Brain Waves Analysis Via a Non-Parametric Bayesian Mixture of Autoregressive Kernels

Area of Science:

• Neuroscience
• Computational Neuroscience
• Signal Processing

Background:

• Standard analysis of brain electrical activity relies on predefined frequency bands in spectral density functions (SDF).
• This approach is limited as oscillation frequency and bandwidth vary with cognitive demands, making arbitrary band setting suboptimal.
• A data-driven method is needed to dynamically identify relevant spectral characteristics.

Purpose of the Study:

• To introduce the Bayesian mixture auto-regressive decomposition (BMARD) method for analyzing brain electrical activity.
• To overcome the limitations of a priori defined frequency bands in spectral analysis.
• To identify prominent spectral peaks, their locations, and bandwidths in a data-driven manner.

Main Methods:

• Developed the Bayesian mixture auto-regressive decomposition (BMARD) method.
• BMARD uses a Dirichlet process mixture model based on second-order auto-regressive processes to represent the standardized SDF.
• Employed a Metropolis-Hastings within Gibbs algorithm for posterior distribution sampling of mixture parameters.

Main Results:

• Simulations confirmed the robust performance of the proposed BMARD method.
• The method successfully identified the number of spectral peaks, their frequencies, and bandwidths.
• Application to rat hippocampal local field potential (LFP) data revealed specific oscillatory patterns linked to cognitive demands.

Conclusions:

• The BMARD method offers a data-driven alternative to traditional spectral analysis of brain activity.
• It accurately characterizes oscillatory dynamics, including peak frequencies and bandwidths, without arbitrary band definitions.
• This approach facilitates a more precise examination of the relationship between brain oscillations and cognitive processes.