Citation

Roy, Dipon and Khan, Robiul and Sadeque, Md Golam and Ahmed Refat, Kawsar and Sarker, Md Tanjil and Ramasamy, Gobbi (2025) EEG-MHNet: A Hybrid Convolution and Attention Network for Motor Imagery EEG Classification. In: 2025 Multimedia University Engineering Conference, MECON 2025, 21 July 2025 - 23 July 2025, Cyberjaya, Malaysia.

Text 66.pdf - Published Version Restricted to Repository staff only Download (1MB)

Abstract

Effective motor imagery EEG classification requires extracting both fine-grained local patterns and global temporal dependencies from non-stationary neural signals. However, single-scale CNNs are limited in their ability to capture a broad range of spectral features from EEG signals, and conventional multi-scale CNNs often struggle to effectively integrate multi-scale information when relying solely on concatenation-based fusion methods. To overcome these limitations, we propose EEG-MHNet, a hybrid architecture that incorporates a dual-branch convolutional network to encapsulate local and global features, for a long range. The dualbranch structure captures multi-scale temporal features using distinct kernel sizes, while depthwise-separable convolutions enhance spatial resolution with reduced computational overhead. Pointwise convolutions are further employed to project inter-channel relationships, enabling effective feature fusion efficiently. A self-attention mechanism is then employed to capture global dependencies from the locally enriched temporal features. Finally, a lightweight fully connected classifier is utilized to predict the corresponding motor imagery classes from the encoded representations. The proposed scheme achieves an average accuracy of 72.70% on the BCI Competition IV 2a dataset, demonstrating that EEG-MHNet consistently outperforms baseline models. These results underscore the importance of advanced convolutional strategies and attention-based modeling for reliable and efficient EEG decoding in BCI applications.

Item Type:	Conference or Workshop Item (Paper)
Uncontrolled Keywords:	Attention mechanism, motor imagery
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5101-6720 Telecommunication. Including telegraphy, telephone, radio, radar, television
Divisions:	Faculty of Artificial Intelligence & Engineering (FAIE)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	18 Mar 2026 07:06
Last Modified:	18 Mar 2026 07:06
URII:	http://shdl.mmu.edu.my/id/eprint/15555

Downloads

Edit (login required)