Breakdown of effective information flow in disorders of consciousness: Insights from TMS-EEG

Background: The complexity of the neurophysiological mechanisms underlying human consciousness is widely acknowledged, with information processing and flow originating in cortex conceived as a core mechanism of consciousness emergence. Combination of transcranial magnetic stimulation and electroencephalography (TMS-EEG) is considered as a promising technique to understand the effective information flow associated with consciousness. Objectives: To investigate information flow with TMS-EEG and its relationship to different consciousness states. Methods: We applied an effective information flow analysis by combining time-varying multivariate adaptive autoregressive model and adaptive directed transfer function on TMS-EEG data of frontal, motor and parietal cortex in patients with disorder of consciousness (DOC), including 14 vegetative state/unresponsive wakefulness syndrome (VS/UWS) patients, 21 minimally conscious state (MCS) patients, and 22 healthy subjects. Results: TMS in DOC patients, particularly VS/UWS, induced a significantly weaker effective information flow compared to healthy subjects. The bidirectional directed information flow was lost in DOC patients with TMS of frontal, motor and parietal cortex. The interactive ROI rate of the information flow network induced by TMS of frontal and parietal cortex was significantly lower in VS/UWS than in MCS. The interactive ROI rate correlated with DOC clinical scales. Conclusions: TMS-EEG revealed a physiologically relevant correlation between TMS-induced information flow and levels of consciousness. This suggests that breakdown of effective cortical information flow serves as a viable marker of human consciousness. Significance: Findings offer a unique perspective on the relevance of information flow in DOC, thus providing a novel way of understanding the physiological basis of human consciousness.


Introduction
Information processes in the brain are closely linked to its anatomical and functional structure.Although there is still ongoing debate, the mainstream consciousness theories more or less all relate the emergence of human consciousness to brain information flow [1].For instance, the Global Workspace Theory hypothesizes that consciousness emerges by information processing, the more the information of a state is globally available, the more that state can be consciously accessed [2].The Information Integration Theory highlights the crucial role of information integration in consciousness function, the more information integrated by a system, the higher the level of consciousness entailed by it [3].These hypotheses form a commonly held view that information flow between widely separated regions of the cerebral cortex is a necessary component in the generation of human consciousness [4,5].Thus, in order to understand how consciousness emerges or disappears, information flow networks underlying manifestations of consciousness must be elucidated.
The combination of transcranial magnetic stimulation and electroencephalography (TMS-EEG), a non-invasive technique, provides a way to explore electrophysiological processes in human cortex, permitting assessment of brain states, such as global function, regional responsiveness, and cortico-cortical interactions [6].In contrast to TMS-evoked electromyography and peripheral event-related-potentials, TMS-EEG does not require the intactness of peripheral neural pathways and does not necessitate any participation from the patient, enabling its broad application in clinical populations and making it a viable option to conduct bedside studies [7].It directly probing consciousness related neural circuits without any task-or stimulus-related confounding factors, is seen as a promising tool to gain insight into the mechanisms of consciousness fading and emerging [8,9].In unconsciousness states induced by sleep or anesthesia, TMS-EEG has revealed simplified neural response patterns and a global break-down of effective connectivity, compared to when consciousness is fully present [10,11].These findings have closely linked cortical reactivity evoked by TMS to brain consciousness.
Disorders of consciousness (DOC) are altered consciousness states caused by an injury or dysfunction of the neural systems regulating arousal and awareness [12,13].DOC comprise a vegetative state/unresponsive wakefulness syndrome (VS/UWS) [14] and a minimally conscious state (MCS) [15].DOC patients are an important but still underexplored entity in neurology.Researchers are increasingly turning to TMS-EEG measurements to gain a better understanding of DOC.Cortical responses to TMS in DOC are similar to those seen in unconsciousness states, such as sleep and anesthesia, suggesting a breakdown of effective connectivity [16].Furthermore, the complexity of the temporal-spatial distribution of the TMS-evoked potentials allows differentiation between MCS and VS/UWS [17].
Information theoretic analysis on TMS-EEG provides the opportunity to measure a cortical effective information flow induced by TMS.For instance, applying symbolic transfer entropy, vector autoregression and adapted directed transfer function could reveal effective connectivity between the TMS target and other brain regions, that construct the TMS induced information network [2,3,18].Identifying information flow patterns during different states of consciousness is greatly needed, which will contribute to disentangling the causal information flow that is critical for human consciousness.

Participants
A cohort of 35 DOC patients (15 females; mean age ± SD: 53.18 ± 17.88 years) were enrolled into the study based on the inclusion criteria: (i) persistence of a DOC state for at least 1 month after the acute brain insult; (ii) stable vital signs and free of acute medical complications (e. g., acute pneumonia).Exclusion criteria included any contraindications to MRI or TMS (e.g., pacemakers, intracranial stents, epilepsy, incomplete skull) and no detectable motor evoked potentials (MEP) in hand muscles with TMS of the primary motor cortex (M1) of either hemisphere, even at maximum stimulator output.The DOC patients included 21 patients with MCS (9 females, mean age ± 1SD: 47.31 ± 10.86 years) and 14 patients with VS/UWS (6 females, mean age ± 1SD: 55.02 ± 20.81 years), based on the Coma Recovery Scale-Revised (CRS-R) [19].Twenty-two age-matched healthy subjects (13 females) were included to form the healthy control group for this study.All participants were screened for possible contraindications to TMS and met the criteria specified in the TMS safety screening questionnaire [20].Written informed consent to participate in the study was obtained from the legal custodians of the DOC patients and healthy participants.This study was approved by the Ethics Committee of the First Affiliated Hospital of Nanchang University (No. IIT-2022-222).

TMS-EEG data recording
A Brainsight® neuronavigation system (Rogue Research Inc., Montreal, Canada) and individual anatomical MRI (three-dimensional gradient echo T1) was used to enable consistent positioning of TMS coil in the TMS-EEG recordings.The participants' head was co-registered to individual MRI data and was mapped to the Montreal Neurological Institute (MNI) coordinate system.TMS-EEG data was acquired by a TMS-compatible EEG system (BrainAmp 64 MR Plus, BrainProducts GmbH, Munich, Germany).The EEG cap was equipped with 64 TMS-compatible C-ring slit Ag/AgCl pin electrodes arranged in the International 10-20 montage.The EEG amplifier was set with a hardware filter at DC to 10 kHz and a sampling rate of 5 kHz.The skin/electrode impedances of all electrodes were maintained below 5 kΩ throughout the data recordings.
Prior to the TMS-EEG recordings, the individual resting motor threshold (RMT) was determined by motor evoked potential (MEP) recording from the abductor pollicis brevis muscle using surface EMG Ag-AgCl cup electrodes in a belly-tendon montage.The EMG raw signal was amplified, bandpass filtered (20 Hz-2 kHz) with a 2-channel EMG device built in the Brainsight® system.Individual RMT was defined as the minimum intensity that was sufficient to elicit an MEP of >50 μV peak-to-peak amplitude in at least five out of ten subsequent trials [21].
TMS-EEG data were recorded with the TMS coil placed at righthemispheric middle frontal gyrus (x = 43, y = 21, z = 38), motor cortex (x = 51, y = − 8, z = 44) or superior parietal lobule (x = 30, y = − 67, z = 60) on the individually MNI-fitted images.The corresponding coil positions were set on the surface, tangential to the scalp and by orienting the coil such that the induced electric field was perpendicular to the main axis of the target gyrus.TMS pulses were delivered through a 70 mm figure-of-eight coil connecting a Magstim R 2 stimulator (Magstim Company Limited, Whitland, UK) with a monophasic current waveform.200 single TMS pulses at an intensity of 90 % RMT were delivered with jittered inter-trial intervals of on average 3 s at each target.The participants were supplied with earphones with white noise during the TMS-EEG recordings to minimize EEG contamination by auditory evoked potentials caused by the TMS coil click [22].

Data processing
The MRI pre-processing and mesh extraction were performed based on the FieldTrip toolbox running in the MATLAB (Version 2017b, MathWorks Inc., Natick, USA) environment and FreeSurfer, following pipeline of our previous study [23].Preprocessing of TMS-EEG were performed using customized analysis scripts on MATLAB and EEGLAB 14.1.2b.The continuous TMS-EEG data were segmented into epochs with respect to the TMS trigger markers.The epochs were defined from − 1000 ms to 1000 ms and baseline corrected with − 500 ms to − 20 ms.Data from − 2 ms to 10 ms around the TMS pulses were excluded and cubic interpolated to eliminate the high-amplitude TMS artifact.EEG data were then visually inspected to identify and exclude the epochs containing major artifacts and the channels that showed prominent noise in most of the epochs.Afterwards, data was down-sampled from 5 kHz to 1 kHz and submitted to a two-step ICA procedure.In a first step, ICA components representing high-amplitude TMS-related artifacts were inspected and removed based on the topography, single-trial time-course and average time-course.Then, data were filtered with a 1-100 Hz zero-phase Butterworth band-pass filter and a 48-52 Hz notch filter.As a second step, ICA was used to remove artifacts containing eye blinks, eye movement and persistent scalp muscle activity.Finally, the channels that were discarded during the visual inspection were spline-interpolated using signal of the neighbor channels and data were then average-referenced.
The individual lead field matrix was calculated based on the aligned electrodes, individual mesh, and head model.The source reconstruction of TMS-evoked potentials (TEPs) was performed using a linearly constrained minimum variance beamforming method [24].We then reconstructed 12 regions of interest (ROIs), including inferior parietal lobule (IPL), superior parietal lobule (SPL), inferior frontal gyrus (IFG), middle frontal gyrus (MFG), precentral gyrus (PrG) and postcentral gyrus (PoG) of both hemispheres.The activity for each of these ROIs was estimated by taking the average from source TEPs of dipoles included in the ROIs [25].

Information flow
TMS-induced information flow was estimated on TEPs (− 500 ms before to 1000 ms post TMS pulse) by combining time-varying multivariate adaptive autoregressive (TV-MVAAR) model and adaptive directed transfer function.
Firstly, for each time-series TEP X(t), the temporal-spatial relationship was estimated by: P is the model order.It signifies the number of historical time points needed to estimate the present time point.A(i, t) are the matrices of time-varying model coefficients, defining relationship between the historical time points and the present time point.E(t) is the multivariate independent white noise.Kalman filter algorithm was used to describe the behavior of the multivariate signals by the observations equation, i. e., Eq. ( 1), and with the following state equation: We conducted the Akaike information criterion (AIC) to automatically determine the model order P in this study with the following equation: where M is the number of the time series, P is the optimal model order defined within the order range 1-30 samples, N is the time point and S is the covariance matrix.The observation and state equations were solved by the recursive least squares algorithm with forgetting factor.Secondly, the time-frequency information flow was derived from the frequency domain expression of A: where A k=0 = I, X(f, t) and E(f, t) are the frequency transformation of X(t) and E(t).In this way, the information flow H(f, t) could be obtained with elements H ij representing the information flow with a source at jth node and a sink at ith node.Finally, the TMS-induced information flow was normalized by zscore standardization relative to the baseline t b (− 500 ms to − 20 ms before TMS pulse):

Directed information flow
The directed information flow was calculated by the degree difference of information flow-out and flow-in across each sample time point in the kth ROI: Accordingly, the ROIs could only act as either an information sender, with positive values of directed information flow, or a receiver, with negative values of directed information flow, at any given time in the directed information flow network.To identify when each ROI acted as a sender or receiver, time windows with significant (values compared with 0) positive or negative values of directed information flow were determined, respectively.The directed information flow networks contain ROIs which can reverse their roles over time, from either a sender to a receiver or from a receiver to a sender.These ROIs are referred to as information interactive ROIs, and the ratio of information interactive ROIs to the total number of ROIs is defined as interactive rate for each subject.

Data availability
The data and all custom written MATLAB codes that support the findings of this study are available from the corresponding author upon reasonable request.

Statistical analysis
To avoid spurious information flow, we created surrogates to establish a null hypothesis (no information flow between source and sink nodes).We ran surrogates for source and sink nodes 1000 times to create a reference dataset and reference time-frequency information flow distribution.The TEPs time-frequency information flow was then adjusted by only keeping the values that were significantly higher (permutation test, p < 0.05) compared to the reference distribution.
TMS-induced total information flow in group contrasts (healthy vs. MCS, healthy vs. VS/UWS and MCS vs. VS/UWS) was calculated by summarizing 1-30 Hz from the average time-frequency information flow matrix of all ROIs.To assess the significance of the information flow, two-tailed independent t-tests with FDR correction (q < 0.05) were conducted across time (21-1000 ms).Subsequently, one-way ANOVA with group (3 levels: healthy, MCS and VS/UWS) as main effect was used to compare the average of induced total information flow across the time points.Bonferroni correction (3 comparisons, p < 0.05) was applied to adjust the p values after multiple comparisons, to identify significance in group contrasts.
TMS-induced average information flow-out and flow-in of ROIs, as well as the degree of information flow between ROIs (averaged across the time-frequency information flow matrix in 1-30 Hz and 21-1000 ms), were compared in group contrasts (healthy vs. MCS, healthy vs. VS/ UWS and MCS vs. VS/UWS) by two-tailed independent t-tests with FDR correction (12 comparisons, q < 0.05).The TMS-induced information flow between ROIs were compared in each group contrasts by two-tailed independent t-tests with FDR correction (total 65 comparisons, q < 0.05).
Global mean field amplitude (GMFA) of directed information flow was evaluated and compared between groups over time (21-1000 ms) using two-tailed independent t-tests with FDR correction (q < 0.05).The interactive rate of ROIs was compared between groups (healthy, MCS and VS/UWS) using one-way ANOVA p values after multiple comparisons were adjusted using Bonferroni correction (3 comparisons, p < 0.05) to identify significant group contrasts.Accuracy (ACC) and area under receiver operating characteristic curve (AUC) was used to measure the classification capacity of interactive rate in group contrasts.The relationships between the interactive rates of ROIs tested by TMS of frontal, motor and parietal cortex, and the CRS-R scores of the DOC patients were examined using Pearson correlation.

Results
The surrogate analysis validated that the method of combining timevarying multivariate adaptive autoregressive model and adaptive directed transfer function was capable of extracting information flow from TMS-EEG data (Figs.S1-S2 in Supplementary Materials).We Y. Bai et al. extracted TEPs from TMS-EEG recordings of healthy and DOC patients.TEPs of a representative subject, MCS and VS/UWS patient are shown in Fig. S3 (Supplementary Materials).The time-frequency information flow spectrum was obtained by applying time-varying multivariate adaptive autoregressive model and adaptive directed transfer function on TEPs.Healthy subjects displayed a distinct pattern of information flow (average of all ROIs) in the 21-800 ms post TMS pulse time window, while DOC patients (both MCS and VS/UWS) exhibited a weaker information flow (Fig. 1A).The degree of information flow (summary over the frequency band of 1-30 Hz) showed significant differences (twotailed independent t-tests with FDR correction, p < 0.05) between groups over time with TMS of frontal cortex (Fig. 1B).The total degree of information flow (summary of 1-30 Hz and average of 21-800 ms) was significantly different (F (2,73) = 13.99,p < 0.001) between groups.It was significantly higher in the healthy group compared to the MCS group (mean ± 1SD, healthy: 12.72 ± 10.80 vs. MCS: 3.74 ± 5.46, p < 0.001 after multi-comparison correction) and VS/UWS (VS/UWS: 3.29 ± 2.37, p < 0.001) (Fig. 1C).
TMS of motor and parietal cortex both induced distinct information flow in the healthy group, while DOC patients showed a weaker information flow (Fig. 1D-G).The degree of information flow showed significant differences over time (two-tailed independent t-tests with FDR correction, p < 0.05) between healthy vs. MCS and healthy vs. VS/UWS with TMS of parietal cortex (Fig. 1H), and a significant difference between MCS and VS/UWS with TMS of motor cortex (Fig. 1E).The total degree of information flow was significantly higher (F (2,73) = 6.78,p = 0.002) in the healthy group than VS/UWS (mean ± 1SD, healthy: 7.63 ± 4.95 vs. VS/UWS: 2.66 ± 2.81, p < 0.001 after multiple comparison correction) with TMS of parietal cortex (Fig. 1I), but without a significant difference in the group contrasts (p > 0.05) with TMS of motor cortex (Fig. 1F).
The average information flow in and out of the defined ROIs induced by TMS of frontal, motor and parietal cortex was compared between healthy subjects, MCS and VS/UWS patients.In comparison to healthy subjects, TMS-induced information flow in DOC patients was significantly lower in both directions (Fig. S4, Supplementary Materials).
With TMS of frontal cortex, healthy subjects had significantly higher (two-tailed independent t-tests with FDR correction) information outflow (summary of information-out from each ROI to all other ROIs) in 8/12 ROIs (except the right PrG, left IPL and bilateral SPL) than VS/ UWS, and in 8/12 ROIs (except the left PrG, left IPL and bilateral SPL) than MCS (Fig. 2A).There was no significant difference of outflow between MCS and VS/UWS.Information inflow was significantly higher in healthy subjects in 9/12 ROIs (except left PrG, SPL and right IFG) than VS/UWS and in 4/12 ROIs (left MFG, IFG, IPL and right PoG) than MCS.Information inflow of MCS patients was significantly higher compared to VS/UWS in 8/12 ROIs (except left PrG, PoG and bilateral SPL) (Fig. 2B).
With TMS of motor cortex, healthy subjects showed significantly higher information outflow and inflow than VS/UWS (outflow: right PrG, PoG and bilateral IPL; inflow: bilateral MFG, IFG, right PoG and left IPL) and MCS (outflow: left MFG and right IPL; inflow: bilateral MFG, IFG, right PoG and left IPL).There was no significant difference of information inflow or outflow between MCS and VS/UWS (Fig. 2C and D).
With TMS of parietal cortex, the ROIs of healthy subjects showed significantly higher information inflow/outflow compared to VS/UWS (Fig. 2E and F).
When the direction of information flow is taken into account, TMS induced significantly different (t-test, FDR correction) interhemispheric and intrahemispheric bidirectional information flow between healthy subjects and DOC patients (Fig. 3).This difference between healthy subjects and VS/UWS was most conspicuous with TMS of frontal cortex compared to TMS of motor and parietal cortex.A few differences of directional information flow were also present between MCS and VS/ UWS with TMS of frontal and parietal cortex but not with TMS of motor cortex.
The directed information flow was defined by the degree difference of information outflow and inflow over time (for details, see Methods).Flipping of the directed information flow from outflow to inflow, or from inflow to outflow was observed in most of the ROIs, lasting up to 600 ms in healthy subjects, with TMS of frontal (Fig. 4A),motor (Fig. 4B) and parietal cortex (Fig. 4C).Most of the ROIs of MCS patients showed lowamplitude, short-time directed information flow (lasting up to 400 ms) and reduced flipping, in comparison to healthy subjects (Fig. 4A-C).TMS of the frontal and parietal cortex did not induce marked directed information flow in VS/UWS patients, while TMS of motor cortex still induced distinct directed information flow in VS/UWS patients (Fig. 4B).The directed information flow showed significantly higher GMFA values (two-tailed independent t-tests with FDR correction, p < 0.05) in healthy subjects than patients with MCS mainly at 400 ms-700 ms, and patients with VS/UWS mainly at 40 ms-700 ms with TMS of frontal cortex (Fig. 4D).TMS of motor cortex induced significantly higher GMFA values of directed information flow in healthy subjects than in MCS mainly at 50-70 ms, and VS/UWS mainly at 30 ms-400 ms (Fig. 4E).TMS of parietal cortex induced significantly higher GMFA values of directed information flow in healthy subjects than in VS/UWS at 30 ms-600 ms and MCS at 50 ms-200 ms, and significantly higher GMFA values in MCS than in VS/UWS patients at 80 ms-700 ms (Fig. 4F).
The interactive rate of ROIs was influenced by the order of the autoregressive model in the TEP analysis.Specifically, the interactive rate decreased with increasing order with TMS of frontal, motor and parietal cortex (Fig. S6, Supplementary Materials).The interactive rate of ROIs did not show any significant relationship with the local cortical evoked power, which was calculated from TEPs with a time period from 25 to 80 ms post TMS pulse (Fig. S7, Supplementary Materials).

Discussion
Capturing the effective connectivity between regional neural activities induced by TMS would enable us to gain insight into which information is transmitted across brain regions.We explored time-frequency dynamic information flow induced by TMS of different consciousness states, by a method combining time-varying multivariate adaptive autoregressive model and adaptive directed transfer function.The results revealed that the information flow induced by TMS was significantly weaker in DOC patients, especially VS/UWS patients, when compared to healthy subjects.The directed information flow and interactive ROI rate revealed a breakdown of effective information interaction between brain regions in DOC patients.Furthermore, the features of TMS-induced information flow showed significant differences between different consciousness groups (healthy subjects, MCS, VS/UWS) and correlation with DOC clinical scales (CRS-R), indicating a physiological relevant correlation between the capacity of effective information flow and human consciousness.
The human brain is structured in a way that optimizes the exchange of information to enable more complex functions [26].Traditional approaches using undirected functional connectivity have advanced our understanding of the functional architecture of the brain and its relationship with human consciousness [27].The total information flow (Fig. 1), in line with previous studies, demonstrated a suppression of brain network function in DOC patients.The findings revealed a significant reduction in the information outflow and inflow of brain regions in DOC patients (Figs. 2 and 3), compared to healthy conscious subjects.These findings are consistent with the core claim of the Global Workspace Theory, i.e., the more the information of a state is globally available, the more that state can be consciously accessed.However, the Global Workspace Theory assigns consciousness to specific brain areas (in particular, anterior areas) with widespread broadcast neuronal workspace [2].Our findings revealed a widely distributed decreased information flow in DOC patients, with TMS of frontal, motor and parietal cortices.Considering the dependence of TMS-induced brain reactivity on structure and functional network, the loss of effective information propagation (decreased total information flow) directly reflects the breakdown of causal interactions between brain regions.The findings are in agreement with the central claim of the Integrated Information Theory, which links consciousness with the cause-effect structure of a system [3].
However, one important feature of cortical interregional connections is that they are frequently reciprocal in nature, which implies that information should be exchanged in a bidirectional way [28].This bidirectional flow of information has been shown as a crucial feature of the neurobiological systems that support functions relevant to consciousness, such as language [29], cognitive control [30] and attention [31].The present findings increase the evidence that human consciousness is not only driven by the total amount of information exchange, but also by the direction of the information flow.Specifically, DOC patients showed a significant lack of bidirectional information flow (Figs. 3 and 4).The cortical areas of DOC patients only maintained a single-directional function of information transmission, either as a sender or receiver, leading to a spatially restricted and temporally short-lived (Fig. 4) information propagation network triggered by TMS.This could likely be a shared common mechanism that explains the breakdown of TMS effective connectivity in DOC patients [16].Furthermore, the number of ROIs, playing roles of bidirectional information flow, was significantly decreased in DOC patients in comparison to healthy subjects and was positively correlated with the clinical scales of DOC patients (Fig. 5).These findings reveal that directed information flow takes precedence over information processing itself in the generation of consciousness, as pointed out by the Information Flow Theory [32].
The alterations in TMS-induced information flow are associated with abnormal oscillatory cortical responses to TMS in DOC patients, as information flow is essentially a measure of frequency-specific causal interactions.Oscillatory cortical reactivity to TMS is believed to mirror the intrinsic dynamics of the corresponding corticothalamic circuits [33].Thus, the decreased information flow with TMS of frontal, motor and parietal cortex in DOC patients compared to healthy subjects is likely due to changes in the thalamo-frontal, thalamo-motor and thalamo-parietal circuits caused by the brain injury.Among them, the thalamo-frontal and thalamo-parietal circuits were considered more relevant than thalamo-motor circuit in supporting human consciousness.The meso-circuit model posits that the thalamo-frontal and thalamo-parietal circuits are essential pathways for the integration of consciousness-related information processing [34,35].Consistently, our findings revealed that the MCS patients demonstrated higher interactive rate with TMS of frontal and parietal cortices than VS/UWS, but not with TMS of the motor cortex (Fig. 5), supporting a crucial role of frontal and parietal regions with regard to consciousness-related information flow.This is in agreement with the mainstream consciousness theory.For instance, the role of frontal cortex was emphasized in higher-order theories and global workspace theories, given its association in meta-representation of cognitive functions and widespread broadcast neuronal workspace [36].Integrated Information Theory links consciousness primarily with posterior cortical areas, in part on the grounds that these areas exhibit neuroanatomical properties that are supposedly well suited for generating high levels of integrated information [5].
TMS-EEG measurements are recommended in assessing consciousness levels of DOC patients [37].MCS and VS/UWS showed significant difference in the TEP patterns [16] as well as the corresponding quantity indices (e.g., Perturbational Complexity Index, PCI) [17].TEPs and PCI, unlike the information flow analysis, directly represent information interaction but not the causal patterns in the information flow.Time-varying multivariate adaptive autoregressive models measures the causal interaction between brain regions, enabling a better understanding of the pathways of TMS-induced information propagation and how information is integrated within a brain network.Furthermore, the information flow unravels the information transmission caused by TMS from a frequency-temporal-spatial perspective, and offers a potential avenue to explore the neuro-electrophysiological mechanisms related to consciousness from multiple aspects.It is especially vital in exploring DOC, since consciousness impairment could lead to abnormal frequency spectra, temporal dynamics and spatial networks [37].Another directed connectivity measure, i.e., symbolic transfer entropy, has been proposed to evaluate the TMS-induced interactions across brain regions [18].However, symbolic transfer entropy has the limitation that this bivariate measure could be potentially misleading when applied to a multivariate system [38], because it only investigates the statistical relationship of two signals but does not take into account the influence of other parts of the system.With TMS-EEG analysis, the complexity of the situation is immediately apparent, since the route of TMS-induced information transmission is intricate and unpredictable.Another drawback of the symbolic transfer entropy method is that it assumes that the connectivity pattern between signals is unchanged over the analyzed time period.Given the highly dynamic nature of neural activity, such an assumption might not always be valid, especially for events with short duration, as is the case with TMS-induced neural activity [39].Thus, the symbolic transfer entropy method may not catch time-variant connectivity, and violation of such an assumption may lead to a possible erroneous reconstruction of connectivity patterns and create a misinterpretation of the propagation of TMS-induced activity.In such a situation, utilizing the information flow method, which is derived from the multivariate autoregressive model fitting to the data, is an appropriate approach to measure TMS-induced neural responses.

Limitations
We applied time-varying multivariate adaptive autoregressive models to measure information flow with TMS of frontal, motor and parietal brain regions, and have provided the initial proof that breakdown of effective information flow is closely linked with impairment of human consciousness.However, limitations remain: Firstly, timevarying multivariate adaptive autoregressive models impose a linear relationship between variables that could hardly be justified from the standpoint of the nature of TEPs.Secondly, the computation cost of creating time-varying multivariate adaptive autoregressive models to fit the TMS-EEG data is expensive, thus restricting the number of investigated nodes in the TMS information flow network.As a result, the presented information flow patterns depend on the current nodes selection, and may not exactly detangle the real causal relationship between brain regions because of the potential influence from unselected areas.

Conclusions
Measurement of effective information flow induced by TMS offers unique insight into the brain reactivity to external stimuli, and into the pathophysiological mechanisms of brain functions.With the growing scientific and clinical interest into DOC, effective information flow provides a new perspective on the information interaction networks underlying human consciousness, and yields potentially meaningful biomarkers for assessing DOC.

Fig. 1 .
Fig. 1.TMS-induced information flow with TMS targeting frontal, motor and parietal cortex in healthy subjects, and patients with minimally conscious state (MCS) and vegetative state/unresponsive wakefulness syndrome (VS/UWS).(A) Average time-frequency information flow in healthy subjects, MCS and VS/UWS, tested with TMS of frontal cortex.Pink colors represent power values of the information flow.Vertical grey bars indicate 0-20 ms following the TMS pulse.(B) Summary (1-30 Hz) of information flow over time with frontal TMS in healthy subjects (red), MCS (grey) and VS/UWS (yellow).Curves and shades indicate mean ± 1SD of each group.Color lines above the curves represent significant (t-tests with FDR correction) differences in the group contrasts.(C) Bar plots (mean ± 1SD) of total information flow (summary of 1-30 Hz and average of 21-800 ms) tested with frontal TMS in healthy, MCS and VS/UWS group.Asterisks indicate significant differences (one-way ANOVA, post-hoc tests with Bonferroni correction) in the group contrasts.(D-F) Same arrangement and conventions as for (A-C), but with TMS of motor cortex.(G-I) Same arrangement and conventions as for (A-C), but with TMS of parietal cortex.

Fig. 2 .
Fig. 2. TMS-induced information flow of the regions of interests (ROIs) by TMS of frontal, motor and parietal cortex in healthy subjects, minimally conscious state (MCS) and vegetative state/unresponsive wakefulness syndrome (VS/UWS).(A) ROIs (red dots) with significantly different (t-test, FDR correction) information outflow (summary of information outflow from each ROI to all other ROIs) in the indicated group contrasts, with TMS of frontal cortex.(B) ROIs (red dots) with significantly different (t-test, FDR correction) information inflow (summary of information inflow of each ROI from all other ROIs) in the indicated group contrasts, with TMS of frontal cortex.(C-D) Same arrangement and conventions as in (A-B), but with TMS of motor cortex.(D-E) Same arrangement and conventions as in (A-B), but with TMS of parietal cortex.Abbreviations: IPL -inferior parietal lobule; SPL -superior parietal lobule; IFG -inferior frontal gyrus; MFG -middle frontal gyrus; PrG -precentral gyrus; PoG -postcentral gyrus; L -left hemisphere; R -right hemisphere.

Fig. 3 .
Fig. 3. Significant differences of information flow (with direction) between regions of interest in the indicated group contrasts.(A) Significantly higher (t-test, FDR correction) information flow in healthy subjects than patients in vegetative state/unresponsive wakefulness syndrome (VS/UWS) (left) or in minimally conscious state (MCS) (middle), and in MCS compared to VS/UWS (right), with TMS of frontal cortex, as indicated by the red lines.(B-C) Same arrangement and conventions as in (A), but with TMS of motor cortex (B) and parietal cortex (C).Abbreviations: IPL -inferior parietal lobule; SPL -superior parietal lobule; IFG -inferior frontal gyrus; MFG -middle frontal gyrus; PrG -precentral gyrus; PoG -postcentral gyrus; L -left hemisphere; R -right hemisphere.

Fig. 4 .
Fig. 4. Directed information flow of regions of interests (ROIs).Average directed information flow of each ROI across the frequency band of 1-30 Hz in healthy (left), minimally conscious state (MCS) (middle) and vegetative state/unresponsive wakefulness syndrome (VS/UWS) (right) with TMS of frontal (A), motor (B) and parietal cortex (C).Each curve represents one ROI with positive values indicating outflow, and negative values indicating inflow.Global mean field amplitude (GMFA) (mean ± 1SD) of directed information flow in healthy subjects (red), MCS (grey) and VS/UWS (yellow) with TMS of frontal (D), motor (E) and parietal cortex (F).Color lines above the curves represent significant differences (t-tests with FDR correction) in the indicated group contrasts.Abbreviations: IPL -inferior parietal lobule; SPLsuperior parietal lobule; IFG -inferior frontal gyrus; MFG -middle frontal gyrus; PrG -precentral gyrus; PoG -postcentral gyrus; L -left hemisphere; Rright hemisphere.

Fig. 5 .
Fig. 5. Information interactive rate of regions of interests (ROIs).(A) Interactive rate of ROIs (mean ± 1SD) in healthy, minimally conscious state (MCS) and vegetative state/unresponsive wakefulness syndrome (VS/UWS) groups tested with TMS of frontal cortex.(B) The scatter plot shows the regression line between interactive rate tested with TMS of frontal cortex and the Coma Recovery Scale-Revised scores (CRS-R).(C-D) Same arrangement and conventions as for (A-B), but with TMS of motor cortex.(E-F) Same arrangement and conventions as for (A-B), but with TMS of parietal cortex.In (A, C, E) Asterisks indicate significant differences (one-way ANOVA, post-hoc tests with Bonferroni correction) in group contrasts.In (B, D, F), each circle represents data of one DOC patient.Abbreviations: ACC = accuracy; AUC = area under the curve; RMSE = root mean square error.