Neural Manifold Constraint for Spike Prediction Models Under Behavioral Reinforcement

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biomedical Engineering

Background:

• Spike prediction models are vital for neural prostheses to restore communication by predicting downstream neural activity from upstream signals.
• Reinforcement learning (RL) is necessary for training these models when ground truth is unavailable, but existing methods lack constraints, leading to unrealistic outputs.
• Current models neglect neural firing pattern constraints and correlations, raising concerns for clinical applications.

Purpose of the Study:

• To introduce and evaluate a neural manifold constraint for shaping RL-generated spike trains in feature space.
• To improve the biological plausibility and clinical viability of spike prediction models for neural prostheses.
• To ensure that predicted neural activity remains within natural ranges and maintains realistic correlations.

Main Methods:

• Proposed a neural manifold constraint using first and second-order statistics from neural recordings during free movement.
• Integrated constraint terms into RL optimization for models predicting primary motor cortex (M1) spikes from medial prefrontal cortex (mPFC) spikes in rats performing a discrimination task.
• Trained models using behavioral reinforcement within the estimated neural manifold.

Main Results:

• Constrained models generated M1 spike trains that closely resembled real recordings.
• Achieved comparable behavioral success rates to unconstrained models while reducing mean squared error of neural firing by 61%.
• Demonstrated increased model robustness across data segments and induced realistic neural correlations.

Conclusions:

• The neural manifold constraint is a promising tool for enhancing spike prediction models in neural prostheses.
• This approach enables the restoration of transregional neural communication with high behavioral performance and realistic microscopic neural patterns.
• The method addresses limitations of existing models by incorporating biological constraints for more clinically relevant predictions.