AMANet: a data-augmented multi-scale temporal attention convolutional network for motor imagery classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery brain-computer interfaces (MI-BCI) show potential for neural plasticity but face challenges with limited data and signal noise.
• High decoding performance in MI-BCI is hindered by insufficient subject-specific samples and susceptibility to artifacts in EEG signals.

Purpose of the Study:

• To propose a novel deep learning network, AMANet, for enhanced motor imagery decoding.
• To address the challenges of limited data and noise in MI-EEG signal processing.
• To improve the accuracy and robustness of brain-computer interfaces.

Main Methods:

• A Data-Augmented Multi-Scale Temporal Attention Convolutional Network (AMANet) was developed, incorporating data augmentation (sliding-window, CSP, linear scaling), multi-scale temporal convolution, ECA attention, and depthwise separable convolution.
• The network integrates spatial and temporal feature extraction with adaptive channel weighting for robust MI-EEG signal classification.
• 10-fold cross-validation was employed on benchmark datasets (BCI Competition IV Datasets 2a and 2b) and the High-Gamma dataset.

Main Results:

• AMANet achieved classification accuracies of 84.06% and 85.09% on BCI Competition IV Datasets 2a and 2b, respectively.
• On the High-Gamma dataset, AMANet attained a classification accuracy of 95.48%.
• The proposed AMANet significantly outperformed baseline models like Incep-EEGNet.

Conclusions:

• The AMANet demonstrates superior performance in motor imagery decoding tasks, effectively overcoming limitations of existing MI-BCI methods.
• The network's architecture facilitates robust extraction and classification of complex EEG features, paving the way for more reliable BCIs.
• This study highlights the potential of advanced deep learning techniques for advancing brain-computer interface technology.