Electroencephalographic α-oscillations in the frontal cortex do not appear during general anesthesia in infants less than 3 to 4 months old
In adults, functional connectivity and brain network integration appear to break down during anesthesia but whether this is true in infants is unknown
In infants younger than four months, slow-wave functional connectivity breaks down during general anesthesia and brain networks are less integrated
Functional disconnections in the cortex might be a common marker of anesthesia-induced unconsciousness in infants and adults
Functional brain connectivity studies can provide important information about changes in brain-state dynamics during general anesthesia. In adults, γ-aminobutyric acid–mediated agents disrupt integration of information from local to the whole-brain scale. Beginning around 3 to 4 months postnatal age, γ-aminobutyric acid–mediated anesthetics such as sevoflurane generate α-electroencephalography oscillations. In previous studies of sevoflurane-anesthetized infants 0 to 3.9 months of age, α-oscillations were absent, and power spectra did not distinguish between anesthetized and emergence from anesthesia conditions. Few studies detailing functional connectivity during general anesthesia in infants exist. This study’s aim was to identify changes in functional connectivity of the infant brain during anesthesia.
A retrospective cohort study was performed using multichannel electroencephalograph recordings of 20 infants aged 0 to 3.9 months old who underwent sevoflurane anesthesia for elective surgery. Whole-brain functional connectivity was evaluated during maintenance of a surgical state of anesthesia and during emergence from anesthesia. Functional connectivity was represented as networks, and network efficiency indices (including complexity and modularity) were computed at the sensor and source levels.
Sevoflurane decreased functional connectivity at the δ-frequency (1 to 4 Hz) in infants 0 to 3.9 months old when comparing anesthesia with emergence. At the sensor level, complexity decreased during anesthesia, showing less whole-brain integration with prominent alterations in the connectivity of frontal and parietal sensors (median difference, 0.0293; 95% CI, −0.0016 to 0.0397). At the source level, similar results were observed (median difference, 0.0201; 95% CI, −0.0025 to 0.0482) with prominent alterations in the connectivity between default-mode and frontoparietal regions. Anesthesia resulted in fragmented modules as modularity increased at the sensor (median difference, 0.0562; 95% CI, 0.0048 to 0.1298) and source (median difference, 0.0548; 95% CI, −0.0040 to 0.1074) levels.
Sevoflurane is associated with decreased capacity for efficient information transfer in the infant brain. Such findings strengthen the hypothesis that conscious processing relies on an efficient system of integrated information transfer across the whole brain.
Each year millions of infants and children are administered general anesthesia for surgery.1,2 Volatile anesthetic drugs bind to multiple targets at the brain and spinal cord, where they exert their physiologic and functional effects.3 Sevoflurane is one of the most commonly used vapor anesthetics in infants and children and is used for its rapid induction, emergence, and recovery profile. Many clinical studies use noninvasive recordings such as electroencephalography as a way to monitor adult brain function during titration of anesthesia. Although several studies have characterized electroencephalography-based dynamics in adults under anesthesia, less is known about the dynamics that occur in the infant brain, particularly during early postnatal development.
Quantifying large-scale functional brain network connectivity reconstructed from neurophysiologic data constitutes a promising method for exploring the brain’s complex dynamics during anesthesia.4 In that regard, graph theory examines network properties of nodes connected by edges, representing brain regions and their functional connections.5
General anesthesia modulates functional connectivity networks at the local, meso, and whole-brain network scales.6 Several recent studies using electroencephalography, functional magnetic resonance imaging, and transmagnetic stimulation indicate that sevoflurane weakens signal correlations or “connectivity” among brain regions that share functionality and specialization during wakefulness.7 Specifically, connectivity studies of resting state networks indicate that within- and between-network functional disconnections occur in the default mode and frontoparietal networks, with connectivity of primary sensory networks maintained.8,9 These alterations in network connectivity indicate that the brain during anesthesia drastically reorganizes toward a less complex configuration in which the brain’s functional systems are more segregated, thus inhibiting whole-brain integration.10
Few studies detailing the sevoflurane-induced alterations in functional connectivity in infants have been described. Early studies suggest that the neurophysiologic responses to sevoflurane are different in infants from those of adults and change as a function of age.11,12 Age-varying changes in the spectral properties indicate that coherent δ-oscillations (1 to 4 Hz) dominate in the first 3 months of life.13 Adult-like features, consisting of a frontal predominance of coherent α-oscillations, emerge in late infancy at ~10 months old.14 However, less is known about anesthesia-associated changes in functional connectivity at the sensor and source levels during early postnatal development.
The study rationale is to provide a comprehensive analysis of how sevoflurane-induced brain δ-oscillations change during anesthesia at the first months of life. The goal is to understand how the functional organization of the infant brain (birth up to 3.9 months old) changes during deep and relatively lighter levels of anesthesia. We characterized infant brain connectivity during a state of maintenance of general anesthesia (where a subject is unresponsive to a noxious surgical incision), and during emergence from general anesthesia (immediately before a behaviorally responsive state). We hypothesized that infant brain networks during maintenance of general anesthesia would display less complex and more fragmented connectivity because general anesthesia-induced loss of consciousness is associated with altered functional connectivity and disruption of whole-brain integration. These findings can provide valuable clinical insights regarding accurate monitoring of anesthesia in the pediatric operating room with analysis of electroencephalography data.
Materials and Methods
The objective of this study was to examine functional connectivity changes in response to different depths of anesthesia in infants aged 0 to 3.9 months old. We analyzed electroencephalography data in 20 infants aged 0 to 3.9 months during administration of sevoflurane general anesthesia for elective surgery. End-tidal anesthetic gas volume and video recordings of behavioral activity were time-locked to the electroencephalography recordings. Functional connectivity measures were evaluated during the maintenance period and during emergence from sevoflurane general anesthesia. This study was approved by the Boston Children’s Hospital (Boston, Massachusetts) Institutional Review Board (protocol No. P000003544) and classified as a “no more than minimal risk” study. Informed written consent was obtained from the parents or legal guardians before each study.
Infants who were scheduled for an elective surgical procedure were recruited from the preoperative clinic at Boston Children’s Hospital from December 2011 to August 2016 (under Institutional Review Board P-3544, with written informed consent obtained from parents/legal guardians). Subjects required surgery below the neck, were clinically stable on the day of study, and had American Society of Anesthesiologists’ physical status I or II. Exclusion criteria were (1) born with congenital malformations or other genetic conditions thought to influence brain development, (2) diagnosed with a neurologic or cardiovascular disorder, (3) born at less than 32 weeks postmenstrual age, and (4) postnatal age of 4 months and older. We included all the infants of this cohort that matched the inclusion criteria and were aged 0 to 3.9 months. The summary demographics of subjects included in the analysis are given in table 1.
Each patient received anesthesia induced with sevoflurane alone or with a combination of sevoflurane and nitrous oxide. Nitrous oxide was added at the discretion of the anesthesiologist. Nitrous oxide was discontinued after placement of an endotracheal tube or laryngeal mask. Epochs used for analysis were comprised of sevoflurane administration with air and oxygen, titrated to clinical signs; end-tidal sevoflurane concentration was adjusted per the anesthesiologist’s impression of clinical need, not a preset end-tidal sevoflurane concentration.
Electroencephalography data were acquired using an electroencephalography cap (WaveGuard electroencephalography cap, Advanced NeuroTechnology, Enschede, The Netherlands). In total, 33 recording electrodes were positioned per the modified international 10/20 electrode placement system. Reference and ground electrodes were located at Fz and AFz, respectively. Electroencephalography activity from 0.1 to 500 Hz was recorded with an Xltek electroencephalography recording system (EMU40EX, Natus Medical Inc., Canada). The signals were digitized at a sampling rate of 1.024 Hz and a resolution of 16 bits.
Clinical Data Collection.
Demographics and clinical information were collected from the electronic medical records and from the in-house anesthesia information management system. End-tidal sevoflurane, oxygen, and nitrous oxide concentrations were downloaded from the anesthetic monitoring device (Dräger Apollo, Dräger Medical Inc., USA) to a recording computer in real-time using ixTrend software (ixcellence, Germany). The signals were recorded at a 1-Hz sampling rate.
Raw electroencephalography signals were exported and processed using custom-built code in MATLAB (MathWorks Inc., USA). Electrodes at M1 and M2 were removed. For conformity, the same channels were taken from all subjects and used for the analysis. The preprocessing pipeline involved: (1) remontaging to a nearest neighbor Laplacian reference using distances along the scalp surface to weight neighboring electrode contributions; (2) bandpass filter (0.1 to 50 Hz); (3) downsampling to 256 Hz; (4) rejection and interpolation of bad channels using EEGLAB15 (https://sccn.ucsd.edu/eeglab/index.php; accessed February 1, 2019) average channel interpolation was 6 ± 2 channels; (5) cleaning of the data using EEGLAB routines; and (6) visually inspecting the remaining data to ensure that no artifact was present.
For each subject, epochs in the electroencephalography recordings were identified based on when the subject was in the maintenance of anesthesia period or in the emergence period. The maintenance of anesthesia (anesthesia) period was defined as a steady state of end-tidal sevoflurane volume ±0.1% required for maintenance of a surgical state of anesthesia. The end-tidal sevoflurane was maintained between 1.8 to 2.5% in all epochs selected (median end-tidal sevoflurane, 1.9 ± 0.1%). The emergence period was defined as the time point in the electroencephalography recording at 2 min before body movement was observed (median end-tidal sevoflurane, 0.3 ± 0.1%). One segment of 100 s was selected for maintenance anesthesia and emergence periods. The choice was made because there were a limited number of clean segments corresponding to maintenance of surgical anesthesia with steady sevoflurane concentration.
Electroencephalography Data Processing.
Electroencephalo-graphy time-frequency decomposition of oscillatory activity into different frequency bands was calculated using multitaper analysis. Multitaper analysis reduces the bias in obtaining the true underlying oscillatory activity caused by standard Fourier techniques.16 Multitaper analysis uses tapers of data and calculates the spectrum within each taper separately by using spectral decomposition functions. For a taper of specific time length and for a frequency band of interest (tile of frequency and time), the time-bandwidth product T × W corresponds to how many such functions will be used in this particular tile of frequency and time. We divided the signal into canonical frequency bands including slow (0.1 to 1 Hz), δ (1 to 4 Hz), θ (4 to 8 Hz), α (8 to 12 Hz), and β (13 to 30 Hz) using a setup of T × W = 3, K = 5 tapers as previously reported.11,13 Quantification of power for each tile of frequency and time was done using “mtspecgramc” function implemented in the Chronux toolbox (http://chronux.org/; accessed February 1, 2019).16
Cross-spectral coherence was calculated by correlating the multitaper spectrums of two sensors i and j. One issue with using cross-spectral coherence relates to the volume conduction problem. This refers to the case when two different electrodes might give spuriously high coherence just because they are measuring the same source.17 One way to overcome this problem is by keeping only the imaginary part of the coherence.18 Two sensors measuring the same source cannot give non-zero imaginary coherence.18 In that regard, we obtained a functional connectivity matrix for each infant where each entry (i, j) represented imaginary coherence between the two channels i and j at the frequency band of interest. Comparing connectivity at the local scale involved statistical comparisons in each pair (i, j) between maintenance of anesthesia and emergence conditions.
Functional Connectivity Networks: Variable 1 Analysis–Complexity.
Each δ-specific matrix was converted to network for each infant by thresholding the most important connections. This was conducted within a range 10 to 50% thresholds T at steps of 2%. The limits were chosen to prevent the network from being severely fragmented and from being random by introducing connections. Graph-theoretic indices such as complexity and modularity were estimated in these thresholded networks as an average of the aforementioned range of thresholds T. Functional connectivity complexity was computed by calculating the Shannon entropy of each network’s degree distribution. The degree sample of each network was computed by using the BCT toolbox (https://sites.google.com/site/bctnet/; accessed February 1, 2019).19 We used a standard estimator to calculate entropy of the degree sample that utilizes the frequency estimates of the sample. Formally, let F=(F(1),F(2),…) the fingerprint of the degree sample of size k, where F(1) is the number of nodes appearing once, F(2) is the number of nodes appearing twice and so on. Degree entropy was then calculated as .20
At the network whole-brain scale, δ-oscillation complexity is a measure of richness of the connectivity repertoire across the whole-brain network and is associated with the brain’s ability to balance integration of information coming from different specialized/segregated modules.21 Complexity was estimated by calculating the entropy of the degree (number of connections of each node of the network) distribution across all scalp electrodes.
Functional Connectivity Networks: Variable 2 Analysis–Modularity.
Functional connectivity was estimated using a modularity index. Modularity is a meso scale graph metric used to calculate the extent to which the nodes of a network can be grouped together in modules.19 Modularity was calculated using the heuristic Louvain algorithm, and the modularity measure derived was averaged more than 50 repetitions of the algorithm.
One way of linking surface electroencephalography activity to source activity in the cortex is to use source reconstruction methods. The two steps for source reconstruction include forward modeling and its inverse solution. Forward modeling refers to using Maxwell’s equations to predict the electromagnetic field produced by the sources at a given electrode or else the “leadfield” matrix. In other words, the leadfield matrix relates the measured activity at the electrode level with the underlying source activity. To calculate the forward model one needs to combine information regarding (1) how electric activity spreads though different tissues (the head model), (2) the position and orientation of different dipoles (the source model), and (3) the electrodes’ locations. As far as the head model is concerned, one approach is to use numerical solutions such as the boundary element model,22 in which the brain is compartmentalized into three tissues (brain, skull, and scalp), with each one being covered by a tessellation. To obtain such a geometrical description, one needs anatomical information from the T1 images such as to segment out the brain, scalp, and skull tissues. Because of a lack of individual data, an alternative way is to use predefined templates. Under this framework, it is important to use age-appropriate brain templates and parameters to accurately quantify the localization and time course in each infant.23 Toward this direction, we used age-specific templates provided by the Richards laboratory (https://jerlab.sc.edu/projects/neurodevelopmental-mri-database/; accessed February 1, 2019).23
In addition to T1 and T2 structural images, scalp, brain, and skull segmentations were also obtained. The specific details are described by Sanchez and colleagues.24
For each template, we used standard conductivity settings and the “bemcp” option in FieldTrip (http://www.fieldtriptoolbox.org/; accessed February 1, 2019)25 to obtain a boundary element head model using the tessellation of the three compartments (brain, skull, and scalp). We used a standard number of vertices for the construction of the head model (3,000, 200, and 1,000, respectively).
The next step required realignment of the electrodes’ positions with the head model. This was performed manually using FieldTrip’s graphical interface. Finally, for the source model, a three-dimensional grid of dipoles with 1-cm resolution was constructed. Electrodes, head model, and source model were aligned and mapped to the same space. To increase the validity of the forward solution, the alignment between the electrodes, head model, and source model was visually inspected. Following this methodology, we obtained the leadfield matrix for each template.
Inverse solution refers to obtaining the source level activity using the leadfield matrix and the obtained electroencephalography data. One popular method for obtaining the inverse solution encompasses beamforming techniques.25 The basic principle of beamforming lies in obtaining a single source’s activity by looking at how it contributes to the measured electroencephalography activity compared with other sources.26 Using the forward solution obtained, alongside the electroencephalography data, we utilized the beamforming technique to obtain the activity of each dipole. We used the FieldTrip toolbox with the option “pcc” to obtain the time course and spectrum of each dipole. We then calculated the dipole × dipole connectivity matrix using the imaginary coherence as with electrode-based networks. It is worth noting that the data here were remontaged to an average reference before source reconstruction. Because of the way a forward solution is obtained, there is usually a small error associated with each channel. By remontaging the electrodes at an average reference, the model error is averaged out across the electrodes, thus allowing more precise inverse solutions.25
After obtaining the dipole matrices, we wanted to group dipoles based on known cortical regions to assist with interpretation of the results. We used a specific parcellation obtained in the Richards templates to group dipoles into specific regions of interest (https://jerlab.sc.edu/projects/neurodevelopmental-mri-database/; accessed February 1, 2019).
The parcellation was based on the macroanatomical Hammers atlas previously used in adult literature.27 We assigned each dipole to a region of interest by overlapping the source model (the dipole positions) and the parcellation image. We then averaged the connectivity values of the dipoles belonging to each region of interest to obtain a region of interest-specific time courses. This resulted in a region of interest × region of interest connectivity matrix for each infant that was eventually used for the source level graph-theoretic analysis presented in the main text.
We excluded regions corresponding to the cerebellum, brainstem, striatum, corpus callosum, hippocampus, thalamus, amygdala, and insula because of known difficulties in obtaining source signals from subcortical regions using beamforming techniques.28 To assess the validity of source reconstruction, we further examined whether signal was obtained from regions where no signal would be expected, such as the ventricles. To do this, we used the ventricle regions defined in the Hammers atlas, and we observed that no signal could be extracted using these regions as regions of interest. Matrix thresholding was conducted by looking at a range of thresholds 10 to 50% in steps of 2%. Network measures (pairwise connectivity, modularity, and complexity) were calculated in source networks as in the sensor-based networks.
There was no power calculation performed a priori. This is a retrospective study with secondary analysis of δ-oscillatory properties in infants aged 0 to 3.9 months. From our large data set of more than 100 infants, we identified a subcohort of infants who were specifically aged 0 to 3.9 months and had acceptable electroencephalography data; this yielded 20 subjects.
Statistics were used for comparing local connectivity changes, changes in modularity and changes in complexity. Statistical analysis was performed using custom-written MATLAB code.
First, for local connectivity changes, statistical comparison between pairs of connections was performed using the Network Based Statistic Toolbox (http://sites.google.com/site/bctnet/comparison/nbs; accessed February 1, 2019).29 Because of the mass-univariate testing, Network Based Statistic Toolbox controls the family-wise error rate by identifying “clusters” of suprathreshold statistically significant connections. Each connection is associated with a test statistic quantifying evidence with respect to whether or not it favors the null hypothesis. Then a test statistical threshold is chosen; connections with a test statistic value exceeding this threshold or suprathreshold connections are considered. In turn, if there is a connected component (from a graph perspective) of suprathreshold statistically significant connections, a P value is assigned by indexing its size with the null distribution of maximal component size (the latter is derived using permutations at which the group to which each subject belongs is randomly exchanged). We identified statistically significant connections using t tests with contrast emergence greater than anesthesia and with a statistical t test threshold of 3, a number of permutations of 5,000, and statistical significance level of 0.05 for the permutations. Visualization of the significant edges was produced using the in-built function of the Network Based Statistic Toolbox. The function uses the regions of interest coordinates as input and produces a network with significant edges and how these project in a standard brain template. The results are also presented using the false discovery rate method. The false discovery rate enables rejection of the null hypothesis at the level of individual connections (as opposed to clusters of connections in the way that the Network Based Statistic Toolbox does) and does not require the use of a prespecified statistical threshold.30
Second, within-subject comparisons were performed to evaluate the difference in complexity and in modularity in the maintenance period compared with the emergence period. Normality tests were performed using the Kolmogorov–Smirnov test. Because modularity and complexity data were not normally distributed, we used a paired Wilcoxon signed rank test (two-tailed). A P value of less than 0.05 was considered statistically significant. CI values were computed via bootstrapping. Specifically we sampled with replacement from the two samples, and we calculated the difference in the medians. This process was repeated 10,000 times. The naïve 95% CI was calculated using the 25th and 97.5th largest median differences. The box plot data are shown as median values alongside the 1.5 interquartile range.
We used multichannel scalp electroencephalograph recordings in infants undergoing sevoflurane general anesthesia for elective surgery. Continuous multichannel electroencephalography recordings were collected during maintenance of a surgical state of anesthesia and at emergence (first body movement) in 20 subjects aged 0 to 3.9 months old. End-tidal sevoflurane (end-tidal sevoflurane) was between 1.8 to 2.5% during the maintenance phase and 0 to 0.3% during the emergence phase. Electroencephalography functional connectivity was obtained using the cross-spectral coherence between sensor signals or region of interest source-reconstructed signals. For each infant and two conditions (emergence and anesthesia), we obtained a functional connectivity graph or network by keeping the strongest connectivity values between all sensors and regions of interest for sensor and source level analysis respectively. Subject demographics and clinical characteristics are provided in table 1. Details concerning the study design and the relevant methods are given under “Materials and Methods” and in figure 1.
Functional Connectivity at the Sensor and Source Levels
Power spectral analysis indicated that δ power (1 to 4 Hz) was the dominant frequency component during maintenance and emergence from anesthesia in all infants (n = 20; fig. 2). During maintenance, peak power was observed at the mean frequency of 1.25 Hz ± 0.62 Hz (mean ± SD). During emergence, peak power was observed at 1.08 Hz ± 0.29. Based on this key feature, δ-oscillation functional connectivity networks properties were evaluated during anesthesia and emergence in subsequent analyses.
At the network local scale, we looked at differences in δ-oscillation connectivity values between maintenance and emergence from anesthesia. Connections that survived statistical significance are presented in figure 3. We observed a decrease between frontal and central/parietal electrodes when comparing anesthesia with emergence (fig. 3A). The full list of edges is presented in table 2. Similar results obtained using the false discovery rate method are presented in table 3. At the source level, we observed decreased connectivity during anesthesia with statistically significant changes in the connectivity within the posterior cingulate with right frontal and parietal regions (fig. 3B). The full list of edges for which connectivity was significantly decreased during anesthesia is presented in table 4. Similar results obtained using the false discovery rate method are presented in table 5.
Modularity at the Sensor and Source Levels
At the network meso scale, δ-oscillation modularity was estimated for the functional connectivity networks at anesthesia and emergence. Modularity is a measure of functional segregation, where large values indicate scattering between many communities or networks, and low values indicate high network integration. We found that modularity during anesthesia was increased when compared with emergence (fig. 4A; Wilcoxon signed rank test z value, 2.53; P = 0.011; difference in medians, 0.0562; 95% CI, 0.0048 to 0.1298). This suggested that networks became more fragmented during anesthesia, thus increasing segregation between different parts of the brain. Similar results were obtained when we remontaged the data to the average reference (Wilcoxon signed rank test z value, 2.15; P = 0.032). This result was confirmed at the source level where we observed that modularity increased during anesthesia compared with emergence (fig. 4B; Wilcoxon signed rank test z value, 2.31; P = 0.021; difference in medians, 0.0548; 95% CI, −0.0040 to 0.1074).
Complexity at the Sensor and Source Levels
We then looked at complexity as a global measure of information integration. δ-Oscillation complexity decreased when comparing anesthesia with the emergence phase (fig. 5A). The change in complexity between anesthetic states was statistically significant (Wilcoxon signed rank test z value, 2.01; P = 0.044; difference in medians, 0.0293; 95% CI, −0.0016 to 0.0397). Similar results were obtained when we remontaged the data to the average reference (Wilcoxon signed rank test z value, 1.9700; P = 0.049).
Complexity was further evaluated for each infant at the source level. δ-Oscillation complexity decreased during anesthesia compared with emergence corroborating the results observed at the sensor level (fig. 5B; Wilcoxon signed rank test z value, 2.50; P = 0.012; difference in medians, 0.0201; 95% CI, −0.0025 to 0.0482). Collectively, functional connectivity was reorganized during maintenance anesthesia, with network connectivity becoming less complex.
Summary of Findings
In this study, we described functional connectivity network changes during anesthesia-induced loss of consciousness in early postnatal development. We examined brain state transitions from the maintenance phase of anesthesia to emergence from anesthesia and quantified modulations of connectivity patterns. Our findings indicated that infant general anesthesia is potentially (1) driven by connectivity changes in key default mode network and frontoparietal regions (fig. 3), (2) associated with more segregation (increase in δ-oscillation modularity; fig. 4); and (3) related to reduction in δ-oscillation complexity (fig. 5), at both the sensor and source levels.
Studies of Electroencephalography Functional Connectivity in Adults under General Anesthesia
A growing number of studies have indicated differences in functional connectivity during general anesthesia31,32 and also at different stages of brain development.33,34 Deeper stages of anesthesia are characterized by reduced functional connectivity in higher-order brain networks such as the default mode network and the frontoparietal network.10,12,35 Crucially, emergence from anesthesia is associated with restoration of functional connectivity between frontoparietal regions and subcortical regions.36
Further evidence for the reconfiguration of functional connectivity in adult anesthesia is evident in studies investigating alterations in graph-theoretic properties. At the local scale, previous studies have shown disruption of highly connected regions located in the default mode network and frontoparietal networks.37,38 At the meso and whole-brain scales, α-based electroencephalography networks in adults during propofol-induced anesthesia were shown to be more fragmented and less efficient.17 Overall, α-oscillation functional connectivity in the anesthetized adult brain is less globally efficient and more segregated as a result of attenuated connectivity between higher-order networks.
Studies of Electroencephalography Functional Connectivity of the Infant Brain under General Anesthesia
However, there are few studies describing alterations of functional connectivity in infants. Previous studies analyzing electroencephalography power in infants less than 6 months of age showed no difference in power between anesthesia and emergence.39,40 Previous studies showed that although δ power was prominent, there were no differences in δ power in infants 0 to 3.9 months when comparing the awake state with anesthesia and between anesthesia and first body movement (a surrogate of emergence).41 In addition, functional connectivity (coherence) between frontal electrodes was weak in δ frequencies in infants 0 to 3.9 months compared with older infants.11
A Mechanistic Explanation for Loss of Consciousness in the Infant Brain
We thus argue that power cannot discriminate between young infants’ brain responses during anesthesia and brain responses during emergence. In contrast, graph-theoretic approaches can provide a holistic representation of how different parts of the brain integrate and how this integration is affected by anesthesia.9 Motivated by this, we asked whether δ-oscillation whole-brain connectivity would provide more information with respect to the infant brain dynamics. Our results showed that brain connectivity was less complex during anesthesia compared with emergence. Complexity is an aggregate measure of how connectivity is distributed across the brain. It captures the coexistence of hubs and other sparsely connected regions that, together, provide a balance between segregation and global integration.21 Reduced complexity implies a more segregated configuration that promotes local efficiency in communicating information and inhibits global integration.10 This result was corroborated at the meso scale where we found that infant brain networks during anesthesia were more segregated into clusters compared with emergence.
How does the shifting to a less complex brain network occur? In light of our findings at the local scale, it is possible that such shifting might take place because of reduction in connectivity between default mode network and frontoparietal regions. Frontoparietal connectivity is modulated by anesthetic-induced loss of consciousness.36 In addition to its relationship with the anesthetic effect, frontoparietal connectivity is important for loss of consciousness because of its extensive connectivity to the rest of the brain.37 We observed loss of connectivity in frontoparietal regions and distant parts of the brain, most notably the posterior cingulate. In that regard, long-range connections are crucial in communicating information between distant regions, thus increasing global integration in the network. Specifically, long-range connections from the posterior cingulate, a key region of the default mode network, to the frontoparietal regions have been deemed important in regulating communication in the whole brain allowing the cross-talk between different specialized regions.42 Therefore, it is possible that impairment of long-range connections could be linked to loss of consciousness in the sense that their alteration causes the whole-brain network to become more disconnected.10
The Role of δ-Oscillations in Loss of Consciousness
The results reported here were for δ-oscillations based on the limited range of oscillatory frequencies generated in the human infant brain in early postnatal development. There is no clear consensus as to how δ-oscillations are produced. They can be generated cortically43 or from parts of the thalamus such as the thalamic reticular nucleus.44 Electrophysiologic studies in adult rodents show that propofol-induced coherent α- and δ-oscillations develop rapidly at loss of consciousness, appearing to mediate the functional disruption of thalamus and cortex, and disappear in a spatiotemporal sequence during emergence from anesthesia.45 However, their role in loss of consciousness in humans is still exploratory. Alterations in δ-oscillations have been shown in intracortical recordings in patients under propofol-induced anesthesia.46 In this study, it was conjectured that loss of consciousness would be associated with whole-brain network fragmentation, whereas local network structure would remain intact.47 Toward this direction, we have provided evidence for δ-based whole-brain alterations in the infant brain with δ-based connectivity becoming less complex and more fragmented.
Clinical Implications for Brain Monitoring under Anesthesia
General anesthetics and hypnotic agents such as sevoflurane, propofol, ketamine, and dexmedetomidine produce stereotyped electroencephalography oscillations that relate fundamentally to neural circuit architecture and function in adults.48 Spectral features of the electroencephalography can be used to monitor brain activity during anesthesia and guide anesthetic dosing. With respect to this, many academic publications, as well as proprietary algorithms used for Bispectral Index, SedLine, and other commercial monitors, place emphasis on frontal α spectral power to monitor the transitions into and recovery from anesthetic-induced unconsciousness.47 Previous studies from our group show an association of α power with an anesthetic state beginning around age 4 to 6 months.11,13,49 However, infants 0 to 3.9 months of age show very little α power, with most of the overall power concentrated in the slow frequency range. In this age range, power spectra during a surgical state of anesthesia (e.g., sevoflurane concentration, 1.8 to 2.5%) shows no discernible difference from power spectra around return of gross body movement (sevoflurane concentration less than 0.4%), unlike in older infants and young children.49 Conversely, studies of age dependence of minimal alveolar concentration required to prevent response to surgical incision have found very weak dependence on age, at least for 0 to 6 months of age.50 It is therefore clinically relevant to determine whether, for infants 0 to 3.9 months of age undergoing surgery, more sophisticated electroencephalography analyses might distinguish features of brain function during a surgical state of anesthesia compared with emergence.
In that regard, combining our results with the literature in infants and adults shows that whole-brain connectivity and its properties at the respective dominant frequencies (for example δ-oscillations for infants and α oscillations for adults and infants) can discriminate between different levels of consciousness. In anesthetized infants, where power is indiscriminable between anesthesia and emergence, a connectivity-based marker might assist perioperatively with monitoring anesthetic depth. With the increased sophistication of hardware and software technology for real-time electroencephalography monitoring of general anesthesia, it is now possible to perform spectral analysis of electroencephalography data online, thus paving the way for real-world application of electroencephalography network-based markers for monitoring infant anesthesia. Similar efforts have begun to emerge in classifying patients with disorders of consciousness,51,52 showing that electroencephalography connectivity can directly contribute to consciousness-level dependent brain monitoring.
Study Limitations and Constraints
It is possible that the observed electroencephalography features could have been confounded by systematic age-related differences in drug administration; further studies will need to address this effect. Although challenging to pursue because of the idiosyncratic nature of the experimental setup (anesthetic management was administered according to the anesthesiologist’s discretion, rather than in a controlled, titrated fashion), there is a need for more detailed electroencephalography studies including controlled sevoflurane administration at a wider range of concentrations and over longer periods of time.36,53 Limitations regarding source reconstruction methods apply to this study54 ; we attempted to alleviate these by focusing on cortical regions only and using age-matched templates to obtain realistic electrophysiologic models for brain activity in the infant head. Finally, the limited sample size and number of clean segments prevented us from further looking into individual variability across time and anesthetic concentrations; thus, our plan is to generalize the findings using a larger cohort.
General anesthetics modulate functional connectivity and reduce brain network integration. We showed that such a process takes place even in young infants where δ power is dominant. Thus connectivity can become an important tool for assessing anesthetic depth in the very young infant brain even in the absence of α-oscillations.
The authors thank the preoperative and operating room staff, Boston Children’s Hospital, Boston, Massachusetts, for their assistance during these studies, as well as the families who took part in the study. The authors also thank Ann-Marie Bergin, M.B., Sc.M., M.R.C.P., Department of Neurology, Boston Children’s Hospital, for reviewing all electroencephalogram recordings for potential incidental findings.
Supported by Downing College, University of Cambridge through a Treherne Studentship (Cambridge, United Kingdom; to Dr. Pappas), Boston Children’s Hospital (Boston, Massachusetts; to Dr. Cornelissen), International Anesthesia Research Society Mentored Research Award (San Francisco, California; to Dr. Cornelissen), a Sara Page Mayo Endowment for Pediatric Pain Research and Treatment (Boston, Massachusetts; to Dr. Berde), the National Institute for Health Research through the Cambridge National Institute for Health Research Biomedical Centre and a Senior Investigator Award (Cambridge, United Kingdom; to Dr. Menon), the Canadian Institute for Advanced Research Brain, Mind and Consciousness Program (Toronto, Canada; to Dr. Menon), and a Stephen Erskine Fellowship from the University of Cambridge (Cambridge, United Kingdom; to Dr. Stamatakis).
The authors declare no competing interests.