Cortical Areas Associated With Mismatch Negativity: A Connectivity Study Using Propofol Anesthesia

Auditory mismatch negativity (MMN) is an event-related potential (ERP) waveform induced by rare deviant stimuli that occur in a stream of regular auditory stimuli. The generators of MMN are believed to include several different cortical regions like the bilateral temporal and the right inferior frontal gyrus (IFG). However, exact cortical regions associated with MMN remain controversial. In this study, we compared the number of long-distance connections induced by the standard and deviant stimuli during awake state and propofol anesthesia state to identify the cortical areas associated with the generation of MMN. In awake state, we find that deviant stimuli synchronize more information between the right frontal and temporal than standard stimuli. Moreover, we find that the deviant stimuli in awake state activate the bilateral frontal, central areas, the left temporal and parietal areas as compared to the anesthesia state, whereas the standard stimuli do not. These results suggest that, in addition to the bilateral temporal and the right IFG, the bilateral frontal and centro-parietal regions also contribute to the generation of MMN.


INTRODUCTION
Auditory MMN is an ERP waveform induced by rare deviant stimuli that occur in a stream of regular auditory stimuli (Näätänen et al., 2001;Garrido et al., 2009a;Paavilainen, 2013). These deviant stimuli can be obtained by changing the frequency, intensity, duration, or other auditory characteristics of the standard stimuli. The peak of MMN usually occurs at 100-200 ms after stimuli with maximal amplitudes over fronto-central scalp regions.
Currently, the predictive coding is the leading hypothesis for explaining the neural mechanism that underlies the generation of MMN (Garrido et al., 2009b). It posits that the generation of MMN involves a hierarchical neural network interconnected with recurrent connections (Schmidt et al., 2013). The low-level brain regions are responsible for the abstraction of the auditory input, whereas the high-level brain regions are responsible for predicting the upcoming stimulus (Friston, 2005). MMN is generated when an input violates the top-down prediction. Moreover, the repeated standard stimuli which reduces the strength of bottom-up connection could also contribute to the observed MMN waveform (Garrido et al., 2009b).
The dynamic casual modeling (DCM) is the primary tool to test the predictive coding hypothesis (Garrido et al., 2008). In DCM, a model with pre-selected brain regions is constructed to fit the observed MMN waveform obtained from neuroimaging data (Garrido et al., 2009a). An accurate DCM model rests on the properly assigned cortical generators of MMN (Friston, 2003). Most of studies using DCM selected the bilateral superior temporal gyrus and the right inferior gyrus to fit the measured MMN waveform (Garrido et al., 2008(Garrido et al., , 2009aCooray et al., 2016;Ranlund et al., 2016). In addition, experimental evidences suggest that the auditory cortex and the frontal cortex are both activated during MMN (Rinne et al., 2000;Opitz et al., 2002;Doeller et al., 2003). The involvement of bilateral superior temporal gyrus has been increasingly supported by experimental evidences, however, the role of the frontal cortex generators remains unclear (Alho, 1995;Opitz et al., 2002;Doeller et al., 2003;Rosburg, 2003;Molholm et al., 2005;Rosburg et al., 2005). Therefore, it is of great significance to locate the exact generators of MMN, which may lead to an accurate DCM model.
A typical MMN waveform composes of a temporal component and a frontal component (Baldeweg, 2007). The temporal component of MMN is originated from the auditory cortex and it represents the auditory feature extraction process, whereas the frontal component may be caused by the high level cognitive function such as attention switch (Joos et al., 2014). It has been shown that as compared to standard stimuli, the deviant stimuli induced more brain activation in both temporal lobe and frontal lobe (Molholm et al., 2005). Hence, we can infer that the deviant stimuli are considered as more cognitively demanding as compared to the standard stimuli.
The workspace theory predicts that cognitive demanding tasks will result in an increased information integration among discrete cortical regions (Dehaene and Naccache, 2001;Baars, 2002). Therefore, according to the workspace theory, the cortical areas that are responsible for the generation of MMN can be identified by comparing the cortical regions activated by standard and deviant stimuli respectively. One of the early attempts found that deviant stimuli induce increased long-distance information synchronization among discrete cortex regions as compared to standard stimuli (Nicol et al., 2012). However, they only analyzed data on sensor level which prohibited them to pinpoint the cortical regions that are responsible for the generation of MMN.
An alternative approach to identify the cortical associated with the generation of MMN is to abolish the high-level cognitive function that is required for generating MMN during the presentation of deviant stimuli. Various anesthetic agents can achieve this target by blocking the information transmission across multiple brain regions (Lee et al., 2009(Lee et al., , 2013Ku et al., 2011). It has been shown that the primary auditory perception of the brain is still active under the deep sedation (Ku et al., 2011;Lee et al., 2013), whereas the auditory stimulation processing is totally broken down at deep level of anesthesia (Heinke et al., 2004;Koelsch et al., 2006).
In the present study, we hypothesized that, based on the workspace theory, deviant stimuli would induce increased longdistance connectivity across distinct brain regions as compared to the standard stimuli, and these connections should be blocked by propofol induced anesthesia. To test this hypothesis, we recorded EEG signal during an MMN paradigm under awake and propofol-induced anesthesia states, respectively. Then, source analysis was conducted to identify the cortical regions with different long/short-distance connections between the brain networks involved in processing the two kinds of stimuli under both awake and propofol-induced anesthesia states. Our findings might provide further evidence on the cortical regions associated with MMN and shed lights on the generation mechanism of MMN.

Subjects
This study was approved by First Affiliated Hospital of Xi'an Jiaotong University and complied with the guidelines of the Declaration of Helsinki. Written informed consent was obtained from all participants. Totally, 25 healthy subjects (age = 35.5 ± 5.3) participated in this study. All the subjects are drug-free and allergy-free to propofol. Exclusion criteria included a history of neurological, psychiatric or hearing diseases. Each participant was asked to fast for at least 8 h before propofol administration.

Data Acquisition
Scalp EEG data was acquired from 64 Ag/AgCl electrodes which were placed according to the International 10-10 system. The reference electrodes were placed at bilateral mastoids. Two electrodes were placed above and below the left eye to record the vertical electro-oculogram (VEO). The impedance of all electrodes was kept below 5 kΩ. The EEG data was recorded using SynAmps amplifier (Neuroscan Inc.,) with a sampling rate of 1,000 Hz.

Stimuli
The MMN paradigm contained three kinds of pure sinusoidal tones with different pitch: a standard tone (500 Hz, n = 200) and two deviant tones (450 Hz, n = 50; 550 Hz, n = 50). There are 300 randomly ordered tones in each block. The duration of tones was 75 ms with a 10 ms rising/falling time, and the inter-stimulus interval between each stimulus was 1000 ms. E-prime software (Pittsburg, PA, USA) was used to design and present these pure tones (80 dB) through a pair of earphones (Sony, MDR EX450).

Anesthesia Infusion
All the experiments were performed by two experienced anesthesiologists. Propofol was employed as the sole anesthetic in the experiment and it was administered by a target-controlled infusion device (TCI, Injectomat TIVA Agilia). During the experiment, vital signs including heart rate, blood pressure, respiratory rate and pulse oximetry (SpO 2 ) were continuously monitored by using intra-operative monitors (MP50, PHILIPS). The anesthesia depth was tracked by using the bispectral index monitor. The bispectral index (BIS) is a processed EEG parameter and it has been widely adopted in the clinical settings for assessing the level of anesthesia during surgery (Myles et al., 2004). The BIS value ranges from 0 (flat line EEG) to 100 (fully awake). In the present study, the awake state was defined as the BIS value is equal or >90, whereas the deep anesthesia state is defined as the BIS within the range of 35-45.

Procedure
Subjects comfortably laid on the bed after EEG cap and headphone were set. They were asked to close their eyes and stay relaxed during the experiment. The experiment was carried out following the procedures as below: (1) Awake state: In this state, subjects were asked to stay awake with eye closed. 2 min resting EEG data and 4 min ERP data were collected.
(2) Anesthesia: The propofol plasma target concentration was first set at 1-2 µg/ml. A laryngeal was placed after subjects lost their consciousness and connected to a ventilator, which was set to the SIMV mode. The tidal volume was set at 8-10 ml/kg. The respiratory rate was 12 breaths/minute. The endtidal carbon dioxide partial pressure was maintained at 35-45 mmHg and the plasma propofol target concentration was set at 4-4.5 µg/ml. 2 min resting EEG data and 4 min ERP data were collected while the BIS value was maintained around 40.
(3) Recovery of anesthesia: The reverse protocol was initiated by setting the plasma concentration of propofol to 0 µg/ml. Then, the subjects were pushed into the anesthesia recovery room to wait for the complete recovery.
In this study, we analyzed the 4 min ERP data from each state, the resting data were kept for other studies.

EEG Preprocessing and ERP Analysis
The EEG signal processing pipeline was given in Figure 1. Raw EEG signal was first band-pass filtered at 0.5-40 Hz to remove baseline drift and power line noise. The segments of data contain excessive fluctuation and eye movement were manually removed. Epochs of standard stimuli (500 Hz) and deviant stimuli (450 Hz; 550 Hz) were extracted with 100 ms before to FIGURE 1 | EEG data analysis pipeline.

Source Reconstruction
Since long-range synchronization among cortical regions was mostly presented in the high frequency band (Nicol et al., 2012), only gamma band was considered in the present study. The clean epochs of EEG signal were first band-pass filtered in the frequency band of gamma band (30-40 Hz). Then, source reconstruction was performed to localize cortical activation by using Brainstorm (Tadel et al., 2011). The MNI model (Colin27) (Collins et al., 1998) was used as the standard head model and the standard Neuroscan Quik-Cap64 registration channel file was used as the electrode registration. The forward model was first calculated using the OpenMEEG boundary element method (Gramfort et al., 2010). We selected the weighted minimum norm estimator (wMNE) (Baillet et al., 2001) to estimate the activation of 15002 dipoles distributed over the entire cortex. Then, these dipoles were divided into 148 regions of interest (ROI) based on the Destrieux atlas (Destrieux et al., 2010). Activity of each ROI was formed by averaging the current source density of all voxels within that region. These 148 ROIs were further divided into 14 areas according to their anatomical location on the cortex including bilateral prefrontal(LPF/RPF), bilateral frontal(LF/RF), bilateral central(LC/FC), bilateral temporal(LT/RT), bilateral occipital(LO/RO), and bilateral cingulum(LCing/RCing).

Network Analysis
Recent studies have shown that the data segment around the maximum amplitude of ERP or MMN is suitable for studying the brain functional network related to its generation (Nicol et al., 2012;Karamzadeh et al., 2013). Hence, we selected the data segment according to the maximum amplitude of MMN that is identified for each subject at the awaken state for constructing the brain functional network. Furthermore, it has been shown that the anesthetics can attenuate the amplitude of ERP, whereas the latency of the ERP is left unchanged (Ozgoren et al., 2010). Therefore, we extracted EEG data in the anesthetic state from the same time window as identified in the awaken state for estimating the brain functional connectivity in the absence of observable MMN waveform. In doing so, the difference in functional connectivity at the same time frame relative to the onset of stimuli between awaken and anesthetic state would reflect the changes that is related to the generation of MMN. The brain activity in a 100 ms temporal window which is centered at the peak of MMN waveform was extracted (marked blue in Figure 2) to construct the brain network. The 148 ROIs were defined as nodes and the connectivity strength between the nodes were defined as edges.
The weighted phase lag index (wPLI) was employed to measure the connectivity between a pair of ROIs. The reason of selecting wPLI is due to its robustness against the influence of volume conduction (Vinck et al., 2011;Lau et al., 2012). The wPLI was defined as: where n is the number of trials, t is the index of trial, S ijt represents a vector of phase differences between region i and region j at each time point within the t-th trial. The phase of brain activity of each ROI was computed from the signal x(t) and its corresponding Hilbert transformx(t).
The obtained wPLI ranges from 0 to 1 in which 1 indicates signal i always leads or lags signal j in all trials and it equals to 0 indicates Frontiers in Human Neuroscience | www.frontiersin.org that the phase difference between two signals are completely random across trials.
To prevent false connections, the obtained wPLI of all possible pairs were pruned by preserving the strongest connections above a threshold. Previous research showed that the average degree of networks must be larger than 2 × ln N to maintain the integrity of network, where N is the number of nodes in the network (Watts and Strogatz, 1998). In the present study, the corresponding lowest threshold is about 2×ln 148 = 6%. Hence, we constructed networks with a wide range of connection densities ranging from 6 to 90%.

Long/Short-Distance Connections
In source analysis, we grouped the 148 ROIs into 14 cortical areas according to its anatomic location. Similar to the previous work (Kitzbichler et al., 2011;Nicol et al., 2012), we then defined the functional connections among these cortical areas as the longdistance connections, and the connections within each cortical area as the short-distance connections.
The weighted value of long/short connections was calculated as the sum of the corresponding wPLI values in the connection matrix. The long-distance connection of the whole brain network was given as The short-distance connection of the whole brain network was defined as: where k ∈ [1, 2, 3, ..., 14] is the index of the 14 cortical areas, L k and S k represent the connection strength of long/short-distance connects, respectively. The ratio between the value of the longdistance connections and the short-distance connections was defined as We first compared the Ratio of the brain networks involved in processing the two kinds of stimuli in awake state. If there was a statistically significant difference on Ratio, then we compared L all and S all obtained from two kinds of stimuli to find out the cause of the difference in Ratio. Then, we compared the obtained measures between awake and anesthesia states. Statistical comparison of the number of long/short-distance connections and Ratio between the two states and between the two kinds of stimuli were conducted by using paired t-test. P < 0.05 was considered as statistically significant.

MMN
Fourteen out of 25 subjects showed a typical MMN waveform during awake state, the other 11 subjects with no clear MMN waveform were excluded from further analysis. The obtained MMN waveform from three representative subjects at Fz electrode during awake state was provided in Figure 2. It can be observed that both standard stimuli and deviant stimuli induced a negative going response after the stimuli. The MMN is the negative differential response between deviant stimuli and standard stimuli at the Fz electrode. As shown in Figure 3, the observed MMN wave vanished at anesthesia state.

Long/Short-Distance Connections
As shown in Figure 4A and Table 1, in awake state, the Ratio between the long/short-distance connections induced by deviant stimuli was significantly higher than that induced by standard stimuli. However, no difference was found during anesthesia state. It was noticed that the Ratio of two kinds of stimuli were both higher during anesthesia than during awake state. We then compared the difference of the S all and the L all of two kinds of stimuli during awake state and anesthesia to investigate the cause of the increase in Ratio. As shown in Figures 4B,C and Table 2, both S all and L all were decreased under anesthesia. However, the reduction on S all was significantly higher than L all , which results in an increase in Ratio during anesthesia state.
Since the changes in Ratio between the standard and deviant stimuli might be caused by the changes of the L all or the S all , we compared the L all and the S all of the two kinds of stimuli in both awake state and anesthesia state. We found that in awake state, the L all of deviant stimuli was significantly higher than the standard stimuli, while the S all of these two kinds of stimuli had no significantly difference ( Figure 5A and Table 3). This finding suggested that increased Ratio induced by deviant stimuli was mainly caused by the increase of long-distance connections. Furthermore, we compared the differences in the number of long-distance connections of each cortical area to identify which region contributed the most to the increased L all of deviant stimuli. As shown in Figure 5B and Table 4, deviant stimuli induced more long-distance connections in the right prefrontal, central and bilateral temporal areas as compared to the standard stimuli. On the other hand, there were no statistically significant differences in L all , S all , and L k at each cortical region between the standard and deviant stimuli during anesthesia state (Figures 5C,D, Tables 3, 4).
We also compared the number of long-distance connections induced by standard stimuli and deviant stimuli between the awake and anesthesia states. As compared to the state of anesthesia, standard stimuli in the awake state induced greater number of long-distance connections at left prefrontal and right central areas ( Figure 6A and Table 5). In contrast, the number of long-distance connections induced by deviant stimuli were higher at the bilateral frontal, central, parietal and cingulate areas in awake state ( Figure 6B and Table 5).

DISCUSSION
In this study, we compared the number of long-distance connections induced by the standard and deviant stimuli during awake state and propofol induced anesthesia state to identify the cortical areas which are associated with the generation of MMN.
We found that in the awake state, deviant stimuli induced a greater number of long-distance connections than standard stimuli due to its high demand on cognitive function. This result was in consistent with the hypothesis of the global workspace theory (Kitzbichler et al., 2011;Nicol et al., 2012). The increment of long-distance connections was mainly distributed over the right prefrontal, bilateral temporal areas, which is in consistent with the previous studies (Rinne et al., 2000;Opitz et al., 2002;Doeller et al., 2003). Additionally, we found that the right central area also contributed to the increment of long-distance connections induced by deviant stimuli.
By comparing the measurements of brain networks involved in processing the two kinds of stimuli between awake and anesthesia state, we found that deviant stimuli induced a greater number of long-distance connections at the bilateral prefrontal, central, parietal and cingulate areas in awake state than that in anesthesia state. However, significant difference between two states induced by standard stimuli only existed on two cortical regions: left prefrontal and right parietal areas. As shown in Figure 5C, no differences existed between the brain network responsible for processing the standard stimuli and deviant stimuli during anesthesia. Hence, we could infer that during anesthesia, the brain networks involved in processing the two kinds of stimulus were suppressed to similar level. Furthermore, there were no significant difference in the number of longdistance connections induced by standard stimuli between the awake state and anesthesia state, while the number of longdistance connections induced by deviant stimuli was significantly suppressed during anesthesia state as compared to the awake  (Figures 4B,C). Hence, the cortical regions which may contributed to the differences of the long-distance connections induced by the deviant stimuli between the awake state and anesthesia state can be considered as the cortical regions associated with the generation of MMN.
Our results revealed that in the awake state, deviant stimuli induced more long-distance connections at bilateral temporal regions. However, the number of long-distance connections induced by deviant stimuli had no significant difference between the two states at bilateral temporal regions. There were several studies suggested that the cortical regions that are responsible for higher-level cognitive process is more sensitive to anesthesia as compared to the lower-level sensory cortex. The temporal cortex is responsible for the perceptual function of the auditory signal, which belongs to the low-level cortex and it is less sensitive to anesthesia, hence the number reduction on longdistance connections at temporal cortex during anesthesia is not as obvious as observed in other cortical areas.
The temporal cortex is responsible for forming sensory memory model, while the frontal cortex is responsible for comparing the memory model to the input stimuli. The sensory memory model will be adjusted when the frontal cortex finds that the current stimulus and the model does not match, and this process would result in the MMN (Garrido et al., 2009b). Our results suggested that the frontal regions were closely related to the generation of MMN, which is in line with the prediction of the predictive coding theory. In other words, the generation of MMN relies on the communication between the frontal and temporal cortex.
Several studies have suggested that the involvement of frontal lobe might be related to a specific MMN paradigm (Näätänen and Alho, 1997;Shestakova et al., 2002;Molholm et al., 2005). The right IFG was activated in the frequency condition, whereas the activation of the left IFG was observed in the duration condition. Most recent researches have selected bilateral temporal and right frontal cortex as the prior model to analyze the MMN response to pitch by using DCM (Garrido et al., 2008;Cooray et al., 2016;Ranlund et al., 2016). In this study, we found that the MMN to pitch did not only activate the right IFG, but also activate the left prefrontal. This finding is in line with previous studies (Lin et al., 2007;Wehner et al., 2007;Hsiao et al., 2010;Cheng et al., 2013).
Our results also suggested the parietal region (Kasai et al., 1999;Schall et al., 2003;Marco-Pallarés et al., 2005;Molholm et al., 2005) and the central region were associated with the generation of MMN (Molholm et al., 2005;Novitski et al., 2006;Fulham et al., 2014). The involvement of the parietal cortex might be supported by the anatomical connection from the parietal and prefrontal to the temporal cortex (Makris et al., 2005;Schmahmann et al., 2008;Kamali et al., 2014). Several fMRI and MEG studies have found that the fronto-temporo-parietal network might play a role in generating MMN (Schall et al., 2003;Cheng et al., 2013). Our results suggested that the generation of MMN activated a wide fronto-temporo-parietal network.
There are several limitations of our study. We used the standard Colin27 head model as the standard head model of every participant in source analysis, which might slightly bias the accuracy of the source reconstruction. Moreover, the EEG signal only record the neural activities on the scalp, the activation of the inner cerebral hemisphere, such as cingulate gyrus, cannot be accurately observed even by using the source analysis. Hence, we excluded the involvement of cingulate gyrus during the generation of MMN.
In conclusion, by comparing the brain networks involved in processing the standard and deviant stimuli in the MMN paradigm during awake and anesthesia state, we observed the increased number of long-distance connections originate from the bilateral prefrontal, temporal, central and parietal cortices, which suggested that the generation of MMN would require communications among prefrontal, temporal, and centroparietal regions.

AUTHOR CONTRIBUTIONS
YZ processed the data and wrote the manuscript. FY recruited subjects, performed data collection, and revised the manuscript. LW performed the data collection and processing and wrote the manuscript. YW performed data processing and revised the manuscript. CW performed the experiments. QW was responsible for designing the study and manuscript editing. LH designed the study and reviewed the manuscript. All authors contributed to the completion of the manuscript.