| dc.description.abstract | Background: Non-invasive Brain-Computer Interface (BCI) studies primarily center on the
motor imagery (MI) concept, where multi-channel Electroencephalogram (EEG) signals are
collected and characterized by patterns for different imagined tasks. Previous studies put extensive
efforts into data-driven techniques to improve classification performance on benchmark datasets;
however, other aspects, such as experimental factors, still lack thorough investigation. This pilot
study aims to evaluate the effect of different cue-based protocols on within-subject MI-BCI
baseline performance to better guide the experimental instructions on a specific group of users.
Materials and Method: An Emotiv EEG headset kit integrated into the Lab-Streaming-Layer
(LSL) was used for data acquisition. Three PsychoPy-based protocols were designed, namely, G1,
G2, and G3, incorporating different visual instructions of image-cue, arrow-cue, and arrow-cuefeedback utilizing Event-Related (de)Synchronization (ERD/ERS) demonstration, respectively.
Imagery data (left/right hand/foot) from 12 healthy college participants (age 20~22, five females)
were collected (15 trials/task/run) and randomly allocated for each designated protocol. A
processing framework was implemented using a conventional Lasso-based sparse Filter Bank
Common Spatial Pattern (SFBCSP) for feature extraction/selection and Linear Discriminant
Analysis (LDA) for classification to assess the baseline performance. Average ROC (5-fold crossvalidation) was calculated for the upper-limb binary model of each run with different nonoverlapping time segments. Statistical non-parametric tests were used for within-group and crossgroup comparative analysis.
Results: In cross-group analysis, an average performance combining all runs of G1, G2, and G3
are 48.8%, 59.8%, and 60.1%, respectively, where it shows significant differences in G1&G2
(p<0.05) and G1&G3 (p<0.05) but not in G2&G3. In within-group analysis, average performance
between run1 & run2 is as follows: G1 (52.7% & 44.8%); G2 (62.0% & 57.8%); G3 (52.5% &
67.7%) where G3 group yields significant improvement (run2 > run1, p<0.05), while no statistical
difference has been found within the G1 or G2 group. In the after-run self-assessment analysis, while few elements strongly correlate with the overall performance, no significant difference was
found between the image-cue and arrow-cue groups.
Conclusion: The preliminary result highlights that different instructions (arrow/image cue &
feedback) may affect the within-session performance between runs while reporting no evidence of
changing the subject’s psychological factors. The statistical analysis also suggests that verbal
feedback with arrow-cue can enhance the model's efficacy, which can be further explained by
orienting the alpha-band ERD/ERS response. A key contribution of this project, along with the
findings, is the creation of a preliminary dataset, which captures the effects of different cueing
systems on motor imagery performance. Future studies may explore other human-based factors
considering the motor response-ability within the more extensive target group of users, potentially
advancing BCI application in a personalized paradigm. | en_US |
