A low-dimensional variational autoencoder (VAE) can model dynamic functional connectivity from resting-state fMRI to capture brain patterns related to consciousness. The VAE balanced reconstruction and classification performance compared to other models. Its latent representations stratified brain patterns and experimental conditions. Receptive field analysis identified latent directions for transitioning between patterns, and an ablation study virtually inactivated brain areas. The model summarized consciousness-specific information in key inter-areal connections, consistent with the global neuronal workspace theory. This framework may support development of an interpretable computational brain model for disorders of consciousness.
A new framework uses a linear latent variable model to identify and quantify resting-state brain networks from fMRI recordings, addressing the atlas selection problem and enabling statistical inference on network activities. Applied to monkey data under different anesthetics with static functional connectivity, the method suggests that two networks—one fronto-parietal and cingular, another posterior (temporo-parieto-occipital)—strongly influence shifts in consciousness, particularly between anesthesia and wakefulness. This aligns with the global neural workspace and integrated information theories of consciousness. The approach can also decode anesthesia level from network activities and may aid studies of disorders of consciousness.