Abstract
The accuracy of the classification of user intentions is essential for motor imagery (MI)-based brain–computer interface (BCI). Effective and appropriate training for users could help us produce the high reliability of mind decision making related with MI tasks. In this study, we aimed to investigate the effects of visual guidance on the classification performance of MI-based BCI. Leveraging both the single-subject and the multi-subject BCI paradigms, we train and classify MI tasks with three different scenarios in a 3D virtual environment: non-object-directed, static-object-directed, and dynamic object-directed. We demonstrate that both static and dynamic object-directed scenarios provide better classification accuracy than the non-object-directed case, shorten response times, and are suitable for small training datasets. Furthermore, the multi-subject BCI paradigm significantly improves performance compared to the single-subject approach.
Introduction
Brain-Computer Interfaces (BCIs) allow users to control external devices through "Motor Imagery" (MI)—the mental rehearsal of physical movement. A significant challenge in this field is "BCI illiteracy," where users struggle to generate consistent EEG signals. Traditional training often relies on abstract cues, which can be disengaging and difficult to map to specific intentions. This research explores whether immersive, object-directed 3D visual guidance can improve signal quality and classification accuracy.
Methods and Materials
The study involved five healthy participants (aged 25–30) using a 32-channel EEG system. The core of the experiment compared three 3D visual guidance scenarios:
- Non-Object-Directed (NOD): Abstract 3D hand movements without a goal.
- Static-Object-Directed (SOD): A hand moving to grasp a stationary object, such as an apple.
- Dynamic-Object-Directed (DOD): A hand attempting to catch a moving target, specifically a flying insect.
Technical Pipeline
The EEG signals underwent a processing chain to isolate relevant motor rhythms (7-30Hz). Common Spatial Pattern (CSP) was employed for feature extraction to distinguish between left and right-hand imagery, followed by Linear Discriminant Analysis (LDA) for classification.
A key innovation was the multi-subject paradigm, which uses a weighted voting system to combine classifiers from multiple users, creating a more robust ensemble model.
Key Results
The study yielded several critical insights:
- Goal-Directed Superiority: Scenarios with clear goals (SOD and DOD) significantly outperformed abstract training. Mean individual accuracies were 78.3% for NOD, 85.1% for SOD, and 82.3% for DOD.
- Data Efficiency: SOD and DOD reached over 75% accuracy using only 50% of the training data required by NOD, suggesting faster user onboarding.
- The Power of the Crowd: The multi-subject paradigm achieved near-perfect results. By pooling data from 5 subjects, accuracies reached 98.2% for SOD and 98.5% for DOD.
- Temporal Optimization: Analysis of trial phases revealed that the "middle phase" of imagery provides the most discriminative data, suggesting optimal windows for BCI "listening."
Discussion and Conclusion
The results confirm that context-rich, object-directed training helps users produce more robust neural patterns. While the small sample size is a limitation, the dramatic performance boost from the multi-subject paradigm offers a promising path for stabilizing BCI systems against individual signal noise. Future work will focus on real-time implementation and personalized frequency band optimization to further enhance responsiveness and reliability in practical applications.
Summary updated