Amortized Bayesian inference on generative dynamical network models of epilepsy using deep neural density estimators.

Area of Science:

• Computational Neuroscience
• Medical Imaging Analysis
• Epilepsy Research

Background:

• Whole-brain modeling integrates anatomical data and dynamical models to simulate epilepsy.
• Stochastic generative processes are crucial for inferring brain dynamics in disorders.
• Calculating likelihood functions at a whole-brain scale is computationally challenging, necessitating likelihood-free algorithms.

Purpose of the Study:

• To introduce simulation-based inference for a virtual epileptic patient model (SBI-VEP).
• To enable efficient estimation of parameters related to epileptogenic and propagation zones.
• To provide a framework for fast and reliable inference on brain disorders using neuroimaging.

Main Methods:

• Utilizing simulation-based inference to amortize the approximate posterior of the generative process.
• Employing deep learning algorithms for conditional density estimation.
• Applying invertible transformations to establish statistical relationships between parameters and observations.

Main Results:

• SBI-VEP efficiently estimates the posterior distribution of parameters defining seizure zones.
• The method successfully utilizes sparse intracranial electroencephalography recordings.
• Demonstrated ability to handle non-linear dynamics and parameter degeneracy.

Conclusions:

• SBI-VEP offers an efficient Bayesian methodology for inferring brain disorder parameters.
• This approach facilitates fast and reliable inference from neuroimaging data.
• Paves the way for improved understanding and prediction of epilepsy.