LMDA-Net:A lightweight multi-dimensional attention network for general EEG-based brain-computer interfaces and...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG)-based brain-computer interfaces (BCIs) face challenges due to low spatial resolution and signal-noise ratio.
• Traditional EEG feature extraction relies on neuroscience knowledge, potentially limiting BCI performance.
• Existing neural network methods struggle with generalization, prediction volatility, and interpretability.

Purpose of the Study:

• To propose a novel lightweight multi-dimensional attention network (LMDA-Net) for improved EEG-based BCI decoding.
• To enhance feature integration and classification performance across diverse BCI tasks.
• To develop interpretable algorithms for understanding LMDA-Net's extracted features.

Main Methods:

• Developed LMDA-Net incorporating novel channel and depth attention modules tailored for EEG signals.
• Evaluated LMDA-Net on four public datasets, including motor imagery (MI) and P300-Speller.
• Employed class-specific neural network feature interpretability algorithms using class activation maps.

Main Results:

• LMDA-Net demonstrated superior classification accuracy and reduced prediction volatility compared to other models across all tested datasets.
• Achieved highest accuracy within 300 training epochs.
• Ablation studies confirmed the effectiveness of the proposed attention modules.

Conclusions:

• LMDA-Net offers a promising solution for overcoming limitations in EEG-based BCI decoding.
• The model effectively integrates multi-dimensional features, leading to enhanced performance and interpretability.
• LMDA-Net shows potential as a general decoding model for various EEG applications.