Towards Balanced Bias-Variance With SincDualFormer: A Dual-Scale Sinc-Filterbank Transformer Model With SR-BandMix...

Motor imagery electroencephalography (MI-EEG) decoding remains challenging because of its low signal-to-noise ratio, substantial inter-subject variability, and limited training data. This study proposes SincDualFormer, a dual-rhythm architecture that combines physiologically constrained Sinc filtering and unconstrained temporal convolution within parallel mu and beta-aligned branches. The two complementary representations are independently recalibrated by BandSE and projected through depthwise spatial convolutions before pointwise fusion. A multi-scale Inception-TCN and a lightweight Transformer encoder are then employed to model local temporal patterns and long-range dependencies, respectively. We further introduce SR-BandMix, a joint time-frequency augmentation strategy that integrates same-class Segmentation-Reconstruction with fine-grained spectral mixing over multiple sub-bands within the MI-related 8-30 Hz range. Experiments on two public benchmarks and one private dataset show that SincDualFormer achieves average accuracies of 82.14%, 87.93%, and 58.50% on BCI Competition IV-2a, BCI Competition IV-2b, and the HCMIU MI Hand-Binary dataset, respectively. It achieves the highest average accuracy among the evaluated methods on all three datasets under their corresponding evaluation protocols. Across representative backbone architectures, SR-BandMix generally improves or maintains competitive performance relative to Segmentation-Reconstruction, BandMix, and no augmentation. In particular, SincDualFormer consistently achieves its best performance with SR-BandMix on both public benchmarks. Ablation, bias-variance, and visualization analyses further demonstrate the complementary contributions of rhythm-constrained filtering, data-driven temporal modeling, and joint time-frequency augmentation to robust MI-EEG decoding.