T-SNN: Temporal Simplicial Neural Network for EEG Decoding

Nikita Malik; Shubhajit Roy; Mohit Kataria; Isuru Herath; Suraj Yadav; Inés García-Redondo; Dhananjay Bhaskar

cs.LGSep 27, 2026

T-SNN: Temporal Simplicial Neural Network for EEG Decoding

Authors: Nikita Malik, Shubhajit Roy, Mohit Kataria, Isuru Herath, Suraj Yadav, Inés García-Redondo, Dhananjay Bhaskar

Organizations: Bharti School of Telecommunication Technology and Management Indian Institute of Technology, Delhi · Department of Computer Science & Engineering Indian Institute of Technology, Gandhinagar · Yardi School of Artificial Intelligence Indian Institute of Technology, Delhi · Medical Scientist Training Program School of Medicine University of Pittsburgh · Hertie Institute for AI in Brain Health University of Tübingen · AIDOS Lab University of Fribourg · Department of Biomedical Engineering, Biophysics Graduate Program, Data Science Institute, Center for Genomic Science Innovation, Wisconsin Institute for Translational Neuroengineering University of Wisconsin–Madison

Abstract

Decoding brain states requires models that capture both the evolution of neural activity and interactions among groups of brain regions. Existing EEG methods often treat recordings as multivariate time series or represent functional connectivity with pairwise graphs, leaving dynamic higher-order interactions largely unmodeled. We introduce the Temporal Simplicial Neural Network (T-SNN), which represents EEG recordings as sequences of evolving simplicial complexes. By combining simplicial convolutions with recurrent updates, T-SNN jointly learns higher-order interactions and their temporal evolution. On the seven-class SEED-VII emotion recognition task, T-SNN outperforms convolutional, recurrent, graph-based, and Transformer methods in both trial-wise and cross-subject evaluations. Incorporating eye-movement features further improves performance, demonstrating the framework's potential for multimodal brain-state decoding.

Figures & tables

Figure 1: Overview of T-SNN. EEG windows are lifted into simplicial complexes and processed by blocks that jointly perform spatial and temporal message passing through simplicial convolutions and recurrent updates. The resulting representations are fused with eye-movement features for emotion recognition.

	Method	Delta band	Theta band	Alpha band	Beta band	Gamma band	All bands
Accuracy	KNN Cover and Hart (1967)	$28.21\pm 5.76$	$28.47\pm 7.02$	$31.92\pm 6.51$	$34.83\pm 5.77$	$37.20\pm 6.06$	$36.43\pm 5.38$
	HCNN Li et al. (2018)	$42.33\pm 6.36$	$42.63\pm 5.28$	$43.62\pm 4.82$	$47.79\pm 6.77$	$48.18\pm 6.10$	$52.42\pm 6.47$
	RGNN Zhong et al. (2022)	$42.31\pm 5.14$	$41.73\pm 5.53$	$43.93\pm 4.98$	$44.68\pm 5.51$	$45.49\pm 4.73$	$48.50\pm 6.83$
	Transformer Vaswani et al. (2017)	$46.87\pm 4.10$	$47.01\pm 4.23$	$48.78\pm 5.39$	$53.10\pm 6.34$	$53.96\pm 6.60$	$56.04\pm 7.82$
	GCNCA Jiang et al. (2021)	$49.97\pm 4.94$	$50.46\pm 4.29$	$53.26\pm 5.45$	$58.36\pm 6.69$	$59.50\pm 5.77$	$58.04\pm 7.78$
	MAET Jiang et al. (2023)	–	–	–	–	–	$58.11\pm 8.78$

Table 1: Trial-wise SEED-VII accuracy and F1 score across EEG frequency bands (Avg.

\pm

Std., %)

Figure 2: Group-averaged EEG simplices across seven emotions in the SEED-VII dataset ( $N=20$ ). Lines represent edges (1-simplices), and filled triangles represent three-way interactions (2-simplices). Blue, green, and red indicate prevalence in early, middle, and late trial segments. Mixed colors indicate prevalence across multiple segments.

Method	Accuracy		F1 score
Method	Avg.	Std.	Avg.	Std.
KNN Cover and Hart (1967)	$40.44$	$6.30$	$37.94$	$6.54$
BDAE Liu et al. (2016)	$61.55$	$8.74$	$59.11$	$8.87$
ETF Wang et al. (2021)	$65.30$	$8.55$	$63.13$	$8.88$
VigilanceNet Cheng et al. (2022)	$62.93$	$7.12$	$60.46$	$7.81$
MAET Jiang et al. (2023)	$71.28$	$7.74$	$69.16$	$8.35$

Table 2: Trial-wise SEED-VII accuracy and F1 score using multimodal EEG and eye movement data (Avg./Std., %)

Method	Accuracy		F1 score
Method	Avg.	Std.	Avg.	Std.
KNN Cover and Hart (1967)	$20.85$	$4.56$	$20.23$	$4.49$
HCNN Li et al. (2018)	$39.88$	$4.94$	$38.18$	$5.06$
RGNN Zhong et al. (2022)	$37.49$	$5.44$	$34.52$	$4.83$
Transformer Vaswani et al. (2017)	$40.36$	$5.22$	$37.76$	$5.54$
GCNCA Jiang et al. (2021)	$38.68$	$\mathbf{3.94}$	$37.25$	$\mathbf{3.65}$

Table 3: Leave-one-subject-out SEED-VII accuracy and F1 score (Avg./Std., %)

Explore similar work

DSAINet: An Efficient Dual-Scale Attentive Interaction Network for General EEG Decoding
Apr 20, 2026Zhiyuan Ma, Zeyuan Li, Zihao Qiu +6Electroencephalography Neuroscience
School of Biomedical Engineering, Tsinghua Laboratory of Brain and Intelligence, and IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing 100084, China · School of Basic Medical Sciences and Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing 100084, China · Department of Psychological and Cognitive Sciences, Tsinghua University, Beijing 100084, China +1
Transformer Based Model for Spatiotemporal Feature Learning in EEG Emotion Recognition
Jun 9, 2026Xinglong Cui, Dian GuElectroencephalography Spatiotemporal Representations
Beijing Neurodeep Technology Co., Ltd, Beijing, 100176, China · University of Pennsylvania, Seattle, 98121, USA
Beyond Flattened Tokens: Structure-Preserving EEG Decoding with Reusable TriDim Blocks
Sep 17, 2026Shiyue Su, Song Wang, Zekai Zhan +7Electroencephalography Decoding Electroencephalography
Department of Biomedical Engineering, Southern University of Science and Technology, Shenzhen, 518055, China. · Omni-Intelligence, Shenzhen, China. · Department of Computer Science, The Chinese University of Hong Kong, Shenzhen, 518172, China. +2

Apr 20, 2026, cs.AIJ/K move · Enter open · S save

DSAINet: An Efficient Dual-Scale Attentive Interaction Network for General EEG Decoding

Zhiyuan Ma, Zeyuan Li, Zihao Qiu, Jinhao Li+5

School of Biomedical Engineering, Tsinghua Laboratory of Brain and Intelligence, and IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing 100084, China · School of Basic Medical Sciences and Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing 100084, China · Department of Psychological and Cognitive Sciences, Tsinghua University, Beijing 100084, China+1

Electroencephalography Neuroscience Auditory Attention Multi-Scale Attention Inductive Bias

Open paper AskPDF arXiv

In real-world applications of noninvasive electroencephalography (EEG), specialized decoders often show limited generalizability across diverse tasks under subject-independent settings. One central challenge is that task-relevant EEG signals often follow different temporal organization patterns across tasks, while many existing methods rely on task-tailored architectural designs that introduce task-specific temporal inductive biases. This mismatch makes it difficult to adapt temporal modeling across tasks without changing the model configuration. To address these challenges, we propose DSAINet, an efficient dual-scale attentive interaction network for general EEG decoding. Specifically, DSAINet constructs shared spatiotemporal token representations from raw EEG signals and models diverse temporal dynamics through parallel convolutional branches at fine and coarse scales. The resulting representations are then adaptively refined by intra-branch attention to emphasize salient scale-specific patterns and by inter-branch attention to integrate task-relevant features across scales, followed by adaptive token aggregation to yield a compact representation for prediction. Extensive experiments on five downstream EEG decoding tasks across ten public datasets show that DSAINet consistently outperforms 13 representative baselines under strict subject-independent evaluation. Notably, this performance is achieved using the same architecture hyperparameters across datasets. Moreover, DSAINet achieves a favorable accuracy-efficiency trade-off with only about 77K trainable parameters and provides interpretable neurophysiological insights. The code is publicly available at https://github.com/zy0929/DSAINet.

Abstract

Explore similar work

DSAINet: An Efficient Dual-Scale Attentive Interaction Network for General EEG Decoding

Transformer Based Model for Spatiotemporal Feature Learning in EEG Emotion Recognition

Beyond Flattened Tokens: Structure-Preserving EEG Decoding with Reusable TriDim Blocks