SAND: Spectral-Attention Neural Decoding of Hand Kinematics from Low-Frequency EEG Dynamics.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Brain-Computer Interface (BCI) technology is crucial for neural rehabilitation but faces challenges in precise hand kinematics decoding using electroencephalography (EEG) due to limited accuracy and adaptability across subjects.
• Existing methods struggle with robust reconstruction of 2D/3D hand trajectories from noisy EEG signals.

Purpose of the Study:

• To introduce the Spectral-Attention Neural Decoder (SAND), a novel hybrid framework designed to enhance the precision and cross-subject adaptability of hand kinematics decoding from EEG.
• To improve robust 2D/3D trajectory reconstruction for non-invasive BCI applications.

Main Methods:

• Developed a dual-branch deep learning architecture: a frequency-domain pathway for spectral embedding and a temporal-attention pathway using transformer networks.
• Leveraged spectral decomposition to identify hand movement information primarily encoded in low-frequency EEG bands.
• Validated the framework on the WAY EEG Grasp-and-Lift dataset and a self-collected dataset using five-fold cross-validation.

Main Results:

• SAND achieved state-of-the-art performance in hand-trajectory decoding, with Pearson correlation coefficients of 0.9595 (x-axis), 0.9534 (y-axis), and 0.9293 (z-axis).
• Demonstrated significant improvements (0.07-0.13) over baseline methods.
• Showcased strong cross-task generalization with average correlation coefficients of 0.90 (x-axis) and 0.96 (y-axis) on a 2D reconstruction task.
• Validated temporal alignment with kinematic data through dynamic time warping analysis.

Conclusions:

• The Spectral-Attention Neural Decoder (SAND) offers a robust and effective solution for precise hand motion decoding from EEG signals.
• SAND advances non-invasive BCI applications by overcoming limitations in precision and cross-subject adaptability.
• The findings highlight the potential of integrating spectral decomposition and adaptive deep learning for improved neural decoding.