A multi-day and high-quality EEG dataset for motor imagery brain-computer interface.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Developing robust electroencephalography (EEG)-based brain-computer interfaces (BCIs) faces challenges with inter-day variability and low signal-to-noise ratios, particularly for motor imagery (MI) tasks.
• Large, reliable datasets are crucial for training BCI models that generalize across sessions and subjects, mitigating inherent EEG signal instability.

Purpose of the Study:

• To introduce a comprehensive motor imagery (MI) dataset collected during the 2019 World Robot Conference Contest-BCI Robot Contest.
• To facilitate research in cross-session and cross-subject motor imagery BCI by providing raw and preprocessed EEG data.

Main Methods:

• Collected EEG data from 62 healthy participants across three recording sessions.
• Included two experimental paradigms: two-class (left/right hand grasping) and three-class (left/right hand grasping, foot hooking).
• Utilized deep learning models (EEGNet for two-class, deepConvNet for three-class) to evaluate dataset performance.

Main Results:

• Achieved an average classification accuracy of 85.32% for two-class MI tasks using EEGNet.
• Attained an average classification accuracy of 76.90% for three-class MI tasks using deepConvNet.
• The dataset includes both raw and preprocessed EEG data for diverse research applications.

Conclusions:

• The presented MI dataset is a valuable resource for advancing EEG-based BCI research.
• This dataset will aid in addressing critical challenges related to cross-session and cross-subject variability in motor imagery BCI systems.
• The availability of this dataset encourages further development and validation of BCI algorithms.