Unified Temporal-Spectral-Spatial Modeling for Robust and Generalizable Motor Imagery Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI)-based brain-computer interfaces (BCIs) show promise for neurorehabilitation and assistive technologies.
• Decoding electroencephalographic (EEG) signals is challenging due to low signal-to-noise ratio and complex neural dynamics.
• Existing deep learning models often focus on single data representations or require extensive calibration.

Purpose of the Study:

• To develop a unified deep learning model for robust and calibration-free MI decoding from EEG signals.
• To jointly learn temporal, spectral, and spatial features for improved EEG signal analysis.
• To enhance cross-subject generalization in BCIs without requiring labeled target-domain data.

Main Methods:

• Introduced NeuroCrossNet, a tri-modal deep learning architecture integrating Temporal HyperMixer, wavelet transformer, and Graph Attention Network.
• Developed Dynamic Residual Attention Gate (DRAG) for adaptive feature stream merging.
• Implemented subject-aware normalization (SAN) for calibration-free cross-subject generalization.

Main Results:

• Achieved a classification accuracy of 91.30% on BCI Competition IV-2a and High-Gamma datasets using a leave-one-subject-out approach.
• Outperformed several state-of-the-art methods including CNN-LSTM, EEGNet, and DeepConvNet.
• Ablation studies confirmed that integrating complementary temporal, spectral, and spatial representations significantly improves robustness and inter-subject consistency.

Conclusions:

• NeuroCrossNet offers a powerful, unified approach for accurate and efficient MI decoding from EEG.
• The model's ability to learn from multiple feature domains and generalize across subjects marks a significant advancement in BCI technology.
• This work paves the way for more reliable and accessible neurorehabilitation and human-computer interaction applications.