The variability in brain activity and dynamical regimes across diverse levels of consciousness have been extensively investigated. However, a comprehensive quantification of these variations is still lacking.This work uses multiple computational strategies to study the mesoscale dynamics and the network topology of neuronal ensembles across different levels of consciousness, focusing on wakefulness, anesthesia, and awakening. The data were acquired by two chronically implanted multielectrode arrays (Utah 96), placed in the dorsal prefrontal cortex (dPFC) and dorsal premotor cortex (PMd) of two rhesus monkeys. We fitted a linear autoregressive model to predict the local field potential (LFP) of both areas by integrating the information derived from recorded spiking activity. We found the accuracy of the reconstruction to be sensitive to the level of consciousness, leading to a more predictable signal during anesthesia compared to wakefulness. Interestingly, the efficiency of each electrode's contribution to the prediction appears to be structured both in space and time. Moreover, the awakening stage displayed a combination of features from the other two consciousness levels, both in predictability and network topology. During anesthesia, we found a disruption of inter-area functional connectivity, despite the ability to make significant inter-area predictions at the mesoscale level. Overall, the findings suggest that a simple linear model can capture essential features of mesoscale activity across different levels of consciousness, shedding light on the network dynamics and topology specific to each stage.