Characterization of Brain Activity Patterns Across States of Consciousness Based on Variational Auto-Encoders

Abstract

Decoding the levels of consciousness from cortical activity recording is a major challenge in neuroscience. The spontaneous fluctuations of brain activity through different patterns across time are monitored using resting-state functional MRI. The different dynamic functional configurations of the brain during resting-state are also called “brain states”. The specific structure of each pattern, lifetime, and frequency have already been studied but the overall organization remains unclear. Recent studies showed that low-dimensional models are adequate to capture the correlation structure of neural activity during rest. One remaining question addressed here is the characterization of the latent feature space. We trained a dense Variational Auto-Encoder (dVAE) to find a low two-dimensional representation that maps dynamic functional connectivity to probability distributions. A two-stage approach for latent feature space characterization is proposed to facilitate the results’ interpretation. In this approach, we first dissect the topography of the brain states and then perform a receptive field analysis to track the effect of each connection. The proposed framework instill interpretability and explainability of the latent space, unveiling biological insights on the states of consciousness. It is applied to a non-human primate dataset acquired under different experimental conditions (awake state, anesthesia induced loss of consciousness).