Background: Motor imagery (MI)-based Brain–Computer Interfaces (BCIs) depend on efficientextraction of oscillatory EEG features and, critically, on the ability of classifiers to generalize acrosssubjects. Despite the widespread use of classical approaches such as Common Spatial Patterns withLinear Discriminant Analysis (CSP-LDA), cross-subject variability remains a major obstacle topractical deployment.New method: This study presents a systematic comparative evaluation between a classical pipeline(CSP-LDA) and a compact deep learning approach (EEGNet) for MI classification under a rigorouscross-subject setting. Using the BCI Competition IV-2b dataset, we implemented a leakage-freepipeline with training-fold-only normalization, Leave-One-Subject-Out (LOSO) cross-validation, andmulti-seed robustness analysis to ensure a correct and reproducible comparison.Results: EEGNet achieved a mean accuracy of 71.10% with Cohen’s kappa of 0.4220, whereasCSP-LDA achieved a mean accuracy of 68.23% with kappa of 0.3646. Multi-seed analysis confirmedthe consistency of the EEGNet advantage across different random initializations.Comparison with existing methods: Compared with the classical CSP-LDA baseline, EEGNetshowed superior cross-subject generalization, with an absolute gain of 2.87 percentage points inaccuracy and improved agreement as measured by Cohen’s kappa. These findings indicate thatcompact convolutional neural networks can better capture inter-subject invariant representations thancovariance-based classical methods under correctly controlled training conditions.Conclusions: Under a rigorous LOSO protocol with explicit prevention of data leakage, EEGNetoutperformed CSP-LDA in cross-subject motor imagery classification. The results support the use ofcompact CNN-based models as a stronger zero-calibration baseline for MI-based BCIs and reinforcethe importance of implementation correctness in comparative evaluations
Bruno Rocha Guimarães (Tue,) studied this question.