Finding Kinematics-Driven Latent Neural States From Neuronal Population Activity for Motor Decoding

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Intracortical brain-machine interfaces (BMIs) show promise for restoring mobility but face challenges in maintaining high decoding performance.
• Traditional decoding methods struggle with noise and direct neural-to-physical quantity mapping.
• Latent neural state models offer robust population activity readout but may lack direct kinematic relevance.

Purpose of the Study:

• To develop a novel approach for extracting kinematics-dependent latent factors from neural population activity.
• To improve the decoding accuracy of intracortical BMIs for restoring motor function.

Main Methods:

• Proposed a method to identify kinematics-dependent components within latent neural factors using linear regression.
• Estimated these components from population activity via nonlinear mapping.
• Compared the decoding performance of the proposed model against established methods like FA, GPFA, LFADS, PSID, and firing rates.

Main Results:

• The proposed kinematics-dependent latent factors generated neural trajectories that effectively distinguished pre- and post-motion onset states.
• The novel analysis model achieved higher decoding accuracy compared to all other tested models, showing an average improvement of %.

Conclusions:

• The developed approach successfully extracts latent neural states specific to kinematic information.
• This method holds potential for significantly enhancing decoding performance in online intracortical BMIs, aiding mobility restoration.