ECSA-EEGNet: A Lightweight EEG Decoding Model with Efficient Channel-SpatialAttention and Multi-Scale Fusion
Yang Yabing1,2, Dong Zheng1,2, Bao Xueliang1,2#*, Zhang Peng1,2*
1(School of Information Engineering, Ningxia University, Yinchuan 750021, China) 2(Key Laboratory of Artificial Intelligence and Information Security for Channeling Computing Resources from the East to the West of Ningxia, Yinchuan 750021, China)
Abstract：To address challenges of low spatial-channel feature fusion efficiency, high computational complexity, and weak cross-subject generalization in existing motor imagery (MI) electroencephalogram (EEG) decoding models, this study aimed to construct a novel decoding architecture that balances high accuracy with lightweight design for resource-constrained embedded applications. We proposed the ECSA-EEGNet, a lightweight model incorporating three key designs. First, a dynamic channel-space attention module was designed, utilizing parallel dual-pooling and adaptive weighting mechanisms to replace traditional fully connected dimensionality reduction, thereby enhancing the perception of key brain region features without information loss. Second, the attention sub-modules were reconstructed by adopting a fully convolutional structure to reduce redundant parameters (approximately 40% reduction compared to traditional mechanisms) and introducing a convolution-batch normalization-mish (CBM) unit to strengthen non-linear modeling capabilities. Finally, a multi-scale channel-space fusion framework was constructed to improve robustness against individual differences through cross-dimensional feature interaction. The model was validated on the BCI Competition IV-2a (4-class) and IV-2b (2-class) datasets involving nine healthy subjects. Results indicated that ECSA-EEGNet achieved an average accuracy of 80.29% on the IV-2a dataset (a 7.89% improvement over the baseline EEGNet) with a Kappa coefficient of 0.72. On the IV-2b dataset, it reached an accuracy of 85.22% (a 9.50% improvement) with a Kappa coefficient of 0.70. The model comprises only 5,022 parameters, approximately one-third of those in TCNet-Fusion. In fine-grained evaluation, Subject 3 in the IV-2a dataset achieved recognition rates exceeding 90% for all four tasks (left hand, right hand, foot, and tongue), with a peak accuracy of 95.83%, confirming the model's stability in complex tasks. These findings demonstrated that ECSA-EEGNet achieved a significant leap in decoding performance with minimal parameter overhead, effectively balancing accuracy and efficiency, thus providing a viable lightweight solution for the clinical deployment of real-time brain-computer interface systems.
杨亚兵, 董政, 鲍学亮, 张鹏. ECSA-EEGNet:基于高效通道空间注意力与多尺度融合的轻量化脑电解码模型[J]. 中国生物医学工程学报, 2026, 45(2): 178-187.
Yang Yabing, Dong Zheng, Bao Xueliang, Zhang Peng. ECSA-EEGNet: A Lightweight EEG Decoding Model with Efficient Channel-SpatialAttention and Multi-Scale Fusion. Chinese Journal of Biomedical Engineering, 2026, 45(2): 178-187.
[1] Nishiura R, Tsuru K. Development of BCI application for the physically challenged using EEG measured with multi-pin electrodes[C]//2024 IEEE International Technical Conference on Circuits/Systems, Computers, and Communications (ITC-CSCC). Hong Kong: IEEE, 2024: 1-6. [2] Kutlu İ Ç, Tashan W, Shayea I, et al. An introductory guide on creating a Pandas-based EEG analysis and action prediction tool for BCI systems[C]//2024 IEEE 13th International Conference on Communication Systems and Network Technologies (CSNT). Dubai: IEEE, 2024: 1372-1378. [3] Akuthota S, Rajkumar K, Janapati R. Svelte EEG Net: Cross-session motor imagery classification for BCI-controlled prosthetic arms using novel compact CNN[C]//2024 IEEE 3rd World Conference on Applied Intelligence and Computing (AIC). Raipur: IEEE, 2024: 1285-1289. [4] Naser MYM, Bhattacharya S. Towards practical BCI-driven wheelchairs: A systematic review study[J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2023, 31: 1030-1044. [5] Wolpaw JR, McFarland DJ, Neat GW, et al. An EEG-based brain-computer interface for cursor control[J]. Electroencephalography and Clinical Neurophysiology, 1991, 78(3): 252-259. [6] Carlson T, Millán JDR. Brain-controlled wheelchairs: A robotic architecture[J]. IEEE Robotics & Automation Magazine, 2013, 20(1): 65-73. [7] Koles ZJ. The quantitative extraction and topographic mapping of the abnormal components in the clinical EEG[J]. Electroencephalography and Clinical Neurophysiology, 1991, 79(6): 440-447. [8] Müller-Gerking J, Pfurtscheller G, Flyvbjerg H. Designing optimal spatial filters for single-trial EEG classification in a movement task[J]. Clinical Neurophysiology, 1999, 110(5): 787-798. [9] Ang Kaikeng, Chin Zhengyang, Zhang Haihong, et al. Filter bank common spatial pattern (FBCSP) in brain-computer interface[C]//2008 IEEE International Joint Conference on Neural Networks (IEEE World Congress on Computational Intelligence). Hong Kong: IEEE, 2008: 2390-2397. [10] Schirrmeister RT, Springenberg JT, Fiederer LDJ, et al. Deep learning with convolutional neural networks for EEG decoding and visualization[J]. Human Brain Mapping, 2017, 38(11): 5391-5420. [11] Lawhern VJ, Solon AJ, Waytowich NR, et al. EEGNet: A compact convolutional neural network for EEG-based brain-computer interfaces[J]. Journal of Neural Engineering, 2018, 15(5): 056013. [12] 李红利,刘浩雨,陈虹宇,等。基于时间序列数据增强的运动想象脑电多尺度特征提取分类 [J]. 生物医学工程学杂志,2023, 40(3): 418-425. [13] Ingolfsson TM, Hersche M, Wang X, et al. EEG-TCNet: An accurate temporal convolutional network for embedded motor-imagery brain-machine interfaces[C]//2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC). Toronto: IEEE, 2020: 2958-2965. [14] Altaheri H, Muhammad G, Alsulaiman M. Physics-informed attention temporal convolutional network for EEG-based motor imagery classification[J]. IEEE Transactions on Industrial Informatics, 2022, 19(2): 2249-2258. [15] Zhao Ruoqi, Wang Yuwen, Cheng Xiangxin, et al. A multi-scale spatial-temporal convolutional neural network with contrastive learning for motor imagery EEG classification[J]. Medicine in Novel Technology and Devices, 2023, 17: 100215. [16] Woo S, Park J, Lee JY, et al. CBAM: convolutional block attention module[C]//Proceedings of the European Conference on Computer Vision (ECCV). Cham: Springer, 2018: 3-19. [17] Wang Yihan, Wang Jiaxing, Wang Weiqun, et al. TFTL: a task-free transfer learning strategy for EEG-based cross-subject and cross-dataset motor imagery BCI[J]. IEEE Transactions on Biomedical Engineering, 2024, 72(2): 810-821. [18] Liu Chang, You Jia, Wang Kun, et al. Decoding the EEG patterns induced by sequential finger movement for brain-computer interfaces[J]. Frontiers in Neuroscience, 2023, 17: 1180471. [19] Xiong Shiming, Wang Li, Xia Guoxian, et al. MBCNN-EATCFNet: a multi-branch neural network with efficient attention mechanism for decoding EEG-based motor imagery[J]. Robotics and Autonomous Systems, 2024: 104899. [20] Zhu Lei, Wang Yunsheng, Huang Aiai, et al. A multi-branch, multi-scale, and multi-view CNN with lightweight temporal attention mechanism for EEG-based motor imagery decoding[J]. Computer Methods in Biomechanics and Biomedical Engineering, 2024,27: 1-15. [21] Li Linglong, Cao Guangzhong, Zhang Yuepeng, et al. MACNet: A multidimensional attention-based convolutional neural network for lower-limb motor imagery classification[J]. Sensors, 2024, 24(23): 7611. [22] Deng Haodong, Li Mengfan, Li Jundi, et al. A robust multi-branch multi-attention-mechanism EEGNet for motor imagery BCI decoding[J]. Journal of Neuroscience Methods, 2024, 405: 110108. [23] Wang Qilong, Wu Banggu, Zhu Pengfei, et al. Efficient channel attention for deep convolutional neural networks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2020: 11534-11542. [24] Bashivan P, Rish I, Yeasin M, et al. Learning representations from EEG with deep recurrent-convolutional neural networks[C]//2016 International Conference on Learning Representations (ICLR). San Juan: ICLR, 2016. [25] Musallam YK, AlFassam NI, Muhammad G, et al. Electroencephalography-based motor imagery classification using temporal convolutional network fusion[J]. Biomedical Signal Processing and Control, 2021, 69: 102826. [26] Huang Qiangfeng, Yang Yuanpo, Li Jun, et al. A multi-level attention-based multi-scale fusion temporal convolutional network for decoding MI-EEG signals[J]. Biomedical Signal Processing and Control, 2025, 108: 107916.
