Macaque’s Cortical Functional Connectivity Dynamics at the Onset of Propofol-Induced Anesthesia

The estimation of functional neural networks was performed according to the following steps:

• •

  Databases of MDR-ECoG electrophysiological neural activity recordings were used.

• •

  Each electrode of the matrix array was considered a node of the network and resembled the cortical region in which it was positioned.

• •

  Granger causality in the frequency domain was used as a neural connectivity estimator to infer functional connectivity interactions among the electrode’s records (time series).

• •

  Pairwise association values between the electrodes were saved into adjacency matrices.

• •

  Matrices were estimated serially over time at every five seconds of the neural recording experiment in five physiological frequency bands.

1. 1.

   An IIR-notch filter was applied to attenuate components of the signal at 50Hz.

2. 2.

   The signal was down-sampled from 1KHz to 200Hz.

3. 3.

   The electrophysiological time series were divided into windows of 1000 points, equivalent to a five-second recording of neural activity.

4. 4.

   On each time series, the trend was removed, and the average was subtracted.

5. 5.

   The tests KPSS (Kwiatkowski et al., 1992) and ADF [Augmented Dickey Fuller] (Hamilton, 1989) were applied to verify the time series stationary conditions.

For the computation of Granger causality in the frequency domain functional interactions, with some adaptations, the following libraries were used: MVGC GRANGER TOOLBOX (Seth, 2010) and the library BSMART toolbox [Brain-System for Multivariate AutoRegressive Timeseries toolbox] (Cui et al., 2008).

1. 1.

   Model Order

   The criteria of model selection from Akaike (AIC) and Bayes/Schwartz (BIC) were used to infer the model order. Both methods returned a model order parameter equal to seven.

2. 2.

   Causal Interactions

   At each window of 1000 points, Granger causality in the frequency domain interactions was pair-wise computed among the 128-time series by the use of the function cca_pwcausal() (MVGC GRANGER TOOLBOX) in five physiological frequency bands: Delta (0-4Hz), Theta (4-8Hz), Alpha (8-12Hz), Beta (13-30Hz) and Gamma (25-100Hz). The interaction values obtained were saved in adjacency matrices.

The functional connectivity matrices serially estimated over time along with the anesthetic induction experiment were projected into a bi-dimensional representation using a t-Stochastic Neighborhood Embedding algorithm (Van der Maaten and Hinton, 2008). The respective entries vectors of the weighted directed functional connectivity matrices were used as input features into the t-SNE algorithm (Van der Maaten and Hinton, 2008) implemented in the R-CRAN package Rtsne (Krijthe et al., 2018), with the following parameters: perplexity=30, exaggeration factor=12, and the maximum number of interactions=500.

To obtain directed unweighted networks, for each sequence of graphs respective to a physiological frequency band, a threshold was applied, and only interactions with magnitude values higher than the threshold were considered edges of the graphs. The threshold was set to 5% higher interaction values of all matrices of the sequence corresponding to each frequency band. This procedure finally resulted in directed, unweighted functional brain networks.

To ensure that our results were not exclusively dependent on the chosen threshold, twelve distinct values of threshold respective from 0.5% to 15% higher interaction values were evaluated (0.5%, 1%, 2%, 3%, 4%, 5%, 6%, 7%, 8%, 9%, 10%, and 15% ). The same dynamic behavior on the network’s average degree over time across the anesthetic induction experiment was verified along with the different thresholds; see complementary material¹¹1Complementary material is not currently available online.. The manuscript shows results respective to one of the four experiments performed; the other three experiments were used to validate the results presented. See complementary material to compare the results and figures of all four independent experiments.

The present research investigated unweighted-directed large-scale functional brain networks obtained by applying a threshold over the Granger causality association matrices. The analysis of functional connectivity and Granger causality flow among brain areas was evaluated in terms of the number of connections in these directed binary networks. Thus, the analysis did not differentially take into account the exact magnitude of association values returned by the Granger causality method. The anesthetic induction experiments in the macaque animal models involved three distinct conditions: alert with eyes closed, transition, and general anesthesia. To describe and verify the alterations due to the administration of propofol in the brain state, we compared the populations of functional brain networks respective to the transition and general anesthesia phases to those estimated during the resting state with eyes closed conditions, a control state without the presence of propofol, which supposedly reflects natural brain activity in the absence of cognitive demands and visual-evoked responses.

Once the recording technique and methodology applied provided several networks throughout the experiment, populations of networks respective to each stage were considered. Therefore, it permitted us to make quantitatively legitimate comparisons among the different physiological states. Properties were evaluated under a statistical test to verify whether significant distinctions occurred or not due to the administration of propofol throughout the transition and general anesthesia conditions. For each property evaluated, the Wilcoxon signed-rank test with a 5 percent p-value was independently applied in each of the four experiments on both conditions [greater] and [smaller]. In the manuscript, only alterations confirmed by the statistical test simultaneously in both four independent realized experiments were considered and reported as results.

The first remarkable alterations in functional connectivity occurred within 20-25 seconds after propofol administration when the average degree of functional brain networks diminished for about 20-25 seconds. As it is possible to observe in $Figure\cdot 1$, a decrease in the mean functional connectivity occurred on the Delta (0-4Hz), Theta (4-8Hz), Alpha (8-12Hz), and Beta (13-30Hz) frequency bands almost right after the administration of propofol, which is indicated by the first red vertical line, time=11 minutes (see $Figure\cdot 1$, Sub-Figures 1 to 4). The same conclusions can be drawn by analyzing $Figure\cdot 2$, where it is possible to verify the respective period when functional connectivity decreased (see $Figure\cdot 2$, Sub-Figures 12 to 16). As each Sub-Figure corresponds to 5 seconds of neurophysiological records, it is evidenced that this first decrease in functional connectivity lasted for about 20-25 seconds. The results shown in $Figure\cdot 3$ reveal large areas of the cortex, mainly over the occipital lobe, colored in white, thus demonstrating that this cortical region did not have a resultant Granger causality flow, supposedly due to the absence or reduced number of connections. This result also corroborates the conclusions obtained from Figures 1 and 2.

Following the decrease in functional connectivity that occurred within 25 seconds after the drug administration, functional connectivity then started to gradually increase for about 4-5 minutes, achieving a maximum peak and later started to progressively decrease for about 6-7 minutes until the loss of consciousness point (LOC) was achieved. From $Figure\cdot 1$, it is verified that this increase in functional connectivity occurred in the five physiological frequency bands analyzed, and the most pronounced increase occurred in the frequency band Gamma (25-100Hz). By analyzing $Figure\cdot 2$, one can verify on the frequency band Beta (13-30 Hz) how this phenomenon of gradual increase in functional connectivity manifested on individual nodes over brain areas (see $Figure\cdot 2$, Sub-Figures 20-35), first 80 seconds of the gradual increase. From ($Figure\cdot 2$, Sub-Figures 20-35), it was also verified that almost all electrodes were highly connected, revealing that this high connectivity effect involved all cortical regions in which electrodes were positioned.

Granger Causality is a functional connectivity estimator aimed at inferring the flow of information from one cortical area to another. As a measure that offers the direction in which the functional interaction occurs, the resultant OUT-DEGREE minus the IN-DEGREE of each node can be interpreted as a measure of causal flow action associated with the respective node. This resultant causal flow action profile identifies at a given moment the nature of preferentially affecting (assuming a role of a source of Granger causality) or being affected by other nodes (assuming an action as a sink of Granger causality). Therefore, provide an estimate of the respective cortical area in preferentially sending or receiving information from other cortical regions. $Figure\cdot 3$ displays the resultant causal flow profile of the electrodes over their respective cortical positional coordinates along time, at every five seconds, right after propofol administering. From $Figure\cdot 3$, one can notice that when the brain entered the regime characterized by high connectivity, which was established about 110 seconds after the drug injection (see $Figure\cdot 3$, Sub-Figures 22-35), electrodes correspondent to the occipital and temporal lobes had a high positive causal flow. This result reveals that these brain areas tended to exert a strong causal influence and assumed predominant action as a source of Granger causality. On the other hand, electrodes of the parietal and frontal lobes predominantly had a negative resultant causal flow profile. Nodes with negative causal flow reflect on cortical areas predominantly affected by other regions. The obtained results, therefore, evidenced that areas of the frontal and parietal lobes acted mainly as sinks of Granger causality during the transition.

This pattern took place during the transition state when functional connectivity significantly increased over the entire cortex. Once presented as one main recurrent pattern, as opposed to resting-state conditions in which the brain presented a much more dynamic and diverse repertoire of states, it showed that the brain entered into a very particular state in which cortical areas assumed one predominant definite behavior of the Granger causality flow dynamics.

To verify the alterations caused by the administration of propofol on the exchange process of Granger causality flow among the cortical brain lobes, for each time-resolved functional network, we have counted the number of directed connections leaving one cortical lobe to another on all combinations (frontal → parietal, frontal → temporal, frontal → occipital, parietal → frontal, parietal → temporal, parietal → occipital, temporal → frontal, temporal → parietal, temporal → occipital, occipital → frontal, occipital → parietal and occipital→temporal). Once the average degree varied considerably throughout the experiment, we then normalized it by the average degree of each time-resolved network. Then we compared the network populations of the transition and general anesthesia phases to the resting state networks. Finally, by applying the Wilcoxon signed-rank test with a 5 percent p-value on both conditions [greater] and [smaller], we reported the results, which were confirmed by the tests concurrently in all four experiments independently realized.

• •

  Transition Phase

  When compared to the resting state conditions, the main alterations observed were increased interactions from inferior-posterior (temporal and occipital lobes) to anterior parts of the brain (frontal and parietal lobes), see ($Figure\cdot 4$, Sub-Figures A, B, C, and D).

  In the Delta (0-4Hz) frequency band, it was verified that there was an increase in the interactions from the occipital lobe to the frontal and parietal lobes and also from the temporal lobe toward the parietal lobe (see $Figure\cdot 4$, Sub-Figure A). In Theta (4-8Hz), interactions from the occipital lobe to the frontal and parietal lobes and also interactions from the temporal lobe to the frontal lobe increased (see $Figure\cdot 4$, Sub-Figure B). In Alpha (8-12Hz), an increase in the number of interactions from the temporal lobe towards the frontal and parietal lobes and from the occipital lobe to the frontal lobe occurred (see $Figure\cdot 4$, Sub-Figure C). Finally, in Beta (13-30Hz), an increase in the interactions was verified from the occipital lobe toward the frontal and parietal lobes and from the temporal lobe to the parietal lobe (see $Figure\cdot 4$, Sub-Figure D). It was not verified any interactions among brain lobes that statistically significantly decreased during the transition phase.
• •

  General Anesthesia Phase

  In general, interactions between the occipital and temporal lobes increased in both directions in the higher frequency bands Beta (13-30Hz) and Gamma (25-100Hz). Connectivity from the parietal to the frontal lobe also increased in the Theta (4-8Hz) and Alpha (8-12Hz) frequency bands. Conversely, the interactions from the parietal lobe to the occipital lobe decreased in the lower and medium frequency bands. In the Delta (0-4Hz) frequency band, there was a decrease in the interactions from the parietal lobe to the occipital lobe (see $Figure\cdot 4$, Sub-Figure F). In the Theta (4-8Hz), an increase in interactions from the occipital lobe to the frontal lobe and from the parietal lobe to the frontal lobe occurred. However, functional connectivity from the parietal to the occipital lobe diminished (see $Figure\cdot 4$, Sub-Figure G). In the Alpha (8-12Hz) frequency band interactions increased from the parietal areas to the frontal lobe. However, there was a decrease in connectivity from the frontal to the occipital lobe and from the parietal to the occipital lobe (see $Figure\cdot 4$, Sub-Figure H). In the Beta (13-30Hz), a decrease in the interactions from the frontal to temporal lobe was verified, and an increase in the interactions from the temporal lobe to the occipital lobe and from the occipital regions to the temporal areas was also observed (see $Figure\cdot 4$, Sub-Figure I). Finally, in the Gamma (25-100Hz), a decrease in functional connectivity from the temporal lobe to the parietal lobe. Additionally, an increase in the number of interactions from the temporal lobe to the occipital lobe and from the occipital lobe to the temporal lobe was also verified (see $Figure\cdot 4$, Sub-Figure H). Particularly different from the transition phase, where only an increase in interactions was verified, during the general anesthesia state, a decrease in certain brain functional connections established among brain lobes was verified on all frequency bands analyzed. The functional connectivity interactions that were most compromised were those from the anterior (fronto-parietal) to the inferior posterior (temporal-occipital) cortical areas.

The transitional state that occurred between the resting-awake and general anesthesia conditions, which are regarded as two physiologically stable and coherent states, did not demonstrate that it took place as a single-step change or as a continuous unraveling of one single neural process. Instead, functional connectivity first decreased considerably, then started to increase for a few minutes, and then later decreased again until LOC was achieved. This succession of changes in functional connectivity dynamics seemed consistent with the unfolding of a series of different neural processes.

Once the MDR-ECoG technique provided at the same time, great coverage of the cortical surface along with high spatial resolution permitted us to evaluate functional connectivity at distinct levels.

Functional connectivity at the global level was evaluated to ascertain how the functional connectivity of the whole large-scale cortical networks changed according to the distinct experimental conditions. We evaluated the average degree of the complete network (all nodes and all connections). Reflected on how connected or disconnected the brain networks were in each phase of the experiment.

• •

  Transition Phase

  The Wilcoxon test with a p-value of 5% confirmed that, during the transition, the global functional connectivity increased on all frequency bands analyzed in both four experiments.
  

  In Delta (0-4Hz), the average connectivity in alert conditions was about 12; during the transition, it increased and achieved a peak by order of 50 (see $Figure\cdot 1$, Sub-Figure A. In Theta (4-8Hz), Alpha (8-12Hz), and Beta (13-30Hz), the average connectivity in the resting state was on the order of 10-12, and during the transition, it achieved a peak of around 60. This observation reveals that functional connectivity increased up to six times during the transition at those frequency bands (see $Figure\cdot 1$, Sub-Figures B, C, and D).

  In the Gamma frequency band (25-100Hz), the average degree was about 5-6 and increased to around 80 during the peak, thus verifying that functional connectivity increased up to about 15 times in the Gamma frequency band (see $Figure\cdot 1$, Sub-Figure E).

• •

  General Anesthesia Phase

  The statistical test confirmed that, during general anesthesia, functional connectivity at the global level increased in the Gamma frequency band (25-100Hz). However, on the frequency bands Delta (0-4Hz), Theta (4-8Hz), Alpha (8-12Hz), and Beta (13-30Hz), the statistical tests were inconclusive. This result is very interesting once it demonstrates that propofol-induced general anesthesia is not necessarily given by means of a substantial decrease in cortical functional connectivity at global levels (see $Figure\cdot 1$).

Our results demonstrate that propofol-induced anesthesia is not given through a complete cortical functional network failure. It was verified that during general anesthesia, the brain keeps active and cortical areas still functionally interact. Compared to the resting state networks, the statistical test did not confirm that the mean functional connectivity of general anesthesia networks increased or decreased at global levels. This result reveals that propofol-induced unconsciousness is not necessarily given by means of a significant reduction in global functional connectivity. We projected the time-resolved functional networks into a bi-dimensional space using a non-linear distance preserving t-SNE algorithm (Van der Maaten and Hinton, 2008). We then verified that populations of networks respective to the resting, transition, and general anesthesia states were located in particular regions of the bi-dimensional space, being reasonably separated, thus revealing that brain activity is distinct and specific during each state. This result does not go in accordance with hypotheses that, during general anesthesia, brain functional networks should be similar to networks respective to the resting state, typically resembling the structural brain networks that underlie functional connectivity. Instead, our results befit that functional networks under general anesthesia are characteristic and specific; certain connections were affected, while others might have been favored and configured in a reasonably distinct state from the brain’s natural resting conditions. By analyzing the t-SNE projections, it was also noticed that general anesthesia functional networks were displayed closer to the resting state networks than to those of the transition state; this outcome might be due to the mean functional connectivity, which did not vary too much among these two states, while it remarkably increased during the transition.

With the MDR-ECoG recording technique and Granger causality in the frequency domain methodology, we could contemplate the unfolding of the propofol anesthetic induction in the macaque’s functional brain networks. The most remarkable phenomenon observed in the research was an expressive increase in functional connectivity, which occurred after the anesthetic drug administration during the transition phase. This state was also characterized by a specific Granger causality flow profile, from the occipital and temporal to frontal-parietal cortical lobes. It was also verified that during general anesthesia, the magnitude of the mean functional connectivity was about the same as during the resting state. However, the functional connections established among certain brain regions had been compromised while others had been favored, verifying alterations in the five frequency bands analyzed. The present research complements the actual scenery and brings novel knowledge toward the characterization of the functional connectivity dynamics throughout propofol anesthetic induction in non-human primates.

The exact mechanisms of general anesthesia are still incompletely known, and the characterization of functional connectivity in the anesthetized states induced by different drugs is, for sure, one essential step for the comprehension of neural correlates of general anesthesia and the elucidation of neuronal pathways and circuits that are involved with unconsciousness, amnesia, and immobility. This elucidation could help in the design of novel drugs or a combination of drugs, that could promote a better and safer anesthetic induction that would best fit the needs of clinical applications. It could also offer new knowledge about mechanisms that involve the depression of the central nervous system, such as sleep. The characterization of functional connectivity during general anesthesia induced by distinct anesthetic agents can also constitute novel relevant knowledge for neuroscience, as it may help the comprehension of neural correlates of consciousness and offer more insights into essential neural processes that account for the brain to perform its activities.