Cortical Control of Virtual Self-Motion Using Task-Specific Subspaces.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Robotics

Background:

• Brain-machine interfaces (BMIs) have shown success in controlling reaching movements.
• Existing BMIs often assume consistent neural activity patterns across different movement types.
• There is a need for BMIs that can decode diverse motor tasks beyond simple reaching.

Purpose of the Study:

• To develop and evaluate a self-motion BMI for a cycling task in a virtual environment.
• To investigate if decoding strategies need to adapt based on movement class.
• To explore novel decoding approaches that leverage the intrinsic structure of neural activity.

Main Methods:

• Developed a self-motion BMI using cortical activity recorded from monkeys pedaling a virtual track.
• Analyzed neural activity patterns, focusing on covariance structure and correlations with kinematic variables.
• Employed nonlinear decoding methods to leverage multidimensional neural activity subspaces.

Main Results:

• Found that high-variance neural dimensions did not directly correlate with kinematic variables during cycling, unlike in reaching.
• Successfully decoded self-motion by nonlinearly integrating information across multiple high-variance neural dimensions.
• Achieved online BMI control performance comparable to manual control during the cycling task.

Conclusions:

• Decoding algorithms must be tailored to specific movement classes, as neural activity structure varies.
• Correlations with hand velocity are not universally present in motor cortex activity for all tasks.
• Leveraging multidimensional neural subspaces with nonlinear methods can yield high-performance BMIs for complex movements.