Dynamic network modeling and dimensionality reduction for human ECoG activity.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biomedical Engineering

Background:

• Dynamic network models are crucial for studying neural representations using electrocorticogram (ECoG) data.
• Existing models primarily focus on spike recordings, not ECoG, necessitating models that capture ECoG's temporal dynamics and reduce dimensionality.
• ECoG offers clinical promise for neurotechnologies, highlighting the need for effective modeling techniques.

Purpose of the Study:

• To devise and evaluate linear and nonlinear dynamic models for ECoG power features.
• To assess the accuracy of linear state-space models (LSSMs) in predicting ECoG dynamics and achieving dimensionality reduction.
• To compare the predictive performance of LSSMs against nonlinear radial basis function (RBF) auto-regressive (AR) models and analyze frequency band differences.

Main Methods:

• Developed linear state-space models (LSSMs) and nonlinear RBF-AR models for ECoG power features.
• Evaluated model accuracy in predicting feature dynamics using numerical simulations and human ECoG data from 10 epilepsy subjects.
• Analyzed dimensionality reduction capabilities of LSSMs and differences in dynamics across ECoG frequency bands.

Main Results:

• LSSMs significantly predicted ECoG power feature dynamics using lower-dimensional latent states.
• Nonlinear RBF-AR models did not offer improved prediction accuracy over LSSMs for human ECoG data.
• ECoG power features in the 1-8 Hz (delta + theta) band showed significantly better predictability and were dominated by slow dynamics.

Conclusions:

• LSSMs with low-dimensional latent states effectively capture dynamics in large-scale human ECoG power features, enabling dynamic modeling and dimensionality reduction.
• The linear assumption in LSSMs is suitable for describing ECoG dynamics.
• Findings support the use of LSSMs for studying brain function/dysfunction and designing closed-loop neurotechnologies.