<p>Attention-based neural networks have advanced EEG decoding for motor imagery (MI) in Brain–Computer Interfaces (BCIs) by capturing global temporal dependencies, whereas convolutional neural networks (CNNs) are particularly effective in capturing localized spatio–temporal patterns. Combining both enables an effective extraction of local and global features. However, fixed kernel sizes in traditional CNNs limit feature diversity and accuracy. To address this, we propose the Multi-Scale Convolutional Self-Attention Network (MCSANet), which integrates multi-scale CNNs with attention mechanisms and introduces the Temporal Segment Shuffling (TSS) data augmentation strategy to enhance model performance further. Trained on BCIC-IV-2a and BCIC-IV-2b datasets in both subject-dependent and subject-independent setups, MCSANet achieves <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$78.11\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>78.11</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> accuracy (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\kappa = 0.72$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>κ</mi> <mo>=</mo> <mn>0.72</mn> </mrow> </math></EquationSource> </InlineEquation>) on BCIC-IV-2a and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$85.56\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>85.56</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> accuracy (<InlineEquation ID="IEq4"> <EquationSource Format="TEX">$\kappa = 0.72$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>κ</mi> <mo>=</mo> <mn>0.72</mn> </mrow> </math></EquationSource> </InlineEquation>) on BCIC-IV-2b, outperforming state-of-the-art methods. TSS further boosts accuracy by <InlineEquation ID="IEq5"> <EquationSource Format="TEX">$5.89\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>5.89</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> and <InlineEquation ID="IEq6"> <EquationSource Format="TEX">$2.8\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>2.8</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> on the respective datasets. These results highlight the potential of MCSANet and TSS for more accurate and robust real-world BCI applications.</p>

MCSANet: A multiscale convolutional self attention network with data augmentation for motor imagery EEG classification

期刊论文
发表日期：

2026年3月16日

查看全文

Seema Devi,
Kamal Singh,
Nitin Singh Singha,
Mahesh K. Singh

摘要

Attention-based neural networks have advanced EEG decoding for motor imagery (MI) in Brain–Computer Interfaces (BCIs) by capturing global temporal dependencies, whereas convolutional neural networks (CNNs) are particularly effective in capturing localized spatio–temporal patterns. Combining both enables an effective extraction of local and global features. However, fixed kernel sizes in traditional CNNs limit feature diversity and accuracy. To address this, we propose the Multi-Scale Convolutional Self-Attention Network (MCSANet), which integrates multi-scale CNNs with attention mechanisms and introduces the Temporal Segment Shuffling (TSS) data augmentation strategy to enhance model performance further. Trained on BCIC-IV-2a and BCIC-IV-2b datasets in both subject-dependent and subject-independent setups, MCSANet achieves $78.11\%$ $78.11 %$ accuracy ( $\kappa = 0.72$ $κ = 0.72$ ) on BCIC-IV-2a and $85.56\%$ $85.56 %$ accuracy ( $\kappa = 0.72$ $κ = 0.72$ ) on BCIC-IV-2b, outperforming state-of-the-art methods. TSS further boosts accuracy by $5.89\%$ $5.89 %$ and $2.8\%$ $2.8 %$ on the respective datasets. These results highlight the potential of MCSANet and TSS for more accurate and robust real-world BCI applications.