Dynamic estimation of auditory temporal response functions via state-space models with Gaussian mixture process noise

Area of Science:

• Computational Biology
• Neuroscience
• Signal Processing

Background:

• State-space models with Gaussian statistics are standard for biological latent dynamics but fail to capture abrupt changes.
• Biological processes, like brain dynamics, exhibit complex features not addressed by traditional Gaussian models.
• Gaussian mixture process noise models offer potential but lack established data-driven inference methods.

Purpose of the Study:

• Develop efficient algorithms for inferring parameters of Gaussian mixture process noise models.
• Apply these algorithms to extract neural dynamics from magnetoencephalography (MEG) data in a cocktail party setting.
• Improve the estimation of dynamic Temporal Response Functions (TRFs) for auditory processing.

Main Methods:

• Developed an Expectation-Maximization algorithm for estimating process noise parameters from state-space observations.
• Applied the algorithm to simulated and experimentally-recorded MEG data.
• Utilized Gaussian mixture models to represent process noise, capturing heterogeneity.

Main Results:

• Richer Gaussian mixture process noise models significantly improved state estimation and TRF dynamics heterogeneity.
• The proposed method demonstrated improvements over existing TRF estimation techniques for MEG data.
• Successful estimation of neural dynamics in a cocktail party scenario was achieved.

Conclusions:

• The developed methodology provides a framework for efficient inference of Gaussian mixture process noise models.
• This approach offers a reliable alternative for probing neural dynamics in complex auditory settings.
• Potential applications include attention decoding for smart hearing aids and analysis of diverse biological data.