Bayesian multilevel hidden Markov models identify stable state dynamics in longitudinal recordings from macaque...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Cortical computation may rely on neural populations, not single neurons.
• Analyzing chronic neural population activity is complex due to high dimensionality and signal changes.
• Existing Hidden Markov Models (HMMs) have limitations in analyzing neural spiking data, longitudinal data, and condition-specific differences.

Purpose of the Study:

• To develop a novel multilevel Bayesian HMM to analyze chronic neural population activity.
• To address limitations of previous HMM approaches by incorporating statistical properties of neural data and longitudinal analysis.
• To model condition-specific differences in neural population activity.

Main Methods:

• Developed a multilevel Bayesian HMM with multivariate Poisson log-normal emission probabilities.
• Incorporated multilevel parameter estimation and trial-specific condition covariates.
• Applied the framework to multi-unit neural spiking data from macaque primary motor cortex during a reaching, grasping, and placing task.

Main Results:

• The model successfully identified latent neural population states associated with behavioral events, even without explicit timing information.
• The identified neural states and their behavioral associations remained consistent across multiple days of recording.
• A single-level HMM failed to generalize across different recording sessions, highlighting the advantage of the multilevel approach.

Conclusions:

• The multilevel Bayesian HMM provides a robust framework for analyzing chronic neural population activity.
• This approach demonstrates stability and utility in identifying behavior-linked neural states over time.
• The framework is well-suited for future studies investigating long-term neural plasticity.