STaRNet: A spatio-temporal and Riemannian network for high-performance motor imagery decoding

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) are crucial for human-computer interaction via brain signals.
• Existing BCIs require high accuracy, robustness, and end-to-end processing for motor imagery (MI).

Purpose of the Study:

• Introduce STaRNet, a novel model for accurate and robust MI-based BCIs.
• Evaluate STaRNet's performance against state-of-the-art models.

Main Methods:

• Integrate multi-scale spatio-temporal CNNs with Riemannian geometry.
• Employ matrix logarithm for tangent space transformation and dense layer for classification.
• No preprocessing required.

Main Results:

• Achieved 83.29% accuracy and 0.777 kappa on BCI Competition IV 2a.
• Attained 95.45% accuracy and 0.939 kappa on the High Gamma Dataset.
• Demonstrated superior robustness on challenging subjects.

Conclusions:

• STaRNet offers accurate, robust, and end-to-end capabilities for MI-BCIs.
• The model effectively extracts relevant spatio-temporal features and frequency bands.
• STaRNet advancements can accelerate BCI development.