A dynamic subject-invariant fragment mixing strategy to suppress subject variability in EEG imagined speech...

This study presents the first systematic benchmark on a recently released 31-class Arabic imagined-speech EEG dataset for which no prior computational analysis has been reported. The task is exceptionally challenging due to pronounced inter- and intra-subject variability and frequent signal degradations, yielding very low performance when existing architectures are applied directly (baseline accuracy typically within the ∼ 7-15% range). To address these limitations, we propose NeuroSilentia, an EEG-tailored spatio-temporal model that integrates channel reweighting and efficient multi-scale temporal modeling, improving accuracy to 20.45%. We then examine a wide set of subject-invariant and domain-alignment strategies including contrastive objectives, DANN, MMD, CORAL, center/prototype losses, and Riemannian geometry-based methods, showing that global alignment alone provides limited gains in this high-cardinality setting. Building on these findings, we introduce Dynamic Subject-Invariant Fragment Mixing (DSIFM), an epoch-wise fragment-level mixing strategy that disrupts subject- and session-specific shortcuts while preserving class structure. Compared with the NeuroSilentia baseline, DSIFM substantially improves generalization, and when combined with contrastive learning reaches 51.60% accuracy. Data cleaning further increases performance, achieving 60.29% accuracy in the full 31-class setting. Extensive evaluations including 10-fold cross-validation, subject/class/channel-wise analyses, confusion-matrix diagnostics, t-SNE representation studies, and reduced-class transfer experiments (16/8/4 classes) consistently confirm the robustness of the proposed approach. Overall, this work delivers the first comprehensive benchmark for 31-class Arabic imagined speech EEG decoding and establishes DSIFM as an effective strategy for mitigating subject variability in complex multi-class EEG classification.