A manifold of heterogeneous vigilance states across cortical areas

Vigilance states are conventionally divided into wake, slow-wave sleep (SWS) and rapid eye-movement (REM) sleep based on distinct patterns of neural activity and muscle tone. These states are traditionally thought to be global. However, there is growing evidence for a departure from a simple model of three basic states. Sleep and wake may be more continuous rather than discrete states. Short periods of wake-like activity in sleep and vice versa can occur regularly. This diversity of state expression is further complicated by evidence of local sleep across the brain. To account for the complexity and heterogeneity of vigilance states across the cortex, we turned to an unsupervised deep learning technique, a variational autoencoder (VAE), to discover a low-dimensional manifold on which the brain states evolve. We recorded local field potentials simultaneously on 14 channels across the cortex of mice during normal sleep-wake cycle. The manifold discovered by our unsupervised model from these data captured the three basic states of vigilance, but also revealed micro-states and transition dynamics. When applied to data from multiple electrodes across the cortex, our model uncovered heterogeneity of vigilance state expression across cortical areas, such as absence of REM in certain areas. We also discovered that micro-states can occur locally and characterized the spatiotemporal dynamics of global transitions between states, such as patterns of SWS to REM propagation from the visual cortex or persistence of slow rhythm in the frontal cortex in transition to wake. These results provide strong evidence of the non-global nature of brain states, refuting the accepted notion of global uniform states and adding to the growing literature of heterogeneous brain states.