An Explainable 3D-Deep Learning Model for EEG Decoding in Brain-Computer Interface Applications

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Decoding electroencephalographic (EEG) signals is crucial for brain-computer interface (BCI) development.
• High inter-subject variability in EEG necessitates time-consuming user-specific calibration, limiting deep learning applications.
• Existing methods often require large datasets for effective model training.

Purpose of the Study:

• To propose a multidimensional and explainable deep learning framework for fast and interpretable EEG decoding.
• To address the challenges of inter-subject variability and reduce adaptation time in BCI systems.
• To introduce a novel deep learning model, EEGCubeNet, for enhanced EEG signal processing.

Main Methods:

• EEG signals were projected into the spatial-spectral-temporal domain.
• A custom three-dimensional (3D) Convolutional Neural Network (EEGCubeNet) was developed.
• A global-to-subject-specific fine-tuning approach was employed, combining population-level training with individual user fine-tuning.
• A 3D occlusion sensitivity analysis-based explainability method (3D xAI-OSA) was introduced.

Main Results:

• EEGCubeNet achieved state-of-the-art performance in discriminating motor imagery tasks (hand open/close) from a resting state.
• The model demonstrated high accuracy ([Formula: see text] and [Formula: see text] for HC vs. RE and HO vs. RE, respectively).
• Reduced framework complexity and training time were observed compared to existing methods.
• The 3D xAI-OSA method provided relevance maps for enhanced prediction transparency.

Conclusions:

• The proposed EEGCubeNet framework offers a significant advancement in fast and interpretable EEG decoding for BCI applications.
• The global-to-subject-specific fine-tuning strategy effectively reduces adaptation time and improves model efficiency.
• The integration of explainability features enhances the transparency and trustworthiness of the deep learning model.
• The developed framework shows promise for broader adoption of deep learning in BCI research and development.