dual-stream eeg decoding

This study addresses the simultaneous neural decoding of three-dimensional object identity and spatial orientation from electroencephalography (EEG) signals to enable high-fidelity 3D reconstruction based on brain activity. Inspired by the ventral and dorsal pathways of the visual system, the authors propose a dual-stream brain decoding model that separately processes identity and orientation information. By integrating circular regression with an EEG-conditioned multi-view diffusion model, the framework achieves dynamic decoding under continuous rotational conditions. This work represents the first application of a biologically inspired dual-stream architecture to EEG-based decoding, revealing the temporal coordination among ventral, dorsal, and motor-related brain regions during 3D perception. It challenges the conventional assumption of static ventral dominance and substantially advances the performance of neural decoding for 3D visual representation.

This work addresses the challenges of signal heterogeneity and task interference in multi-task electroencephalography (EEG) analysis, which hinder effective sharing of a single pre-trained model across diverse tasks. To overcome these limitations, the authors propose MTEEG, a unified framework for multi-task EEG analysis that leverages a self-supervised pre-trained EEG backbone augmented with task-specific Low-Rank Adaptation (LoRA) modules. This design decouples task-specific parameter spaces while maintaining a shared encoder, thereby mitigating inter-task interference. Comprehensive experiments across six downstream tasks demonstrate that MTEEG consistently outperforms state-of-the-art single-task methods on most evaluation metrics, substantiating the efficacy and potential of joint multi-task optimization in EEG representation learning.

Existing two-stream networks for EEG decoding typically process spatial and temporal features independently, fusing them only at later stages, which limits their ability to capture deep couplings between these modalities. To address this, this work proposes an inter-layer interactive two-stream network that enables progressive, dynamic fusion of spatial and temporal features at every layer through a Temporal-Spatial Integrated Attention (TSIA) mechanism. The TSIA leverages a spatial affinity correlation matrix and a cosine-gated temporal channel aggregation matrix for guided interaction, complemented by an adaptive fusion strategy with learnable channel weights. Evaluated across eight EEG datasets, the proposed method significantly outperforms thirteen state-of-the-art models, demonstrating superior decoding accuracy and robustness in motor imagery, emotion recognition, and steady-state visual evoked potential tasks.

Existing approaches struggle to simultaneously decode multiple cognitive processes—such as event perception, response preparation, and vigilance—from continuous electroencephalography (EEG) signals in human–machine collaboration scenarios. This work proposes DS-MTNet, a novel framework that, for the first time, unifies three types of EEG evidence into a structured slot-filling representation. It integrates EEG waveforms, time–frequency power, and task-routing source embeddings through a multi-stream neural architecture, enhanced by a dual-gating mechanism to enable joint multi-task decoding. Evaluated on a sustained-attention driving dataset, DS-MTNet significantly outperforms both single-task and multi-task baselines, with the most pronounced gains observed during steering-response phases. The method achieves reusable, high-precision structured decoding of brain activity, advancing the feasibility of real-time cognitive state inference in collaborative settings.

This work identifies a pervasive data leakage issue in current cross-subject brain-to-text decoding research (using fMRI/EEG): prevailing data splitting strategies fail to enforce strict subject-level isolation, allowing test-subject information to contaminate the training set—thereby inflating performance estimates and compromising generalization assessment. To address this, we propose the first subject-level strictly isolated data splitting protocol specifically designed for brain-to-text decoding, along with a unified multimodal splitting framework. Leveraging this framework, we rigorously re-evaluate state-of-the-art BERT-based decoding models across multiple public datasets, demonstrating that their reported cross-subject generalization capabilities are systematically overestimated. Our work eliminates evaluation bias, establishes a trustworthy cross-subject benchmark, and provides the field with a methodological standard and reproducible evaluation protocol for fair and reliable model assessment.

This study addresses the challenge of simultaneously decoding multiple kinematic and kinetic parameters during grasping and lifting tasks from non-invasive electroencephalography (EEG) signals to enhance the control dimensionality and practicality of brain–computer interfaces (BCIs). To this end, three regression models—partial least squares regression, multilayer perceptron, and a novel attention-based regressor—are proposed and comparatively evaluated for single-model, multi-parameter decoding under both subject-dependent and subject-independent conditions. Notably, this work introduces an attention mechanism into this decoding task for the first time, achieving a decoding accuracy of R² = 0.8 with a low latency of 29.2 ms—significantly outperforming baseline methods—and thereby demonstrating its strong potential for real-time, multi-command BCIs.

This study addresses the challenge of cross-subject generalization in electroencephalography (EEG) decoding, which is hindered by high inter-subject variability. The problem is formalized as a multi-source domain generalization task, and a rigorous subject-independent evaluation protocol is introduced. The work establishes the first systematic taxonomy of deep learning approaches for cross-subject EEG decoding, categorizing existing methods into four paradigms: feature alignment, adversarial learning, feature disentanglement, and contrastive learning. Furthermore, it clarifies key limitations of current approaches concerning theoretical foundations, utilization of subject identity information, and underexplored potential of foundational models. The paper concludes with a forward-looking discussion on pathways toward EEG foundation models, offering a clear roadmap for developing robust and practical EEG decoding systems.

This work addresses the limitations of conventional EEG classification models that rely on separate one-dimensional spatial and temporal convolutions, which struggle to efficiently capture joint spatiotemporal features and suffer from poor training efficiency in high-dimensional tasks. To overcome this, the authors propose a lightweight two-dimensional spatiotemporal convolutional architecture and systematically investigate the representational differences between 1D and 2D convolutions. Evaluated on a 22-channel motor imagery EEG task, the proposed method achieves comparable classification accuracy while significantly improving both training and inference efficiency. Through representational similarity analysis, the study further reveals—for the first time—that 1D and 2D models yield markedly distinct internal representational geometries, underscoring the critical role of architectural design in encoding multivariate neural signals.

This study addresses the poor generalization of electroencephalography (EEG) decoding in cross-subject and cross-task settings by proposing a zero-shot cross-subject EEG decoding framework. It introduces, for the first time, a Transformer-based foundation model to EEG regression tasks and incorporates a progressive unfreezing fine-tuning strategy that effectively mitigates catastrophic forgetting without requiring calibration data from target subjects. Experimental results on the large-scale Healthy Brain Network dataset demonstrate that the proposed approach significantly outperforms CNN and LSTM baselines. After fine-tuning, the Transformer achieves a normalized root mean square error (nRMSE) of 0.9799 on unseen subjects, markedly improving upon the baseline performance of 0.9991, thereby advancing scalable, calibration-free EEG decoding.

This work addresses the challenge of learning transferable representations from electroencephalography (EEG) signals, which exhibit complex multi-channel coupling and non-stationarity. To this end, the authors propose TRACE, a framework that leverages autoregressive prediction of future EEG segments and introduces a causal cross-channel dynamic expert routing mechanism at each time step. This design enables time-adaptive activation of heterogeneous computational pathways while preserving instantaneous channel consistency. By integrating autoregressive pretraining, causal context modeling, and cross-channel collaborative computation, TRACE supports unsupervised pretraining on unlabeled, heterogeneous EEG data. Evaluated across eight downstream EEG benchmark tasks, the method achieves state-of-the-art or competitive performance, demonstrating particular strength in out-of-distribution scenarios and when relying solely on unlabeled pretraining data.