Uncovering Brain Network Insights for Prognosis in Disorders of Consciousness: EEG Source Space Analysis and Brain Dynamics

Accurate prognostic prediction in patients with disorders of consciousness (DOC) is a core clinical concern and a formidable challenge in neuroscience. Resting-state EEG has shown promise in identifying electrophysiological prognostic markers and may be easily deployed at the bedside. However, the lack of brain dynamic modeling and the spatial mixture of signals in scalp EEG have constrained our exploration of biomarkers and comprehension of the mechanisms underlying consciousness recovery. Here, we introduce EEG source space analysis and brain dynamics to investigate the brain networks of patients with DOC (n = 178) with different outcomes (six-month follow-up), followed by graph theory and high-order topological analysis to explore the relationship between network structure and prognosis, and finally assess the importance of features. We show that a positive prognosis is associated with large-scale lower levels of low-frequency hypersynchrony. Moreover, we provide evidence that this pattern is driven not by all brain states but only by specific states. Analyses reveal that the positive prognosis is attributed to the network retaining lower segregation, higher integration, and stronger stability compared to the negative prognosis. Furthermore, our results highlight the importance of brain networks derived from brain dynamics in prognosis. The prognosis models based on clinical and neural features can achieve acceptable and even excellent performance under different outcome definitions (AUC = 0.714–0.893). Overall, our study offers new perspectives for the identification of prognostic biomarkers and provides avenues for profound insights into the mechanisms underlying consciousness improvement or recovery.

Resting-state EEG has shown promise in identifying electrophysiological prognostic markers and may be easily deployed at the bedside.However, the lack of brain dynamic modeling and the spatial mixture of signals in scalp EEG have constrained our exploration of biomarkers and comprehension of the mechanisms underlying consciousness recovery.Here, we introduce EEG source space analysis and brain dynamics to investigate the brain networks of patients with DOC (n = 178) with different outcomes (sixmonth follow-up), followed by graph theory and high-order topological analysis to explore the relationship between network structure and prognosis, and finally assess the importance of features.We show that a positive prognosis is associated with large-scale lower levels of low-frequency hypersynchrony.Moreover, we provide evidence that this pattern is driven not by all brain states but only by specific states.Analyses reveal that the positive prognosis is attributed to the network retaining lower segregation, higher integration, and stronger stability compared to the negative prognosis.Furthermore, our results highlight the importance of brain networks derived from brain dynamics in prognosis.The prognosis models based on clinical and neural features can achieve acceptable and even excellent performance under different outcome definitions (AUC = 0.714-0.893).Overall, our study offers new perspectives for the identification of prognostic biomarkers and provides avenues for profound insights into the mechanisms underlying consciousness improvement or recovery.

I. INTRODUCTION
S EVERE brain injuries, often resulting from traumatic brain injury, cardiac arrest, and stroke, frequently lead to prolonged disorders of consciousness (DOC), characterized by unconsciousness lasting longer than 28 days [1], [2].Patients with DOC are diagnosed with either a vegetative state/unresponsive wakefulness syndrome (VS/UWS) or a minimally conscious state (MCS; subclassification: MCS− and MCS+), based on repeated assessments of the level of consciousness [3].Accurate and individualized prognostication of outcomes for patients with DOC is a critical clinical priority.It has direct implications for treatment decisions, care strategies, and family education [4], [5].However, the neurobiological heterogeneity in DOC and the limited understanding of the mechanisms of consciousness recovery hinder the identification of reliable prognostic markers [4].Various outcome prediction tools still face substantial challenges in their translation into clinical practice [6], [7].
Demographic and clinical characteristics have been extensively investigated, and certain factors were considered to impact the prognosis of DOC: Younger age [5], [8], [9], traumatic (vs.nontraumatic) etiology [2], [8], [9], shorter time postinjury [5], [8], [9], a MCS (vs.VS/UWS) diagnosis [2], [8] and a higher Coma Recovery Scale-Revised (CRS-R) total score [5], [8] were associated with a better prognosis/outcome.Recently, a variable named the consciousness domain index, defined by cluster labels obtained through k-means clustering (two clusters) of the six CRS-R subscores, may improve the prediction of recovery of consciousness compared to clinical diagnosis and the CRS-R total score [10].However, demographic and clinical candidate prognostic markers have not been consistently identified across studies nor have they provided insight into brain function and mechanisms of recovery in consciousness.
Thanks to the applications and advances in neuroimaging, there is a growing consensus that brain functions emerge from complex interactions between brain regions [11], [12].Previous research has suggested that consciousness rests on the brain's ability to sustain dynamic signal coordination and a balance between segregation and integration [11], [13], [14].Severe brain injury often leads to large-scale structural disconnections and abnormal functional connectivity (FC).This impairs the neural infrastructures that support consciousness and reduces the capacity of brain-wide information spreading and dynamic emergence [4], [8], [13], [14].Therefore, low consciousness is thought to be associated with increased network segregation [13], decreased network integration [13], [14], and reduced dynamic richness [11] and functional diversity [14].Moreover, the spatiotemporal evolution of brain signals (referred to as brain dynamics) can be characterized by several states (e.g., identified by clustering analysis [11], [15] or hidden Markov model [16]; referred to as brain states) that transition and reoccur in time.There is a common finding [11], [17], [18] that different brain states correspond to distinct FC patterns (i.e., state-specific FC [18]).However, brain injuries cause damage/alterations in brain dynamics that are not uniform across different states and can even be state-specific [11], [16], [17].In the context of prognosis, numerous neural indicators derived from neuroimaging data, such as EEG and functional magnetic resonance imaging (fMRI), have demonstrated prognostic predictive potential for patients with DOC, including EEG qualitative [5] and quantitative features [8], fMRI-based FC [9], among others [6], [7].Several prognostic prediction models for patients with DOC using clinical and/or neural features have shown promising performance, such as the use of fMRI FC [19], clinical features + fMRI FC [9], fMRI FC + EEG features [20], clinical + EEG qualitative features [21], the TMS-EEG perturbational complexity index [22].However, the relatively small sample sizes and large differences among patients in studies (e.g., baseline diagnoses, time postinjury, and definition of the outcomes) limit the objective evaluation of the performance of different models.
Overall, noninvasive multimodal neural features are of particular interest for developing personalized prognostic prediction models for patients with DOC.Compared to other modalities or research paradigms, resting-state EEG represents an attractive choice owing to its safety, cost-effectiveness, direct detection of high temporal-resolution neural activity, and ease of deployment at the bedside.Previous studies have demonstrated the fundamental predictive power of resting-state EEG features [6], [7].However, there are two critical issues that are largely unexplored, and they are important for insights into the discovery of new prognostic markers and mechanisms underlying the recovery of consciousness.First, the relationship between the EEG source space brain network and prognosis remains unclear.Numerous efforts have been made to study scalp EEG brain networks in relation to the prognosis of patients with DOC [6], [7], [8].However, the spatial mixture of signals caused by volume conduction [23] and the limited spatial resolution substantially constrain the biological interpretation of the findings and the integration of information with other modalities.Second, as a dynamic system, the brain presents a research gap regarding whether brain networks underlying different brain states have the same or similar prognostic utility.Previous studies have shown that brain injuries can lead to significant alterations in brain dynamic parameters (e.g., mean state duration and mean state interval [18], [24]), but they do not necessarily result in marked alterations to the EEG brain state templates (e.g., EEG topographic maps corresponding to cluster centroids [18], [24]).Research on the prognostic abilities of brain networks underlying different brain states will provide opportunities for the identification of novel biomarkers and a deeper understanding of brain dynamics.
In this study, we retrospectively recruited 254 patients with DOC to investigate the key factors predicting prognosis and explore the neural underpinnings of different prognoses.In addition, the analysis in this study focused on the delta band (1-4 Hz) for the following three main reasons: (1) A firm link between delta cortex activity and consciousness [25], [26].(2) Reducing the number of statistical tests to improve the statistical power.(3) Strong independence of temporal sequences of brain states between the spectral bands [27].In general, the novel contributions of our study are as follows: (1) We show that source-space FC during only specific states in brain dynamics can predict outcomes, and large-scale hypersynchrony in the delta band has a negative effect on the improvement of consciousness.(2) We employ graph theory and higher-order topological analysis of brain networks to offer insights into the mechanism of consciousness improvement.(3) We present the key clinical and neural features for predicting prognosis based on a large cohort of patients.our sample size was larger than those reported in previous publications [8], [13], [20].

II. MATERIALS AND METHODS
2) Clinical Assessment and Outcome Definition: The level of consciousness in each patient with DOC was diagnosed by specialized physicians based on the highest score obtained from three CRS-R assessments [3] conducted at both admission and the six-month follow-up.Demographic and clinical characteristics for each patient are available at https://osf.io/fjwst/.In this study, the term "prognosis" denotes a prognosis for improvement in awareness.A positive outcome (prognosis) was defined as any improvement in the level of consciousness at the six-month follow-up compared to admission, whereas a negative outcome was defined as a lack of improvement or retrogress in consciousness.

B. EEG Aquasition
Resting-state EEG data were recorded using 62-channel acquisition equipment (BrainAmp 64 MR plus, Brain Products, Germany) with electrodes (Ag/AgCl) placed according to the standard 10-10 system at a sampling rate of 2500 Hz.We used data from 59 EEG channels for subsequent analyses.The reference channel was placed at FCz.The electrode impedance was maintained below 5 k .The arousal procedures were applied (with recording suspended) if needed (e.g., in the presence of signs of drowsiness and sleep onset) during data collection.There is one EEG recording per patient.For most patients with DOC, EEG acquisition occurred between 8-10 am (29.2%) or 2-4 pm (57.9%).During the data recording, we ensured that patients were awake and had their eyes open.The resting-state EEG (59 EEG channels) acquisition protocol for the HC dataset has been described in detail elsewhere [18].The duration of the recordings was over 5 minutes (760 ± 211 s) for the patients and approximately 4 minutes (243 ± 62 s) for the HC group.

C. EEG Preprocessing
All EEG recordings in this study were preprocessed using the same semi-automated procedures based on the EEGLAB toolbox (v2019.0)[18].First, the EEG data were lowpass filtered (cutoff frequency: 45 Hz; passband ending at 42.5 Hz) using a finite impulse response (FIR) filter, resampled to 250 Hz, and highpass filtered (FIR; cutoff frequency: 0.5 Hz; passband starting at 1 Hz).The bad channels in each recording were then identified and removed semiautomatically.Bad-channel signals were interpolated using spherical interpolation but marked as "bad".The EEG signals were then rereferenced to the common average across all channels.Finally, independent component analysis was performed for each recording, and bad components (e.g., eye blinks, movements, muscle activities, channel noise, and heart artifacts) were removed with the assistance of EEGLAB plugin extensions (ICLabel v1.3 and DIPFIT v3.3).We selected the middle 5 minutes of the preprocessed EEG data for subsequent analyses to maintain consistent data length across all patients.

D. EEG Source Reconstruction and Region-Level Time Series
Different source localization algorithms and FC measures may induce variability in the results.In our study, we chose to use the weighted minimum norm estimate (wMNE) [28] for source localization and the weighted phase-lag index (wPLI) [29] for FC, based on three primary considerations.(1) The wMNE is less sensitive to inaccuracies in the forward model, when using template anatomy and digitized electrode positions are not available [30], [31]; (2) The wPLI is insensitive to zero-lag spurious correlations induced by field spread or volume conduction.(3) Recent research suggests that the combination of wMNE with wPLI yields an optimal performance in some contexts [32].
For each EEG recording, source reconstruction was carried out using the wMNE method [28] in Brainstorm [33].This process converted the EEG signals from sensor space to source space, encompassing 15 002 dipoles distributed across the cerebral cortex.We established a realistic three-layer head model (scalp, skull, and brain) based on the FreeSurfer average anatomy [34], employing the boundary element method with the OpenMEEG plugin (v2.4) [35].The conductivity values for the scalp, skull, and brain were set at 0.33, 0.0066, and 0.33 S/m, respectively [36].We constructed the noise covariance matrix as an identity matrix and constrained dipole orientations to be perpendicular to the cortical surface.Subsequently, the lead field matrix (i.e., the transfer gain from each source to each channel) was computed by solving the forward problem.Finally, the wMNE method (using default parameters) was applied to estimate the inverse solution (i.e., the imaging kernel).The current source densities were calculated by converting the sensor-space EEG signals using the obtained imaging kernel.
We next utilized the "Yan2023" 100-area cortex atlas [37] to extract the region-level time series.Specifically, we averaged source-space signals within each region after flipping the signs of the sources oriented in opposition to the main direction [38].Notably, each brain region can be assigned to one of seven well-characterized resting state networks (RSNs) [39]: visual network (VSN), sensorimotor network (SMN), dorsal attention network (DAN), salience attention network (SAN), limbic network (LIM), frontoparietal network (FPN), and default mode network (DMN).In other words, each RSN consists of several brain regions assigned based on specific algorithms [37], [39].Additionally, we leveraged FC to characterize brain networks, where nodes represent brain regions and edge values denote connectivity strength.

E. Static Functional Connectivity
First, we investigated brain networks from a static perspective.To this end, region-to-region FC was constructed by calculating the wPLI between signals of each region pair [29].The analytic signal for each brain region was obtained by the convolution of the complex Morlet wavelets and regional data, from which the phase information was extracted.Given our specific interest in the delta band (i.e., 1-4 Hz), we employed three wavelets with peak frequencies at 2.0, 2.6, and 3.5 Hz.These corresponded to 3.0, 3.8, and 4.6 cycles, respectively.These wavelets exhibited full-width half maximum values in the time of 0.56, 0.47, and 0.40 s and in the frequency of 1.6, 1.9, and 2.2 Hz, respectively [40].The analytic signals were segmented into non-overlapping epochs of 500 samples (i.e., 2 s), and connectivity was computed for each frequency, epoch, and region pair.Finally, the FC was averaged across frequencies and epochs.The connectivity within each time window between two regional signalsis computed as in (1).
where N is the window size (i.e., the number of time points in the window) and S x y denotes the cross-spectral density derived from the analytic signals of regions x and y.
To study FC at the network leveland improve the statistical power, we calculated the average intra-network connectivity (n = 7) and average inter-network connectivity (n = 21) from each region-to-region FC matrix (100 × 100).

F. State-Specific Functional Connectivity Derived From Brain Dynamics
Our investigation of brain networks expanded to a dynamic perspective by applying microstate analysis [15] to derive the temporal sequence of brain states and computing the statespecific FC [18].
First, each sensor-space recording was filtered into the traditional delta band (1-4 Hz) using an FIR filter.Subsequently, 1000 EEG maps, each corresponding to the a peak (local maxima) of the global field powers, were randomly sampled from each recording.The selected maps from all patientswere merged, and a modified k-means clustering method (ignoring polarity) [41] was employed to determine the centroids of k (ranging from 2 to 10) clusters.The optimal number of clusters k * was determined using the Krzanowski-Lai criterion [41], [42].The centroids corresponding to the k * cluster were regarded as group-level microstate maps [41].Next, each time point in a recording was assigned as one of the k * microstate classes, based on the absolute spatial correlation between the EEG map at that specific time point and the microstate maps.Finally, state-specific FC was computed using a procedure similar to that described in Section II-E.However, here, FC was calculated separately for each microstate class (concatenating all samples within a given microstate class).

G. Graph Theory and Higher-Order Topological Analysis
Furthermore, to explore differences in brain networks between patients with different prognoses of DOC, we employed two graph-theoretic properties, namely, the clustering coefficient (CC) and participation coefficient (PC) [43], to investigate between-group differences in terms of functional segregation and integration.Additionally, we utilized a higher-order topological feature, namely, the 1-dimensional Betti number (β 1 ) [44], [45], to assess differences in network stability between groups.The nodelevel CC and PC were calculated using the Brain Connectivity Toolbox [43], followed by the calculation of average CC and PCacross 100 brain regions.The β 1 captures the number of cycles/loops within the brain network.For further explanation of these measures, please see Section S1-B.Before computing the CC, PC, and β 1 metrics, each FC matrix was z-scored and then binarized with a threshold value of one (similar to proportional thresholds).For comparison, the pooled FC of all the subjects was z-scored and then binarized with a threshold value of one (similar to the absolute threshold).For the PC, we employed the Newman [46] and Louvain [47] community detection algorithms to verify the consistency of our findings.

H. Statistical Analysis
All statistical analyses were conducted using custom MAT-LAB (R2022b) codes.The chi-square test was employed for goodness of fit (one-way) and to assess the independence between two dichotomous variables (two-way).The twosample t-test, Mann-Whitney U test, one-way analysis of variance (ANOVA), and Kruskal-Wallis test were selected based on their applicability to examine the group differences in clinical and network-level measures.The z-statistics corresponding to the U statistics were computed to indicate the direction of the effect.False discovery rate (FDR) correction [48] was applied to reduce the type I error rate by controlling the FDR in multiple tests in clinical and network-level measures.In addition to statistical inference at the network level, a threshold-free cluster-based statistical inference procedure [49] was used to obtain edge-level corrected p-value for difference testing in FC between groups.See Section S1-C for the parameter configurations.All aforementioned tests were two-sided, where applicable, with a significance level of 0.05.We also reported the effect size measures corresponding to each test type (see Section S1-C).Bootstrap confidence intervals [95% CI] were computed using a bias-corrected and accelerated percentile method with 5000 resamples.

I. Feature Importance Analysis
In addition to univariate analyses, we evaluated the importance of a feature when considered in combination with other features using two approaches.First, we employed multivariate logistic regression to examine whether a feature remained significantly associated with the outcome after accounting for the linear effects of other features.The effect size for each feature was measured using the odds ratio (OR; small: 1.68/0.60;medium: 3.47/0.29;large: 6.71/0.15[50]).See Section II-H for the computation of the 95% CI.
Second, we employed a model-agnostic approach, the Shapley additive explanations (SHAP), to estimate the feature importance using the Python package shap (v0.42.1).This game-theoretic approach quantifies the marginal contribution of each feature to the individual prediction using Shapley values (also referred to as SHAP values [51]).The average absolute SHAP value for each feature across subjects illustrates the magnitude of the contribution of each feature to the model output [52].To this end, we developed deep neural network (DNN) models to appraise the predictive power of baseline clinical and neural features concerning patient outcomes.These models incorporated SHAP values to quantify the feature contributions.Model performance somewhat determines the clinical value of feature importance analysis.We used DNN models because they excel at identifying complex nonlinear relationships.In this study, DNN models (feedforward neural networks) were equipped with two hidden layers containing 64 and 16 nodes, respectively.Both hidden layers incorporated Batch Normalization, the LeakyReLU activation function, and a dropout rate of 20%.Given the potential class imbalances, we utilized the class-weighted cross-entropy loss function.To maximize data utilization and estimate the model performance with low bias, the leave-oneout cross-validation (LOOCV) procedure was employed.The DNN model was trained with the Adam optimizer (AMSGrad variant) for 100 epochs but with an early stopping mechanism.The batch size was set to 32.The area under the receiver operating characteristic curve (AUC) was used to assess model performance.Python libraries PyTorch (v1.13.1) and Scikitlearn (v0.24.1) were used for the above procedures.

A. Patients at the Admission and Clinical Evolution Trajectory
Table I shows the demographic and clinical characteristics of the patients with DOC at the admission.Section S2-A provides a detailed description of the results.Fig. 1 illustrates the clinical evolution and prognosis of patients with DOC in different etiology and baseline diagnoses.Following the sixmonth follow-up, 64 patients showed improved consciousness, transitioning to higher levels.Among them, 32 in MCS (48.5%) and 32 in VS/UWS (28.6%) at baseline progressed to a higher consciousness level: three (25.0%) in MCS+, 12 (22%) in MCS−, and five in VS/UWS (4.5%) emerged from MCS (eMCS); 17 moved from MCS− (31.5%) to MCS+; 14 from VS/UWS (12.5%) to MCS+; and 13 from VS/UWS (11.6%) to MCS−.Among the 114 not improved patients, 111 (96.5%) did not change their consciousness level (eight in MCS+, 66.7%; 23 in MCS−, 42.6%; 80 in VS/UWS, 71.4%), whereas three patients exhibited retrogression (one from MCS+ to MCS−, two from MCS− to VS/UWS).See Section S2-B for statistical results on comparisons between improved and not improved patients in terms of baseline demographic and clinical characteristics.

B. Large-Scale Low-Frequency Hypersynchrony Indicates Negative Outcome
We first examined whether significant differences existed in resting-state static intra-and inter-network connectivity between HC and patients with DOC.The results demonstrate that low-frequency neural synchronization was significantly lower in HC than in patients with DOC across all network-level connectivity (Fig. S1; all corrected p < 0.001; r 2 : mean, 0.167, range, [0.068, 0.219]).Remarkably, as depicted in Fig. 2, the improved group exhibits significantly lower low-frequency synchrony in all intra-and inter-network connectivity compared to the not improved group (all corrected p < 0.05; r 2 : mean, 0.044, range, [0.027, 0.071]).Low-frequency hypersynchrony is found to be associated with low levels of consciousness, and low-frequency phase synchronization-based FC has been suggested as a potential neural marker for differentiating or diagnosing levels of consciousness [25], [53].However, in our study, there were no significant differences in network-level connectivity after FDR correction between patients in MCS and VS/UWS, albeit patients in MCS showed lower delta synchrony (Fig. S1; all corrected p > 0.1; r 2 : mean, 0.017, range, [0.001, 0.043]).These results suggest that a better prognosis is associated with lower low-frequency neural synchronization, which cannot be simply determined by the baseline level of consciousness.This finding was further confirmed by subgroup analyses.For example, in both baseline MCS and VS/UWS subgroups, the improved group exhibited significantly lower mean VSN-SMN connectivity values than the not improved group (see Fig. 2; p < 0.05).

C. Not All Brain States but Only Functional Connectivity of Microstate B Predicts Prognosis
We then employed analyses of state-specific FC to explore whether the improved group exhibited significantly lower delta synchrony across all brain states compared to the not improved group.In our study, the optimal number of microstate classes was determined as four (i.e., k * = 4).The four microstate maps (Fig. 3) were labeled as microstate A, B, C, and D, according to the previous studies [15].Intriguingly, only the FC of microstate B (FC B ) showed significant differences between the improved and not improved groups (83 edges with corrected p < 0.05; r 2 for the 83 edges: mean, 0.063, range, [0.054, 0.110]; see Fig. 3).Significant differences were observed in all intra-and inter-network connectivity (r 2 : mean, 0.029, range, [0.017, 0.038]), except for intra-network connectivity within the LIM (corrected p = 0.080) and intra-network connectivity between the LIM and VSN (corrected p = 0.057).Connectivity related to the DMN, SAN, and VSN exhibited a larger average effect size.Again, the differences between patients in MCS and VS/UWS were not significant in any state-specific FC (all corrected p > 0.05).Our findings highlight that the distribution of effects for differences between the improved and not improved groups does not remain at the same level at all moments, instead is more pronounced in specific brain states.
In addition, our study revealed a remarkable spatial similarity between the delta band and broadband (1-45 Hz) microstate maps (Pearson correlation coefficients > 0.98).However, when these delta-band microstate maps were applied to the delta-band and broadband signals, the average overlap of the microstate temporal sequences was only 58.4% (SD = 13.1%).These results indicate that microstate temporal sequences within distinct frequency bands exhibit high variability, underscoring the limited suitability of microstate temporal sequences obtained from one band for application in other bands.

D. Stronger Network Integration and Stability in Improved Patients
We further elucidated the neural underpinnings under the brain network associated with the prognosis of patients with DOC.Compared to the not improved patients, the improved patients showed a significantly lower CC (U = 2850; corrected p = 0.023; r 2 = 0.033 [0.002, 0.104]), higher PC (Newman [46]; U = 2937; corrected p = 0.031; r 2 = 0.026 [0.001, 0.095]), and larger β 1 (U = 2765; corrected p = 0.022; r 2 = 0.040 [0.003, 0.111]) in terms of FC B (Fig. 4).No significant differences were observed in the other three brain states.For static brain networks, we found similar patterns for CC, PC, and β 1 as in microstate B, but only PC (Newman; corrected p = 0.024) survived the FDR correction.Moreover, we obtained similar significant results in CC, PC, β 1 of both static FC and FC B when using absolute thresholds (Section II-G), except that β 1 of microstate B is not significant (corrected p = 0.512).Furthermore, for the PC, the findings using the Louvain algorithm [47] are consistent with those obtained using the Newman algorithm in terms of significance.

E. Key Clinical and EEG Prognostic Factors
To assess the importance of a feature when considering other features, we used a model-dependent approach (i.e., multivariable logistic regression) and a model-agnostic approach (i.e., SHAP values).
Moreover, we built DNN models to predict outcomes using baseline clinical and EEG network-level connectivity and used the SHAP approach to assess the importance of each feature (Section II-I).See Section S1-D for the features used in the DNN models.As depicted in Fig. 5, the top five most important features for prognostic prediction, as indicated by SHAP values, includes two clinical features (etiology and baseline diagnosis) and three average connectivity of FC B (DMN-SMN, VSN-DAN, and DMN-VSN).Among the top ten features, FC features accounted for eight, with only two being static FC features, ranking sixth (VSN-FPN) and eighth (SMN-VAN), respectively.Notably, our results demonstrated that the DNN model achieved an acceptable discrimination capability [55] (AUC = 0.714) for improved vs. not improved.Our findings underscore the importance of etiology and state-specific FC (FC B ), particularly the connectivity related to the DMN and VSN, in prognostic predictions.
Finally, various definitions exist regarding the outcomes of patients with DOC [7].In this study, we also used the LOOCV procedure to evaluate the performance of the models under three alternative outcome definitions: eMCS vs. not eMCS, ≥ MCS+ vs. < MCS+, and ≥ MCS vs. VS/UWS.The results demonstrated their respective abilities to achieve acceptable (AUC = 0.728), acceptable (AUC = 0.786), and excellent (AUC = 0.893) discrimination of prediction outcomes [55].Thus, we show the predictive capacity of baseline clinical and brain networks for outcomes under different definitions.We repeated the analyses of this study using another source localization approach (sLORETA [56]), which yielded similar results, with no change in the main findings.

IV. DISSCUSION
In this study, we investigated static and dynamic EEG source-space brain networks using resting-state EEG in a large cohort of patients with DOC.Our aim was to identify biomarkers associated with prognosis and gain mechanistic insights.We found large-scale differences in low-frequency brain networks between patients with positive (improved) and negative (not improved) outcomes.Improved patients displayed significantly lower levels of hypersynchrony in brain networks compared to not improved patients.Notably, among the brain states, only the brain network of microstate B was significantly associated with DOC prognosis.Moreover, we expounded on the impact of hypersynchrony on brain network segregation, integration, and stability in terms of network organization and topology.Furthermore, our findings demonstrate the importance of specific clinical indicators and brain states for individualized prognosis prediction.Our study provides a promising new perspective for developing prognosis prediction tools using EEG source-space brain networks and for gaining insights into mechanisms of consciousness improvement or recovery.
Our findings highlight that weaker hypersynchrony in the delta band indicates a better prognosis.From the perspective of brain dynamics, prognosis is primarily driven by the brain networks of microstate B. Hypersynchrony manifests as a significant enhancement in phase synchronization-based FC in patients with DOC compared to healthy individuals (Fig. S1) [8], [25].In our study, evidence from both static FC and state-specific FC converges on the findings that improved patients exhibit widespread, significantly lower hypersynchrony than not-improved patients, involving nearly all RSN-based intra-and inter-network connectivity.Our findings are in line with those of the previous study using scalp EEG [8].Moreover, we show that patients in MCS exhibit lower synchrony relative to patients in VS/UWS [25], but to a lesser degree (not significant) relative to improved vs. not improved patients.This suggests that delta FC may have greater prognostic potential relative to diagnosis.Consciousness and higher cognitive functions are dependent on interactions across distributed brain networks [12].The extent to which severe brain injury impairs between-and within-network interactions directly affects the patterns of brain segregation and integration and the potential for consciousness improvement.We attribute the explanation of the impact of low-frequency hypersynchrony on prognosis to alterations in network topology, subsequently affecting the neural substrates of functional emergence and consciousness recovery.
We substantially advanced our understanding of the impact of hypersynchrony on networks through graph theory and higher-order topological analysis of EEG source space brain networks.Brain-wide excessive synchrony may indicate a reduction in the information content of the thalamocortical system [26], leading to a phenomenon in which "everybody behaves like everybody else" [57].Our findings indicate that excessive hypersynchrony leads to higher segregation (CC), lower integration (PC), and weaker stability (β 1 ) within the brain network (Fig. 4), resulting in a poorer prognosis in patients with DOC.In addition, a larger CC is related to worse network synchronizability (a factor affecting collective behaviors of a dynamical network) as suggested in [58].Therefore, patients with a worse prognosis exhibit more localized information processing, hindered integration and coordination between brain regions, reduced "redundant" pathways within brain networks, decreased resilience to attacks, and impaired information propagation.These results intuitively illustrate the fundamental link between prognosis and neurobiological underpinnings following severe brain injuries.Presumably, our findings are logically consistent with those of active brain-computer interface paradigms [59].Most (87.5%) patients with covert consciousness (brain damage in them is likely less severe [60]) showed improvements in CRS-R scores at three months of follow-up [59].To some extent, this alleviates our concerns about the problem of covert consciousness in investigating relatively long-term (e.g., ≥ 6 months) prognoses.
Through feature importance analysis, we demonstrate that state-specific FC derived from brain dynamics (i.e., FC B ) has greater prognostic utility than static FC.Moreover, our results underscore the contributions of connectivity related to the DMN and VSN.Numerous fMRI studies have indicated that connectivity within the DMN and anticorrelations between the DMN and other brain regions are associated with the recovery of consciousness [4].Our research further provides electrophysiological evidence in this regard.Furthermore, it is intriguing to note that current evidence suggests an association between microstate B and many aspects related to the visual system and cognitive abilities, such as self, scene visualization, visual conscious experiences, and autobiographical memory [15], [61].We speculate that microstate B is a brain state crucial for both intrinsic awareness and responses to external stimuli.Our findings reveal that prognostic effects are driven by the specific brain state, akin to investigations of consciousness levels [11].State-specific brain networks in the EEG source space offer powerful tools for exploring mechanisms.
Despite substantial advancements in the identification of prognostic markers in patients with DOC, it is undeniable the research findings exhibit considerable heterogeneity.For instance, in this study, for clinical factors, only etiology and time postinjury continued to show a significant association with prognosis after controlling for the linear effects among variables, partially aligning with previous research [2], [5].The findings of age and baseline diagnoses yielded varying across studies [2], [5].Additionally, our findings highlight that FC B holds as a critical prognostic factor, regardless of whether model-dependent or model-agnostic feature importance analysis methods are employed.This insight will pave the way for the development of novel prognostic markers.Furthermore, our results demonstrate that basic individual prognostic prediction capabilities can be achieved using only clinical and connectivity (with no complex network and topological features).Under specific outcome definitions, this approach can even yield excellent discriminative power.Notably, differences among studies (e.g., the distribution of consciousness levels at baseline, time postinjury, endpoints, and outcome definitions) compromise the foundation for fair comparisons to a certain extent [6], [7].In general, a higher proportion of patients in MCS at baseline and more balanced samples tend to result in better predictive performance [8], [10], [21].Together, our study offers new tools for exploring novel prognostic markers for patients with DOC and for enhancing our understanding of the mechanisms underlying consciousness recovery.It is noteworthy that although our sample size is already at a leading level in the field, the reported effect sizes predominantly fall within the moderate range.Therefore, collaboration among more centers and sites is essential for advancing the discovery of robust prognostic markers and for developing predictive tools that are clinically applicable.
Several limitations/considerations of our study warrant discussion.First, our study focused on the wPLI-based FC in the delta band.Other frequency bands and types of FC need to be explored in the future [62].Second, we did not employ individualized head models because of the absence of high-resolution structural images in some patients and the challenges posed by severe structural damage for accurate head model construction.Third, considering the source-flipping issue, we identified brain states in sensor space, but do not exclude the applicability of directly analyzing in source space (e.g., using the absolute value of the data) in some contexts [63].Fourth, multimodal neuroimaging may capture more prognostic factors.Fusion of multimodal information may further improve the prediction performance [4], [20].We plan to explore the integration of EEG with other modalities in prognostic prediction models in future investigations.

V. CONCLUSION
In summary, we investigated the brain networks of a large cohort of patients with DOC using EEG source-space analysis and brain dynamics, taking a crucial step from sensor space to source space and from diagnostic focus toward prognosis prediction.Our findings revealed that improved patients after the six-month follow-up showed a significantly brain-wide milder degree of low-frequency hypersynchrony at baseline compared to those not improved patients, suggesting a relatively better-preserved network structure.Furthermore, this effect was primarily driven by a specific brain state.Importantly, we demonstrated that state-specific brain networks derived from brain dynamics had superior predictive utility compared to traditional static brain networks.Our research sheds new light on the identification of prognostic biomarkers and provides tools for a more profound understanding of the mechanisms underlying consciousness improvement or recovery.

Fig. 1 .
Fig. 1.Clinical evolution trajectory of patients with disorders of consciousness.Sankey diagram illustrates the clinical classification of etiology, consciousness level (at the admission and six-month followup), and the outcome of all patients (n = 178).The prognosis for improvement is defined as an improved consciousness level (e.g., from VS/UWS to MCS, VS/UWS to eMCS, MCS− to MCS+, or MCS to eMCS) at the six-month follow-up relative to that at the admission (baseline).Pink and blue colors indicate the distribution of improved and not improved patients, respectively.TBI = traumatic brain injury; MCS = minimally conscious state; VS/UWS = vegetative state/unresponsive wakefulness syndrome; eMCS = emerged from MCS.

Fig. 2 .
Fig. 2. Network-level comparisons between improved and improved patients in static FC.Mann-Whitney U test (z-statistic) was performed for the univariate tests.FDR corrected p-values: *p < 0.05 and **p < 0.01.VSN = visual network; SMN = sensorimotor network; DAN = dorsal attention network; SAN = salience attention network; LIM = limbic network; FPN = frontoparietal network; DMN = default mode network; FC = functional connectivity; MCS = minimally conscious state; VS/UWS = vegetative state/unresponsive wakefulness syndrome.The network-level FC is calculated based on the region-to-region FC.The box boundaries represent the 25th and 75th percentiles, with the center line denoting the median.The whiskers extend to the minimum and maximum data values after the outliers (presented as dots) are excluded.

Fig. 3 .
Fig. 3. Microstate maps and comparisons between groups in statespecific FC.Top: Microstate maps in patients with DOC.Bottom: Comparisons between improved and not improved patients in FC B .Mann-Whitney U test (z-statistic) was performed for the univariate tests.Edges with corrected p < 0.05 are displayed.FDR corrected p-values for network-level inferences: *p < 0.05.FC B = functional connectivity of microstate B; VSN = visual network; SMN = sensorimotor network; DAN = dorsal attention network; SAN = salience attention network; LIM = limbic network; FPN = frontoparietal network; DMN = default mode network.The network-level FC is calculated based on the regionto-region FC.

Fig. 4 .
Fig. 4. Comparisons between improved and not improved patients in network measures of FC B .Mann-Whitney U test (z-statistic) was used for the univariate tests.FC B = functional connectivity of microstate B; CC = clustering coefficient; PC = participation coefficient; β 1 = 1dimensional Betti number.The box boundaries represent the 25th and 75th percentiles, with the center line denoting the median.The whiskers extend to the minimum and maximum data values after the outliers (presented as dots) are excluded.FDR corrected p-values: *p < 0.05.

Fig. 5 .
Fig. 5. Feature importance analysis using logistic regression and SHAP values.ET = etiology; TPI = time postinjury; BD = baseline diagnosis; FC = functional connectivity; FC B = functional connectivity of microstate B; OR = odd ratio; VS/UWS = vegetative state/unresponsive wakefulness syndrome; MCS = minimally conscious state; DNN = deep neural network; SHAP = Shapley additive explanations; SMN = sensorimotor network; DMN = default mode network; VSN = visual network; DAN = dorsal attention network.The red color denotes a positive association of the feature with the outcome and the blue color denotes a negative association.

Abstract-Accurate prognostic prediction in patients with disorders of consciousness (DOC) is a core clini- cal concern and a formidable challenge in neuroscience.
Uncovering Brain Network Insights for Prognosis in Disorders of Consciousness: EEG Source Space Analysis and Brain Dynamics Zexuan Hao , Xiaoyu Xia , Yu Pan , Yang Bai, Yong Wang , Bo Peng, and Weibei Dou , Member, IEEE

TABLE I DEMOGRAPHIC
AND CLINICAL CHARACTERISTICS OF PATIENTS WITH DISORDERS OF CONSCIOUSNESS AT THE ADMISSION