Predicting Cortical Signatures of Consciousness using Dynamic Functional Connectivity Graph-Convolutional Neural Networks

Decoding the levels of consciousness from cortical activity recording is a major challenge in neuroscience. Using clustering algorithms, we previously demonstrated that resting-state functional MRI (rsfMRI) data can be split into several clusters also called "brain states" corresponding to "functional configurations" of the brain. Here, we propose to use a selfsupervised machine learning method based on artificial neural networks to predict functional brain states across levels of consciousness from rsfMRI. The Functional Connectivity (FC) matrices reflect the brain-state dynamic at a given time. Because it is key to consider the FC topologies, a specific graph-Convolutional Neural Network (gCNN), namely BrainNetCNN, is considered to predict the brain states in awake and anesthetized nonhuman primates. To avoid the circularity that remains in the training stage, where the target is composed of pseudo-labels, recent self-supervised techniques are implemented. Using a linear probe for the prediction, the network achieves a prediction accuracy consistent with state-of-the-art methods lying in [0.655, 0.759] depending on the experimental settings. To put forward the interest of such a representation, the transition probabilities and the set of connections found to be important for predicting a brain state are computed. This latter is directly linked with the level of consciousness. The results demonstrate that deep learning methods are not only able to predict brain states but also provide additional insight into cortical signatures of consciousness with potential clinical consequences for the monitoring of anesthesia and the diagnosis of disorders of consciousness.