applies joint spatial filtering and is presented as an alternative to existing approaches whose reference optimization still relies partly on sine-cosine waves.
2. Materials and Methods
The study records multichannel EEG during flickering visual stimulation and uses CCA-related methods to recognize SSVEP frequencies. MsetCCA learns training-data-based reference signals from shared features across trials before applying CCA to test data.
- Experimental setup: Ten healthy subjects viewed four stimuli flickering at 6, 8, 9, and 10 Hz during EEG recordings.EEG was recorded from 30 channels at 250 Hz and band-pass filtered from 4 to 45 Hz.
- CCA for SSVEP recognition: CCA recognizes SSVEP frequency by maximizing correlation between multichannel EEG test data and sine-cosine reference signals for candidate stimulus frequencies.The candidate frequency with the maximal correlation coefficient is selected.
- MsetCCA for SSVEP recognition: The method addresses the limitation that sine-cosine references lack important information contained in real EEG data.The proposed calibration procedure replaces artificial references with signals learned from EEG training data.
- MsetCCA for SSVEP recognition: MsetCCA applies the MAXVAR objective to multiple EEG training sets, finding spatial filters whose canonical variates share maximally correlated features.The canonical variates are treated as common features across trials recorded at the same stimulus frequency.
- MsetCCA for SSVEP recognition: The optimized reference set combines MsetCCA canonical variates, and CCA compares a new test set with each frequency-specific reference set.The frequency associated with the maximal correlation is then recognized.
- Comparison methods: MwayCCA optimizes references by correlating EEG tensor dimensions with sine-cosine waves, whereas PCCA embeds training-estimated phase information into those references.Both methods are compared with CCA and MsetCCA for SSVEP frequency recognition.
3. Results
Across harmonic, channel-count, and time-window analyses, MsetCCA achieved the strongest SSVEP recognition accuracy, with especially clear advantages under limited channels and short windows.
- Harmonics: MsetCCA achieved the best frequency recognition accuracy at all four time windows for H = 1, 2, and 3.
- Number of channels: For C = 4, 6, and 8, MsetCCA outperformed CCA, PCCA, and MwayCCA, especially with fewer channels and shorter time windows.
- Number of channels: For C = 4 at a 1 s time window, MsetCCA significantly outperformed CCA (p < 0.001), PCCA (p < 0.005), and MwayCCA (p < 0.05).
- Subject-wise performance: For most subjects, MsetCCA yielded higher accuracy than the other three methods across various time windows when C = 4.
- Stimulus frequencies: MsetCCA consistently outperformed the other three methods at each of the four stimulus frequencies.
- Overall result: The results indicate that MsetCCA is promising for developing high-performance SSVEP-based BCIs.
4. Discussion
The discussion presents MsetCCA as a training-data-based reference-signal optimization method that improves SSVEP recognition, while highlighting training-trial selection and computational considerations.
- Reference-signal optimization: MsetCCA optimizes CCA reference signals from training data rather than directly using pre-constructed sine-cosine waves.It extracts common features through joint spatial filtering of multiple trials at the same stimulus frequency.
- Reference-signal optimization: MsetCCA reference signals more accurately captured harmonic features in test data than sine-cosine waves, supporting more accurate frequency recognition.The comparison is illustrated in Fig. 7.
- Adaptive harmonics: MsetCCA does not require the pre-defined harmonic count H needed by CCA, PCCA, and MwayCCA.Its training procedure automatically estimates subject-specific SSVEP features and accurate harmonics from multitrial EEG data.
- Computational considerations: 0.357 s was required for MsetCCA reference-signal optimization, compared with 1.475 s for MwayCCA and 3.091 s for PCCA at TW = 4 s.Recognition after optimization took 0.0056 s using CCA.
- Training-trial selection: Insufficient training trials may produce ineffective reference signals, whereas too many can introduce redundancy and additional memory and computational requirements.Accuracy changed little beyond ten training trials, motivating further study of the optimal number of reference signals.
MsetCCA CCA
The paper contrasts MsetCCA-learned reference signals with sine-cosine references and emphasizes its accuracy gains, training-data trade-off, and efficient recognition.
- Training-data trade-off: Accuracy changed little when more than ten training trials were used for MsetCCA reference-signal optimization.An appropriate trial count can reduce training-data requirements without significant accuracy decrease.
- Practical choice: CCA remains preferred for zero-training SSVEP BCIs, while MsetCCA is preferred when recognition accuracy is the primary factor.The distinction reflects the cost of requiring training data for reference-signal optimization.
5. Conclusions
The study concludes that MsetCCA improves SSVEP frequency recognition by learning EEG-derived reference signals and outperforms CCA, MwayCCA, and PCCA in experiments with ten healthy subjects.
- Method: MsetCCA extracts common features from multitrial EEG data through joint spatial filtering and uses them to form optimized CCA reference signals.The method maximizes overall correlation among canonical variates.
- Results: Experiments with ten healthy subjects showed that MsetCCA outperformed CCA, MwayCCA, and PCCA in SSVEP frequency recognition.The improvement comes at the cost of using training data.
- Future work: Future work will investigate MsetCCA effectiveness across a wider range of stimulus frequencies.