TSMNet: A comprehensive network based on spatio-temporal representations for SSVEP classification

• Published in: Biomedical signal processing and control Vol. 105; p. 107554
• Main Authors: Deng, Liu, Li, Pengrui, Zhang, Haokai, Zheng, Qingyuan, Liu, Shihong, Ding, Xinmin, Wang, Manqing, Gao, Dongrui
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.07.2025
• Subjects: Electroencephalogram (EEG); Spatial topology converter; Steady-state visual evoked potentials (SSVEP); Temporal feature extractor; TSMNet; Temporal feature extractor; TSMNet; Electroencephalogram (EEG); Spatial topology converter; Steady-state visual evoked potentials (SSVEP)
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2025.107554

Steady-state visual evoked potentials (SSVEP) contributes to the brain’s interaction with external devices due to its advantages of high signal-to-noise ratio and information transmission rate. From the shape of the electroencephalogram (EEG) signals shows an irregular distribution, the current SSVEP deep learning methods have not represent the spatial topology patterns of EEG signals well, thus urgent need for further improvements in extracting spatial features. To this end, we propose a comprehensive network called TSMNet, combining a temporal feature extractor and a spatial topology converter for SSVEP classification. Specifically, the temporal feature extractor constructs a Temporal-conv block to filter the temporal-domain noise to capture pure high-order temporal-dependent representations. The spatial topology converter models the multi-channel representations in EEG signals and dynamically updates the shape of the topology, thus effectively capturing high-order spatial topology representations. We further develop a multigraph subspace module for multi-space mapping of the output of the spatial topology converter. This paper evaluates TSMNet on two SSVEP open datasets with time windows of 0.5s and 1.0s and compares with the state-of-the-art methods based on inter-subject and intra-subject experiments. The experimental results show that TSMNet has significant advantages in performance, it achieves an accuracy improvement of 2.04% and 1.47%. Attractively, this end-to-end framework not only eliminates the time-consuming manual feature engineering process but also speeds up the implementation of practical applications. •A novel deep learning target recognition network is proposed.•We introduce the label smoothing technology to solve potential overfitting issues.•Experiments reveal our proposed method can learn efficient discriminative features.