LI-DSN: A Layer-wise Interactive Dual-Stream Network for EEG Decoding

Chenghao Yue; Zhiyuan Ma; Zhongye Xia; Xinche Zhang; Yisi Zhang; Xinke Shen; Sen Song

cs.LGApr 2, 2026

LI-DSN: A Layer-wise Interactive Dual-Stream Network for EEG Decoding

Authors: Chenghao Yue, Zhiyuan Ma, Zhongye Xia, Xinche Zhang, Yisi Zhang, Xinke Shen, Sen Song

Organizations: School of Life Sciences, Tsinghua University, Beijing 100084, China · School of Biomedical Engineering, Tsinghua Laboratory of Brain and Intelligence, and IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing 100084, China · Department of Biomedical Engineering, Southern University of Science and Technology, Shenzhen 518055, China · Department of Psychological and Cognitive Sciences, Tsinghua University, Beijing, 100084, China

Abstract

Electroencephalography (EEG) provides a non-invasive window into brain activity, offering high temporal resolution crucial for understanding and interacting with neural processes through brain-computer interfaces (BCIs). Current dual-stream neural networks for EEG often process temporal and spatial features independently through parallel branches, delaying their integration until a final, late-stage fusion. This design inherently leads to an "information silo" problem, precluding intermediate cross-stream refinement and hindering spatial-temporal decompositions essential for full feature utilization. We propose LI-DSN, a layer-wise interactive dual-stream network that facilitates progressive, cross-stream communication at each layer, thereby overcoming the limitations of late-fusion paradigms. LI-DSN introduces a novel Temporal-Spatial Integration Attention (TSIA) mechanism, which constructs a Spatial Affinity Correlation Matrix (SACM) to capture inter-electrode spatial structural relationships and a Temporal Channel Aggregation Matrix (TCAM) to integrate cosine-gated temporal dynamics under spatial guidance. Furthermore, we employ an adaptive fusion strategy with learnable channel weights to optimize the integration of dual-stream features. Extensive experiments across eight diverse EEG datasets, encompassing motor imagery (MI) classification, emotion recognition, and steady-state visual evoked potentials (SSVEP), consistently demonstrate that LI-DSN significantly outperforms 13 state-of-the-art (SOTA) baseline models, showcasing its superior robustness and decoding performance. The code will be publicized after acceptance.

Figures & tables

Fig. 1: Exploration of dual-stream interaction strategies for EEG analysis, where $F_{t}$ and $F_{s}$ raw temporal and spatial features respectively, $\tilde{F^{\prime}_{t}}$ and $\tilde{F^{\prime}_{s}}$ denote enhanced features after interaction. (a) Dual-stream architecture with cross-flow interactions between temporal and spatial streams. (b) Baseline with no cross-stream interaction before fusion. (c) Interaction outputs integrated only into the spatial stream. (d) Interaction outputs integrated only into the temporal stream. (e) Bidirectional interaction feeding both streams.

Paradigm	Dataset	Number of Subjects	Number of EEG Channels	Sampling Rate (Hz)	Trial Length (seconds)	Number of Total Trials	Task Types
MI	BNCI2014001	9	22	250	4	1,296	left/right hand
	BNCI2014002	14	15	512	5	1,400	right hand/both feet
	BNCI2014004	9	3	250	5	1,080	left/right hand
	Zhou2016	4	14	250	5	409	left/right hand
Emotion Recognition	SEED	15	62	200	2	7,920	negative/positive
Emotion Recognition	FACED	123	32	250	4	10,332	negative/positive

TABLE I: Summary of datasets and task types across different paradigms.

Fig. 2: The overall architecture of the proposed LI-DSN. The raw EEG signals are first processed by parallel Temporal and Spatial Tokenizers to extract initial embeddings. These embeddings are then fed into $N$ blocks, where each block consists of a Feed-Forward Network (FFN) followed by the Temporal-Spatial Integration Attention (TSIA) module to enable progressive cross-stream interaction. The refined features are subsequently aggregated by the Adaptive Fusion module and passed through an MLP classification head to generate outputs for various EEG tasks (e.g., MI classification, Emotion Recognition, SSVEP). The detailed architectures of the five key components, namely the Spatial Tokenizer, Temporal Tokenizer, FFN Block, TSIA and Adaptive Fusion, are illustrated in the sub-panels.

Task	Dataset	SCNN	DCNN	EEGNet	FBCNet	CTNet	MSVTNet	IFNet	Conformer	DBConformer	LI-DSN
CO	BNCI2014001	73.57 $\pm$ 2.37	59.29 $\pm$ 1.64	69.05 $\pm$ 1.02	68.97 $\pm$ 1.28	73.49 $\pm$ 2.12	74.60 $\pm$ 3.03	77.94 $\pm$ 0.87	78.57 $\pm$ 0.72	80.16 $\pm$ 1.83	80.95 $\pm$ 1.09
	BNCI2014002	79.07 $\pm$ 2.01	64.07 $\pm$ 2.68	66.07 $\pm$ 2.76	69.50 $\pm$ 1.02	71.00 $\pm$ 1.78	78.93 $\pm$ 1.42	78.29 $\pm$ 1.73	76.21 $\pm$ 1.52	79.14 $\pm$ 0.16	81.57 $\pm$ 1.33
	BNCI2014004	68.15 $\pm$ 1.42	62.41 $\pm$ 1.58	68.43 $\pm$ 2.21	65.46 $\pm$ 1.56	71.57 $\pm$ 1.32	71.30 $\pm$ 1.50	73.43 $\pm$ 1.71	72.04 $\pm$ 1.61	75.09 $\pm$ 1.54	75.19 $\pm$ 0.23
	Zhou2016	75.03 $\pm$ 6.18	78.03 $\pm$ 2.41	80.13 $\pm$ 3.35	63.33 $\pm$ 2.31	76.81 $\pm$ 5.14	70.32 $\pm$ 1.42	81.70 $\pm$ 2.14	73.87 $\pm$ 4.45	82.49 $\pm$ 2.42	84.65 $\pm$ 1.87
CV	BNCI2014001	78.22 $\pm$ 0.21	61.59 $\pm$ 1.12	71.86 $\pm$ 1.03	74.74 $\pm$ 0.72	75.97 $\pm$ 0.77	81.28 $\pm$ 0.72	82.62 $\pm$ 0.92	83.62 $\pm$ 1.01	83.66 $\pm$ 0.42	83.79 $\pm$ 0.69
	BNCI2014002	81.24 $\pm$ 0.52	67.17 $\pm$ 0.32	68.90 $\pm$ 1.37	74.53 $\pm$ 0.72	75.50 $\pm$ 0.93	80.91 $\pm$ 0.52	80.21 $\pm$ 0.52	79.30 $\pm$ 0.31	81.36 $\pm$ 0.33	82.57 $\pm$ 0.44

TABLE II: MI classification accuracies (%) of different methods on the BNCI2014001, BNCI2014002, BNCI2014004, and Zhou2016 datasets under CO, CV, and LOSO scenarios. ( Bold : best; Underline : second best)

Method	SEED		FACED
Method	ACC(±std)↑	F1(±std)↑	ACC(±std)↑	F1(±std)↑
GCBNet	$68.42\pm 17.21$	$52.15\pm 35.68$	$57.38\pm 5.23$	$69.12\pm 5.36$
EEGNet	$69.55\pm 11.28$	$65.84\pm 22.08$	$62.60\pm 4.35$	$76.45\pm 3.31$
TSception	$66.23\pm 18.12$	$62.05\pm 28.26$	$61.92\pm 8.83$	$70.19\pm 3.58$
CTNet	$68.71\pm 11.62$	$63.75\pm 25.41$	$62.08\pm 4.20$	$76.10\pm 3.17$
MSVTNet	$64.36\pm 8.21$	$59.86\pm 21.78$	$62.91\pm 6.17$	$76.79\pm 4.84$

TABLE III: Emotion recognition results of different methods on the SEED and FACED datasets. ( Bold : best; Underline : second best)

Method	Nakanishi2015		Kalunga2016
Method	ACC(±std)↑	F1(±std)↑	ACC(±std)↑	F1(±std)↑
EEGNet	$96.09\pm 0.22$	$95.94\pm 0.22$	$69.73\pm 0.68$	$67.87\pm 0.94$
CTNet	$96.19\pm 0.11$	$96.06\pm 0.14$	$69.18\pm 0.45$	$67.05\pm 0.30$
MSVTNet	$90.46\pm 0.27$	$90.37\pm 0.32$	$70.84\pm 0.49$	$68.35\pm 0.54$
Conformer	$94.16\pm 0.34$	$94.16\pm 0.34$	$65.17\pm 0.42$	$63.10\pm 0.52$
Deformer	$\underline{96.26\pm 0.33}$	$\underline{96.19\pm 0.34}$	$70.01\pm 0.35$	$\underline{68.97\pm 0.51}$

TABLE IV: SSVEP results of different methods on the Nakanishi2015 and Kalunga2016 datasets. ( Bold : best; Underline : second best)

Model Variant	2014001	2014002	Zhou2016
w/o Positional Embedding	74.79±1.18	74.67±0.14	84.87±1.19
w/o Interaction (Late Fusion)	73.25±1.32	72.05±0.15	81.51±1.38
w/o TSIA (Simple Concat)	72.37±1.72	70.56±0.23	80.31±1.65
w/o Cosine Gate	75.22±1.15	75.29±0.08	84.01±1.21
w/o Ele-Pos Embedding	75.65±1.20	75.34±0.12	83.92±1.26
w/o Adaptive Fusion	73.69±1.45	73.90±0.14	82.14±1.41

TABLE V: The ablation study of the proposed module and strategy on three MI classification datasets.

Fig. 3: Comparison of integration method under the LOSO protocol across four MI classification datasets.

Fig. 4: Impact of architectural hyperparameters on performance under the LOSO protocol. (a) and (c) illustrate the effect of block depth $(N_{s},N_{t})$ on the SEED and FACED datasets respectively. (b) and (d) display the effect of feature embedding width on the SEED and FACED datasets.

Fig. 5: Comparison of model performance, parameters, and MFLOPS on BNCI2014001, with model size indicated by circle area.

Fig. 6: Visualization of the TSIA matrices and channel-temporal interactions. Subfigures (a), (b), (e), and (f) correspond to the Nakanishi2015 dataset, while (c), (d), (g), and (h) correspond to the Kalunga2016 dataset. Specifically, (a) and (c) illustrate the spatial attention heatmap of SACM; (b) and (d) display the spatial attention heatmap of TCAM. Regarding the feature interaction, (e) and (g) depict the channel-temporal feature matrices before TSIA interaction, whereas (f) and (h) visualize the matrices after interaction.

Fig. 7: Visualization of saliency maps across three representative datasets. Subfigures (a)-(d) correspond to BNCI2014001, (e)-(h) represent the Kalunga2016 dataset, and (i)-(l) depict the SEED dataset. Within each group, the first three subfigures display activation patterns of representative subjects to illustrate individual variability, while the last subfigure presents the mean saliency map averaged across all subjects in the respective dataset.

EEG Decoding

Electroencephalography (EEG) decoding maps brain-signal patterns to mental states, stimuli, speech, or intended actions.

See all 208 papers

DSAINet: An Efficient Dual-Scale Attentive Interaction Network for General EEG Decoding
Apr 20, 2026Zhiyuan Ma, Zeyuan Li, Zihao Qiu +6Cross-Subject EEG Decoding EEG Decoding
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
T-SNN: Temporal Simplicial Neural Network for EEG Decoding
Sep 27, 2026Nikita Malik, Shubhajit Roy, Mohit Kataria +4Cross-Subject EEG Decoding EEG Decoding
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 +4
EEG Decoding Using CNN and LSTM Network
Aug 13, 2026Athanasios KaragounisCross-Subject EEG Decoding MI Classification
School of Science, Department of Digital Industry Technologies GR34400, Psachna, Greece

Apr 20, 2026J/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

Cross-Subject EEG Decoding EEG Decoding Attention Mechanisms Multiscale Time-Series Modeling

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

EEG Decoding

Explore similar work

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

T-SNN: Temporal Simplicial Neural Network for EEG Decoding

EEG Decoding Using CNN and LSTM Network