Answers are grounded in this paper and cite their sources.
Motor imagery (MI) brain–computer interfaces (BCIs) enable communication through brain activity, but reliable decoding of electroencephalography (EEG) remains challenging due to low signal quality, inter-subject variability, and limited training data. This study investigates a lightweight hybrid framework combining Common Spatial Pattern (CSP), Filter Bank CSP (FBCSP), and their feature-level fusion with an exceptionally compact EEGNet classifier. Using a subject-specific, session-to-session evaluation protocol, we assess four classification tasks across BCI Competition IV-2a (9 subjects) and Lee2019_MI (54 subjects). The hybrid CSP–FBCSP framework achieves higher mean classification accuracy than a raw-EEG EEGNet baseline across all four tasks, with improvements of 3.0–11.2 percentage points, statistically significant in three tasks under uncorrected Wilcoxon tests. However, fusion significantly outperforms individual CSP or FBCSP front-ends in only two of eight paired comparisons. A matched-input-size ablation further shows that fusion achieves the highest mean accuracy on the two binary BCI IV-2a tasks, but not on the four-class task or Lee2019_MI. Feature-redundancy analysis provides additional insight into the overlap between spatial representations. These findings demonstrate the potential of classical spatial–spectral feature extraction for compact EEG classification while revealing that the additional benefit of feature fusion is task-dependent.
No takes yet. Share an insight, caveat, or question.
Alidostdargah et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: