Decoding movement kinematics from EEG using an interpretable convolutional neural network.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Electroencephalography (EEG)-based Brain-Computer Interfaces (BCIs) show promise for intuitive control via continuous decoding of hand kinematics.
• Deep neural networks (DNNs) are effective but often lack interpretability and are limited to within-subject decoding.

Purpose of the Study:

• To develop an interpretable convolutional neural network (ICNN) for decoding 2-D hand kinematics from EEG.
• To evaluate within-subject and cross-subject decoding strategies, including transfer learning, to reduce calibration time.
• To interpret the spectral and spatial EEG features learned by the ICNN.

Main Methods:

• An ICNN was developed to decode 2-D hand position and velocity from EEG data during a pursuit tracking task.
• The ICNN was trained using within-subject and cross-subject approaches, with a focus on transfer learning.
• Feature relevance analysis was performed to interpret learned spectral and spatial EEG patterns.

Main Results:

• The ICNN demonstrated superior performance compared to state-of-the-art decoders, offering an optimal balance of performance, size, and training duration.
• Transfer learning significantly enhanced kinematics prediction accuracy, especially with limited data.
• The delta-band was consistently relevant across all subjects, with alpha, beta, and low-gamma bands relevant for some. Sensorimotor, visual, and visuo-motor areas (contralateral central and parieto-occipital sites) were identified as most important.

Conclusions:

• The ICNN provides an interpretable and effective method for decoding hand kinematics from EEG, advancing BCI technology.
• The study highlights the potential of transfer learning to minimize BCI calibration periods.
• Interpretable EEG feature analysis offers insights into neural correlates of motor control and BCI performance.