A general model based on Riemannian manifold for stable decoding movement trajectory from ECoG signals.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) aim to decode neural signals for device control.
• Electrocorticography (ECoG) offers high temporal and spatial resolution for neural decoding.
• Inter-session variability in ECoG signals hinders reliable BCI performance.

Purpose of the Study:

• To develop a robust framework for decoding 3D hand trajectories from ECoG signals.
• To address the challenge of inter-session variability for improved BCI generalization.
• To enable transfer learning across multiple ECoG recording sessions.

Main Methods:

• Utilized Riemannian-based feature extraction from spatial cross-frequency covariance matrices.
• Computed features across 10 frequency band powers within specific brain areas.
• Employed a stacked long short-term memory (LSTM) network trained on extracted geometric and spectral features.
• Applied the framework to ECoG data from monkeys performing reaching tasks.

Main Results:

• The proposed framework demonstrated stable cross-session performance in decoding hand trajectories.
• Achieved superior performance compared to baseline models relying solely on spectral features.
• Extracted features exhibited invariance to session variability, enhancing generalization.

Conclusions:

• Combining Riemannian geometric features with deep learning (LSTM) effectively addresses inter-session variability in ECoG.
• The developed method shows significant potential for generalized decoding in translational BCI applications.
• This approach advances the development of more reliable and adaptable brain-computer interfaces.