Variational Autoencoder with Truncated Mixture of Gaussians for Functional Connectivity Analysis

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning in Neuroimaging

Background:

• Resting-state functional connectivity (rs-fC) states are typically identified by clustering dynamic connectivity patterns.
• Current clustering methods struggle to differentiate major states from minor ones, leading to noise sensitivity.
• Accurate identification of brain states is crucial for understanding neural dynamics.

Purpose of the Study:

• To develop a novel method for robust clustering and outlier detection of brain connectivity states.
• To differentiate major functional connectivity states from rare, minor states.
• To investigate the relationship between identified major states and demographic factors like age.

Main Methods:

• Proposed a truncated Gaussian-Mixture Variational Autoencoder (tGM-VAE) model.
• Modeled major states using a non-linear generative process with Gaussian-mixture priors.
• Modeled minor states using a uniform distribution to handle outliers and rare patterns.
• Applied the tGM-VAE to synthetic data and real rs-fMRI data from 593 healthy adolescents.

Main Results:

• tGM-VAE demonstrated superior accuracy in clustering connectivity patterns compared to existing methods on synthetic data.
• The model successfully identified meaningful major functional connectivity states in adolescent rs-fMRI data.
• The dwell time of these identified major states exhibited a significant correlation with participant age.

Conclusions:

• The tGM-VAE provides a robust framework for joint clustering and outlier detection in dynamic functional connectivity analysis.
• This approach enhances the reliability of identifying major brain states from rs-fMRI data.
• The age-related changes in the dwell time of major connectivity states suggest developmental implications in neural network organization.