Takens-Based Kernel Transfer Entropy Connectivity Network for Motor Imagery Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Decoding motor imagery (MI) from electroencephalographic (EEG) signals is challenging due to signal complexity.
• Existing methods often require extensive preprocessing and struggle with nonlinear, noisy, and non-stationary EEG data.
• Accurate functional connectivity estimation is crucial for robust brain-computer interface (BCI) systems.

Purpose of the Study:

• To develop an end-to-end deep learning model, TEKTE-Net, for inferring directed functional connectivity in MI-based BCI systems.
• To enable reliable decoding of EEG activity without explicit preprocessing by integrating time embeddings and a kernelized Transfer Entropy estimator.
• To enhance the interpretability of deep learning models in BCI applications.

Main Methods:

• Proposed TEKTE-Net, an end-to-end deep learning architecture integrating Takens' embedding via a custom convolutional module.
• Employed a kernelized Transfer Entropy estimator with Rational Quadratic kernels within a differentiable framework to estimate nonlinear, time-delayed interactions.
• Evaluated the model on semi-synthetic causal benchmarks and the BCI Competition IV 2a dataset.

Main Results:

• TEKTE-Net demonstrated robustness in low signal-to-noise ratio conditions.
• The model provided interpretable insights through temporal, spatial, and spectral analyses of functional connectivity.
• Automatic highlighting of contralateral activations and spectral selectivity for beta and gamma bands during MI were observed.

Conclusions:

• TEKTE-Net serves as a fully trainable estimator of functional brain connectivity for decoding EEG activity.
• The model supports advanced motor imagery-based brain-computer interface (MI-BCI) applications.
• TEKTE-Net promotes enhanced interpretability of deep learning models in neuroscience research.