TL;DR
This paper introduces a novel, large-scale open dataset for wearable Steady-State Visual Evoked Potential (SSVEP)-based Brain-Computer Interfaces (BCIs), featuring data from 102 subjects using both wet and dry electrodes. The dataset aims to bridge the gap in practical, long-term wearable BCI applications by providing rich data for algorithm development and electrode comparison.
What Problem Does This Paper Address?
Brain-Computer Interfaces (BCIs) offer a promising avenue for communication and control, particularly for individuals with motor impairments. Among various BCI paradigms, SSVEP-based BCIs are attractive due to their high information transfer rate (ITR) and minimal user training. However, traditional EEG setups are bulky and require conductive gel, hindering real-world applications. Wearable BCI systems offer a solution, but face challenges in data acquisition, user experience, and the need for extensive datasets to develop robust algorithms. Furthermore, a comprehensive comparison of wet and dry electrodes in a wearable SSVEP context, especially considering long-term use and user comfort, has been lacking. This paper addresses this by creating and releasing a comprehensive dataset to facilitate research in practical, wearable SSVEP BCIs.
Methods
The study employed a wearable 8-channel wireless EEG acquisition system to record SSVEPs. A 12-target SSVEP-based BCI speller was used, where targets flickered at specific frequencies and phases. Two types of electrodes were investigated: traditional wet electrodes and novel dry electrodes. Each of the 102 participants underwent experiments with both electrode types, with the order randomized. Data acquisition involved 10 consecutive blocks per electrode type, with each block containing 12 trials. The experiment protocol included visual cues, synchronized triggers, and visual feedback. For offline analysis, the researchers utilized Filter Bank Canonical Correlation Analysis (FBCCA) and Filter Bank Task-Related Component Analysis (FBTRCA) to evaluate BCI performance. They also collected user experience questionnaires regarding electrode comfort and preference.
Datasets
The primary contribution of this paper is the introduction of a new, open dataset. This dataset comprises:
- EEG Data: 102 MATLAB MAT files, each containing an 8-channel EEG recording from a single subject. The data is structured as a 5-D matrix ([Channel index, Time points, Electrode index, Block index, Target index]) and includes 240 epochs per subject (12 targets × 10 blocks × 2 electrodes).
- Stimulation Information: A PDF file detailing the target index, labels, and corresponding stimulation parameters (frequency and phase).
- Subject Information and Questionnaires: A MAT file containing participant demographics and their responses to questionnaires on electrode comfort and preference.
- Impedance Information: A MAT file detailing electrode impedance measurements for each channel, block, electrode type, and subject.
This dataset is made publicly available on FigShare and the authors' website.
Key Insights and Ideas
- Wearable SSVEP is Feasible: The study demonstrates that wearable SSVEP BCIs can maintain stable performance over extended periods (up to 2 hours) without significant degradation, suggesting their viability for practical applications.
- Wet vs. Dry Electrode Trade-offs: While dry electrodes offer superior convenience and ease of use, wet electrodes generally provide better BCI performance (accuracy and ITR). User comfort is a significant factor, with a strong preference for wet electrodes despite their inconvenience.
- Offline Optimization Matters: Offline analysis methods like FBTRCA significantly outperform online algorithms, especially when using dry electrodes, highlighting the potential for improved performance through advanced signal processing.
- Individual Differences are Crucial: BCI performance varies considerably across individuals, and this variability is consistent between wet and dry electrodes, suggesting that individual physiological characteristics play a key role.
- Impedance and Performance Link: While not a direct cause-and-effect, there's a suggestion that reducing impedance in dry electrodes could lead to improved performance, offering a target for future electrode development.
- Cross-Electrode Transfer Learning: The proposed electrode-transfer (ET) method shows promise for reducing calibration time, particularly when transitioning to dry electrodes, by leveraging data from wet electrodes.
Key Results
- Signal Quality: Amplitudes and Signal-to-Noise Ratios (SNRs) of SSVEP signals were comparable between wet and dry electrodes.
- Online BCI Performance:
- Accuracy: Wet electrodes achieved an average classification accuracy of 0.83 ± 0.02, significantly higher than dry electrodes (0.63 ± 0.03).
- ITR: Wet electrodes yielded an average ITR of 76.93 ± 2.95 bits/min, compared to 51.21 ± 3.43 bits/min for dry electrodes.
- Offline BCI Performance (FBTRCA at 2s data length):
- Accuracy: Dry electrodes: 0.73 ± 0.03, Wet electrodes: 0.93 ± 0.01.
- ITR: Dry electrodes: 64.83 ± 3.74 bits/min, Wet electrodes: 92.71 ± 2.06 bits/min.
- User Preference: 83.33% of participants preferred wet electrodes for comfort, while only 20.59% found dry electrodes comfortable. Considering convenience, 68.63% still preferred wet electrodes.
- Impedance: Average impedance for dry electrodes was 261.67 ± 28.74 kΩ, significantly higher than wet electrodes (19.63 ± 1.20 kΩ).
Limitations and Future Work
- Dry Electrode Performance: The study acknowledges that dry electrodes, despite their convenience, exhibit significantly lower accuracy and ITR compared to wet electrodes. Future work should focus on improving dry electrode technology to enhance signal quality and user experience.
- Impedance-Performance Relationship: While a correlation was observed, the exact relationship between impedance and BCI performance for dry electrodes requires further investigation.
- Algorithm Optimization: The paper highlights that offline optimization can greatly improve BCI performance. Further research into optimizing online algorithms is needed for real-world applications.
- Dataset Expansion: The authors suggest the dataset can be used for analyzing BCI demographics and developing more efficient noise removal and channel selection methods for wearable BCIs.
- Long-Term Comfort: While participants reported on comfort, the long-term effects of wearing dry electrodes for extended periods (beyond the experiment duration) warrant further study.
Summary updated