How well do neural signatures of resting‐state EEG detect consciousness? A large‐scale clinical study

Abstract

The assessment of consciousness states, especially distinguishing minimally conscious states (MCS) from unresponsive wakefulness states (UWS), constitutes a pivotal role in clinical therapies. Despite that numerous neural signatures of consciousness have been proposed, the effectiveness and reliability of such signatures for clinical consciousness assessment still remains an intense debate. Through a comprehensive review of the literature, inconsistent findings are observed about the effectiveness of diverse neural signatures. Notably, the majority of existing studies have evaluated neural signatures on a limited number of subjects (usually below 30), which may result in uncertain conclusions due to small data bias. This study presents a systematic evaluation of neural signatures with large‐scale clinical resting‐state electroencephalography (EEG) signals containing 99 UWS, 129 MCS, 36 emergence from the minimally conscious state, and 32 healthy subjects (296 total) collected over 3 years. A total of 380 EEG‐based metrics for consciousness detection, including spectrum features, nonlinear measures, functional connectivity, and graph‐based measures, are summarized and evaluated. To further mitigate the effect of data bias, the evaluation is performed with bootstrap sampling so that reliable measures can be obtained. The results of this study suggest that relative power in alpha and delta serve as dependable indicators of consciousness. With the MCS group, there is a notable increase in the phase lag index‐related connectivity measures and enhanced functional connectivity between brain regions in comparison to the UWS group. A combination of features enables the development of an automatic detector of conscious states.

This study investigates validity of diverse biomarkers for consciousness assessment with a large‐scale clinical resting‐state electroencephalography dataset containing 296 subjects collected over 3 years. Alpha power, phase lag index‐related connectivity measures, and delta power represent the most reliable indicators and feature fusion leads to an enhanced automatic detector of consciousness.

Abbreviation

AC

assortativity coefficient

ApEn

approximate entropy

AUC

area under the curve

BC

betweenness centrality

C

central

CC

clustering coefficient

CMD

cognitive motor dissociation

COH

coherence

CRS‐R

Coma Recovery Scale‐Revised

CTRL

control

D

degree

DOC

disorders of consciousness

dwPLI

debiased estimator of the squared WPLI

EC

eigenvector centrality

EEG

electroencephalogram

EMCS

emergence from the minimally conscious state

EOG

electrooculogram

ERP

event‐related potential

F

frontal

fMRI

functional magnetic resonance imaging

GE

global efficiency

GLM

generalized linear model

HC

healthy control

ICOH

imaginary coherence

LDA

linear discriminant analysis

LIS

locked‐in syndrome

LZC

Lempel–Ziv complexity

M

modularity

MCS

minimally conscious state

MI

motor imagery

PC

participation coefficient

PE

permutation entropy

PET

Positron emission tomography

PLI

phase lag index

PLV

phase locking value

PO

parieto‐occipital

PSD

power spectral density

RBF

radial basis function

ROC

receiver operating characteristic

S

strength

SE

spectral entropy

SFS

sequential forward selection

SVM

support vector machines

TBI

traumatic brain injury

TMS

transcranial magnetic stimulation

uPLI

unbiased estimator of the squared PLI

UWS

unresponsive wakefulness state

WPLI

weighted phase lag index

XGBoost

extreme gradient‐boosted decision trees

1. INTRODUCTION

Many studies in recent years have focused on the assessment of the brain condition of patients with disorders of consciousness (DOC). DOC describes states of unconsciousness induced by brain injury or dysfunction of neural systems in control of arousal and awareness (Giacino et al., 2014; Posner et al., 2007). Some patients with DOC remain in the unresponsive wakefulness state (UWS), in which they open their eyes spontaneously but are unresponsive to external stimuli or show just reflex behavior (Gerstenbrand, 1967; Jennett & Plum, 1972; Kretschmer, 1940; von Wild et al., 2012). Other patients may improve to the minimally conscious state (MCS), in which they show fluctuating but reproducible signs of awareness (i.e., visual pursuit in MCS− and command following in MCS+) of themselves or of their environment (Bruno et al., 2011; Giacino et al., 2014). Patients in emergence from the MCS (EMCS) are capable of functional communication and functional use of objects. Diagnosis of DOC patients is crucial for clinical decision‐making, for example, the effect of some treatments on MCS patients is more significant (Thibaut et al., 2014).

Current consciousness assessment is mostly based on the behavioral evaluation, such as probing for visual pursuit using a mirror (Kondziella et al., 2020). However, the rate of misdiagnosis is possibly as high as about 40% (Giacino et al., 2018; Schnakers et al., 2009) due to impaired motor function of patients, lack of repeated assessment, and so forth. Neuroimaging technologies (Claassen et al., 2019; Comanducci et al., 2020; Monti & Schnakers, 2022) have been introduced to complement behavioral information and find “cognitive motor dissociation” patients (Gosseries et al., 2014). The American and European Academies of Neurology emphasize the importance of multimodal assessments (clinical examination, neuroimaging and electrophysiology‐based techniques such as transcranial magnetic stimulation with electroencephalography [TMS‐EEG]) (Casarotto et al., 2016; Wang et al., 2022; He et al., 2018), functional magnetic resonance imaging (Demertzi et al., 2015; Owen et al., 2006), and positron emission tomography (PET) (Stender et al., n.d.) to detect preserved consciousness (Giacino et al., 2018; Kondziella et al., 2020). EEG is a cost‐effective, noninvasive, portable, and bedside measurement. Many studies investigated EEG‐based biomarkers to evaluate the state of consciousness (Sun et al., 2023). Multiple passive paradigms like event‐related potential (Schnakers et al., 2008) and active paradigms like motor imagery (Cruse et al., 2011) have been adopted in DOC patients. Nevertheless, these paradigms are difficult to complete for some DOC patients who have impaired sensory systems and are easily fatigued. Resting‐state EEG, which collects the brain signals in resting states, is more convenient and feasible for such a condition (Lechinger et al., 2013; Stefan et al., 2018).

One critical problem in EEG‐based consciousness detection lies in that, there is still no general agreement on which EEG features are effective and reliable of assessing consciousness. There are some concordant conclusions, such as MCS states usually contain larger nonlinear measures in brain signals compared with UWS states. However, many metrics reveal inconsistent results. For example, the performance of delta power and alpha power is still unclear and controversial (Sitt et al., 2014; Stefan et al., 2018). Furthermore, as for theta power, some studies (Lechinger et al., 2013; Naro et al., 2018; Sebastiano et al., 2015) indicated it was not significantly different for MCS and UWS patients, while other studies (Piarulli et al., 2016; Sitt et al., 2014) demonstrated MCS patients had higher theta power when compared to UWS patients. Therefore, although efforts have been made, it is still ambiguous about which metrics are reliable for EEG‐based consciousness detection. The inconsistent results may be caused by different factors, such as behavioral tasks, artifact removal procedures of brain signals, and the segmentation of data. Another crucial concern is the sample size of the conducted studies. Owing to the challenges encountered in procuring data, the majority of the studies relied on a confined dataset, which may result in data bias.

To investigate what EEG features can reliably reflect the consciousness states, we collected a large‐scale dataset containing 99 UWS, 129 MCS, 36 EMCS, and 32 healthy subjects (296 total) over about 3 years. Based on the dataset, we did a wide‐ranging analysis with different types of EEG features, including (1) spectrum features, (2) nonlinear measures, (3) functional connectivity, and (4) graph‐based measures (see Figure 1 for details). Using a large‐scale dataset can lead to more robust results that are less affected by data bias. To evaluate the reliability of the features, we tested the features with a random‐subset‐division strategy to further mitigate the effects of dataset bias. The large sample size as well as the random‐subset‐division test method is conducive to general and robust results.

FIGURE 1.

The overview of electroencephalography (EEG)‐based metrics.

2. MATERIALS AND METHODS

2.1. Dataset

Most of studies made use of a limited dataset. Figure 2 illustrates the sample sizes of UWS and MCS patients in the literature. We examined the sample sizes of 65 research papers. It is noteworthy that over 50% (33 papers) of the studies had limited sample sizes, amounting to less than 30 participants, while a meager 6.15% (4 papers) accounted for sample sizes exceeding 100.

FIGURE 2.

Sample sizes of unresponsive wakefulness states (UWS) and minimally conscious states (MCS) patients in the literature. (a) The number of MCS patients and UWS patients in the literature. The size and color intensity of the circle indicate the number of UWS and MCS patients. (b) Distribution of samples of both MCS and UWS patients in the literature.

2.1.1. Participants

The research was performed in Hangzhou Mingzhou Brain Rehabilitation Hospital. Data were collected from 99 UWS and 129 MCS from July 2019 to December 2021. We excluded patients with an unstable consciousness state, characterized by signs of spontaneous recovery or deterioration within 1 week. We also did not include the following patients: (1) patients with schizophrenia, schizoaffective disorder, or primary affective disorder and combined severe heart, brain, kidney, liver and hematopoietic system diseases or other serious primary diseases. (2) Patients with frequent and irregular seizures, long‐term sedative use, severe malnutrition, infection, inflammatory bowel disease, irritable bowel syndrome, and drug or alcohol addiction in the last year. (3) Patients who received sedatives such as nitrazepam. Furthermore, centrally acting drugs, neuromuscular function blockers, and sedative drugs were discontinued for at least 24 h before data acquisition. The recordings included 172 males and 56 females, aged from 18 to 81 years (mean age = 54 ± 14 years). Etiology was traumatic in 97 patients and nontraumatic in 131 patients. The EEG recordings were collected in the resting state of patients, and there were one to three recordings from one patient.

Behavioral diagnosis of the DOC patients was based on the highest score of five CRS‐R assessments (Giacino et al., 2020). The Coma Recovery Scale‐Revised (CRS‐R) is a standardized method for assessing neurobehavioral functions. It consists of six subscales, which examine auditory, visual, oromotor, motor, communication, and arousal functions respectively. The CRS‐R involves a total of 23 items, including tasks like visual pursuit and location to sound. The demographic and clinical information, including age, gender, and clinical diagnosis are shown in Table 2 and categorized by etiology in Figure 3 (see Table S1 for details). EEG data from 36 EMCS patients, and 32 healthy subjects were also collected. Table S2 depicts the mean and the standard deviation of recording time.

TABLE 2.

Demographic and clinical description.

Age	Gender		Clinical diagnosis
Mean ± standard deviation	M	F	UWS	MCS
54 ± 14	172	56	99	129

FIGURE 3.

Demographic and clinical information categorized by etiology. Etiology was traumatic in 97 patients and nontraumatic in 131 patients. (a) Age. Error bars illustrate the standard deviation. (b) Gender. (c) Clinical diagnosis. TBI, traumatic brain injury.

Written informed consent was provided by the participants or their legal guardians for the experiments. The study was approved by the Ethical Committee of the First Affiliated Hospital of Zhejiang University (no.NCT20220621), and by Hangzhou Mingzhou Brain Rehabilitation Hospital.

2.1.2. EEG data collection

The signal was recorded in resting‐state conditions using 64 scalp electrodes with a BrainCap (Brain Products DmbH, Munich, Germany) amplifier at a 1000‐Hz sampling rate. Electrooculogram was recorded using one of these electrodes placed under the right eye. Impedances were kept below 10 kΩ. The signal was online referenced against the FCz channel. The EEG signal was bandpass‐filtered between 1 and 40 Hz to reduce low‐frequency ocular and high‐frequency muscular artifacts, using a zero‐phase, forward‐backward finite impulse response filter. Recordings were segmented into 6‐s epochs. Epochs with voltages exceeding ±180 μV were discarded. Five EEG recordings were discarded due to the absence of non‐artifacted epochs. The remaining epochs were rereferenced using the average reference. The filtering and voltage ranges were chosen according to Wang et al. (2022), Lechinger et al. (2013), and Sitt et al. (2014). Then, 332 EEG recordings from the UWS and MCS, as well as 68 recordings from the EMCS and healthy subjects, were obtained. EEG data from one healthy participant was excluded when evaluating connectivity because it presented only one artifact‐free epoch (recordings with at least two artifact‐free epochs were retained).

2.2. EEG‐based metrics

In Table 1, we provide a summary of a subset of EEG‐based metrics we utilize for consciousness detection. Specifically, we divide these metrics into four groups: spectrum measures, nonlinear measures, functional connectivity, and graph‐based measures. These metrics were predominantly employed in the diagnosis of DOC in resting‐state EEG analysis. Additionally, metrics that have been investigated in other neurological conditions, such as Parkinson's disease (Conti et al., 2022), are also selected for evaluation.

TABLE 1.

Overview of metrics.

Type	Method	Finding in	Literature	Our result
Spectrum	Delta	MCS < UWS	(Sitt et al., 2014; Stefan et al., 2018), (Naro et al., 2018) (***) (Piarulli et al., 2016), (Fz, **) (Lehembre et al., 2012; Sebastiano et al., 2015)	MCS < UWS
		No difference	(Lechinger et al., 2013)
	Theta	No difference	(Lechinger et al., 2013; Naro et al., 2018; Sebastiano et al., 2015)	No difference
		MCS > UWS	(Sitt et al., 2014), (Piarulli et al., 2016) (Fz, *)
	Alpha	MCS > UWS	(Sitt et al., 2014; Stefan et al., 2018), (Naro et al., 2018) (***) (Piarulli et al., 2016), (Fz, **) (Lehembre et al., 2012; Sebastiano et al., 2015)	MCS > UWS
		No difference	(Lechinger et al., 2013)
	Beta	MCS > UWS	(Piarulli et al., 2016) (Fz, *)	MCS > UWS
		No difference	(Lechinger et al., 2013; Lehembre et al., 2012; Naro et al., 2018; Sebastiano et al., 2015; Sitt et al., 2014)
	Gamma	No difference	(Lechinger et al., 2013; Lehembre et al., 2012; Naro et al., 2018; Sitt et al., 2014)	No difference
Nonlinear measures	Approximate entropy	MCS > UWS	(Stefan et al., 2018; Wu et al., 2011) (**)	MCS > UWS
	Lempel–Ziv complexity	MCS > UWS	(Wu et al., 2011) (**)	MCS > UWS
	Spectral entropy	MCS > UWS	(Sitt et al., 2014) (Piarulli et al., 2016), (Fz, **) (Didier Ledoux et al., 2011), (**)	MCS > UWS
	Permutation entropy	MCS > UWS	(Sitt et al., 2014) $\left(\mathrm{\theta },\mathrm{\alpha }\right)$ (Lechinger et al., 2013) ($\mathrm{\theta }$, *)	MCS > UWS
Fluctuation	Spectral entropy	No difference	(Sitt et al., 2014)	No difference
		MCS > UWS	(Piarulli et al., 2016) (Fz, **)
Connectivity	Coherence	MCS > UWS	(Cavinato et al., 2015) (posterior, $\mathrm{\alpha }$)	MCS > UWS
		No difference	(Lehembre et al., 2012; Schorr et al., 2016; Stefan et al., 2018)
		MCS < UWS	(Cavinato et al., 2015) (frontal, fronto‐parietal, $\mathrm{\theta }$)
	ICOH	MCS > UWS	(Lehembre et al., 2012) (frontal‐to‐posterior, $\mathrm{\theta }$)	MCS > UWS
		No difference	(Sitt et al., 2014)
	PLI	MCS > UWS	(Lehembre et al., 2012) (frontal‐to‐posterior, $\mathrm{\theta }$; inter hemisphere, $\mathrm{\alpha }$)	MCS > UWS
	dwPLI	MCS > UWS	(Chennu et al., 2017) $\left(\mathrm{\alpha }\right)$	MCS > UWS
Graph theory	dwPLI	MCS > UWS	(Chennu et al., 2017) (participation coefficient, $\mathrm{\alpha }$)	MCS > UWS

^* p < .05.

^** p < .01.

^*** p < .001.

2.2.1. Spectral analysis

Power spectral density (PSD) on each epoch was obtained using the Welch method (Welch, 1967). At each channel, the PSD within five frequency bands, that is, delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–40 Hz) were converted to relative proportions of the total power spectrum density over all five bands. Further, the theta/alpha and (delta + theta)/(alpha + beta) power ratios, mentioned in a previous publication, were estimated (Lechinger et al., 2013). The total power spectrum density was also recorded.

2.2.2. Nonlinear measures

Nonlinear measures have been applied to measure the disorder of neural activities. We applied several representative entropy and complexity measures, and the settings are given in the following context.

1) Permutation entropy

Permutation entropy (PE) is an appropriate complexity measure for chaotic time series, in particular in the presence of noise (Bandt & Pompe, 2002). Corresponding to Sitt et al. (2014), signals were split into subsequences with amplitudes of m elements separated by a time delay t. The signal was low‐pass filtered before computing PE and cutoff frequencies were defined as f = sampling rate/m/t (in our case, sampling rate = 1000). Each subsequence is considered a symbol. PE can then be computed as,

$$\mathrm{PE}=-\sum _{i=1}^{m!}{p}_i\mathit{log}\left(p_i\right)$$ (1)

where $p_i$ is the probability of the ith symbol.

We calculated PE for each channel in the theta, alpha, beta, and gamma bands separately, using an embedding dimension m = 3 and the time delay t = [32, 16, 8, 4], respectively.

2) Approximate entropy

Approximate entropy (ApEn) is a statistic quantifying regularity and complexity and appears to have potential application to a wide variety of time‐series data (Pincus & Goldberger, 1994). Signals with N amplitudes were split into an array of subsequences [x(1), x(2), …, x(N − m + 1)] with amplitudes of m elements. It is computed as follows:

$${\complement }_i^m\left(r\right)=\frac{\mathit{\text{number of}}x\left(j\right)\mathit{\text{when}}\left|x\left(j\right)-x\left(i\right)\right|<r}{N-m+1}$$ (2)

where r is a filtering level. Next, define

$${\Phi }^m\left(r\right)=\frac{\sum _{i=1}^{N-m+1}\ln {\complement }_i^m\left(r\right)}{N-m+1}$$ (3)

ApEn can be defined as

$$\mathsf{\text{ApEn}}={\Phi }^m\left(r\right)-{\Phi }^{m+1}\left(r\right)$$ (4)

We used the parameters: an embedding dimension m = 2, a filtering level r = 20% of the standard deviation of the amplitude values, which was found by previous research to optimize performance (Bruhn et al., 2003). More details of the ApEn calculation can be found in the literature (Aboy et al., 2006).

3) Lempel–Ziv complexity

Lempel–Ziv complexity (LZC) analysis has been applied widely in the biomedical signal analysis as a metric to estimate the complexity of discrete‐time physiologic signals. The calculation of LZC complexity measure is based on a finite symbol sequence (Aboy et al., 2006). We transformed our signal into a binary sequence in such a way: Amplitudes above the mean of the signal were set as the symbol “1,” whereas the others were set as the symbol “0.” LZC complexity measures the number of different patterns included in the symbol sequence. More details of the LZC calculation can be found in the literature (Aboy et al., 2006).

4) Spectral entropy

The power spectrum contains the proportions of power at each frequency. Spectral entropy is a measure quantifying the uniformity or peakedness of the distribution of the EEG power spectrum over specific frequency bins (Inouye et al., 1991). Spectral entropy can be calculated as,

$$\mathrm{SE}=-\sum _f\mathit{RIR}\left(f\right)\mathit{log}\left(\mathit{RIR}\left(f\right)\right)$$ (5)

where RIR is the relative power of frequency bands f, which are delta, theta, alpha, beta, and gamma, defined in “Spectral analysis” section.

2.2.3. Functional connectivity

Functional connectivity quantifies how neuronal patterns of oscillating brain activities of two distinct brain areas relate to each other over a range of frequencies. We adopted several connectivity measures in this study, and the connectivity analysis was implemented with the MNE‐Python package (Gramfort et al., 2013). Functional connectivity values were computed for each subject for each pair of electrodes in delta, theta, alpha, beta, and gamma bands.

1) Coherence

Coherence (COH) was the earliest connectivity method employed in DOC research (Davey et al., 1949). Coherence measured in the frequency domain is comparable with temporal cross‐correlation (Glaser, 2012).

2) Phase locking value (PLV)

Coherence does not separate the phase and amplitude components, whereas phase locking value (PLV) uses only the phase component to explore neural synchronization (Lachaux et al., 1999).

3) Phase lag index

The above metrics suffer from the problem of common sources due to active reference electrodes and volume conduction (Stam et al., 2007). Signals recorded from distinct electrodes may pick up activity from a single source, rather than two different interacting sources, thus resulting in a fake correlation. The phase lag index (PLI) evaluates real interactions by quantifying the asymmetry of the distribution of phase differences between two signals.

4) imaginary coherence

Imaginary coherence (ICOH) only takes the magnitude of the imaginary component of coherency into consideration to mitigate the effect of the common source (Nolte et al., 2004). Unlike PLI, it relies on both the amplitudes of the signals and the magnitude of the phase delay. The absolute value of ICOH was computed.

5) PLI‐related metrics

Besides, Vinck et al. (2011) developed the weighted PLI (WPLI), an unbiased estimator of the squared PLI (uPLI), and a debiased estimator of the squared WPLI (dwPLI) to deal with the discontinuity and sample‐size bias of PLI (Vinck et al., 2011).

2.2.4. Graph theory

Using each of the above metrics, connectivity between channel pairs resulted in a weighted symmetric connectivity matrix (63 × 63) for each subject for five frequency bands (delta to gamma). Each matrix was converted into a binary matrix with a density threshold to remove weak connections and noise. The density, which was in the range of 30–70% in steps of 5%, was the ratio of the number of preserved connections to the maximum possible number of connections. At each value of the density, the thresholded and binarized connection matrix represented a network with electrodes as nodes and nonzero values as edges. Graph metrics on the network were analyzed using the Brain Connectivity toolbox (Rubinov & Sporns, 2010). These graph metrics were averaged across all connection densities.

1) Global metrics

The global metrics included: (1) the global efficiency, which is the average reciprocal of the shortest path length (Latora & Marchiori, 2001); (2) the assortativity coefficient, which is the correlation coefficient between the degrees of nodes on two opposite ends of an edge (Newman, 2002); and (3) the modularity, which is the degree to which the network can be partitioned into topologically disparate and densely linked groups (using the Newman algorithm) (Newman, 2006).

2) Nodal metrics

The nodal metrics included: (1) the participation coefficient, which is the variety of intermodular connections of a node (Guimera & Amaral, 2005); (2) the clustering coefficient, which is the ratio of triangles around a node (Watts & Strogatz, 1998); (3) the degree, which is the number of neighbors of a node; (4) the strength, which is the sum of all neighboring link weights; (5) the eigenvector centrality, which is the corresponding element in the eigenvector for the largest eigenvalue of the adjacency matrix (Batool & Niazi, 2014); and (6) the betweenness centrality, which is the ratio of all shortest paths in the network that cross over a node (Freeman, 1978).

2.2.5. Statistics of EEG features: Average versus fluctuation across epochs

A few papers (Piarulli et al., 2016; Sitt et al., 2014) described that fluctuations were able to discriminate UWS from MCS patients. For spectral analysis and nonlinear measures, both the mean value and the time variability, measured as the mean and the standard deviation across epochs, were calculated.

2.3. Statistical and classification analysis

2.3.1. Significance test

A pairwise comparison was performed between UWS, MCS and EMCS patients. To evaluate the normality and the homogeneity of variance assumptions of t test for the measure distribution, we applied the Shapiro–Wilk test and Levene's test. Nevertheless, most of the features did not meet these assumptions. Thus, we used the nonparametric two‐tailed Mann–Whitney U test. Channel‐wise measures were summarized as the mean and the standard deviation over the 63 electrodes. For further analysis, the 63 electrode locations were lumped into 3 regions similar to (Naro et al., 1845), namely frontal (F) (FP1, FP2, FPz, F1, F2, F3, F4, F5, F6, F7, F8, Fz, AF3, AF4, AF7, AF8, FT7, FT8, FT9, FT10); central (C) (FC1, FC2, FC3, FC4, FC5, FC6, Cz, C1, C2, C3, C4, C5, C6, T7, T8, CP1, CP2, CP3, CP4, CP5, CP6, TP7, TP8, CPz); and parieto‐occipital (PO) (TP9, TP10, P1, P2, P3, P4, P5, P6, P7, P8, Pz, PO3, PO4, PO7, PO8, POz, O1, O2, Oz). Connectivity features in a given band were averaged within and between the aforementioned three brain regions. To quantify the correlation between the cortical activity and CRS‐R scores, Spearman's correlation analysis was used.

2.3.2. DOC state discrimination ability

The area under the curve (AUC) computed from the receiver operating characteristic was adopted to estimate the discrimination power of measures between clinical groups. AUC was used in two different circumstances: (1) AUC was derived from the aforesaid mean or standard deviation without a classification model and (2) AUC was derived from the predicted probability calculated by classification models. In the former case, model‐free AUC smaller than 0.5 was subtracted from one.

Pairwise discrimination between UWS, MCS and EMCS patients was implemented by the classifier: Extreme Gradient‐Boosted Decision Trees (XGBoost). The between‐subjects evaluation was adopted to test the generalizability of classifiers to predict the states of patients whose datasets were left out of training. Recordings of the same patient were assigned to the same set (the training set or the test set). The input features for training the classifiers corresponded to 380 one‐dimensional or multidimensional metrics. The labels were the CRS‐R‐based diagnosis (UWS/MCS/EMCS). Classifiers were constructed using five‐fold stratified cross‐validation. To determine the optimum number and combination of features, feature selection was based on sequential forward selection (SFS).

3. RESULTS

Here, we evaluated the EEG‐based metrics for consciousness detection with a dataset from 99 UWS, 129 MCS, 36 EMCS, and 32 healthy subjects (296 total). The dataset was collected from July 2019 to December 2021. Brain topography of channel‐wise measures (the spectrum, nonlinear measures, and graph theory) for the UWS, MCS, EMCS, and healthy control groups are plotted in Figure 4 and Figures S1–S3. The figures plot the mean value of each group in each channel. The Mann–Whitney U test was performed on each channel independently in topographical analysis.

FIGURE 4.

Topoplots of differential metrics. The panel plots the group‐wise mean of metrics for patients and healthy controls (columns 1 = VS, 2 = MCS, 3 = EMCS, 4 = Healthy control = CTRL). The fifth to seventh columns show whether there was a significant difference between groups in each channel (black: p < .01, light gray: p < .05, white: not significant).

3.1. Relative delta and alpha power together with power ratios are distinguishable between consciousness states

As illustrated in Tables S3 and S4, the value of relative power in delta, alpha, beta, and gamma significantly discriminated the UWS from the MCS. Overall, relative delta power decreased monotonously from UWS patients to healthy subjects. The central region showed reduced values. Significant differences were observed in entire regions between the UWS group and the other groups of patients. The structure in the EMCS was similar to healthy subjects. High delta power is considered to come from cortical deactivation during troughs of these slow oscillations (Frohlich et al., 2021). Relative alpha power was the highest in EMCS patients, particularly in PO regions. Relative alpha power was fairly low in the UWS group, followed by the MCS and the CTRL. Significant differences were observed in entire regions between groups of patients. Alpha power is thought to play a part in information processing with the inhibition of task‐irrelevant regions (Jensen & Mazaheri, 2010; Klimesch, 1999). Relative beta power in subjects was larger as their state of consciousness improved, especially in bilateral central regions. Significant differences were observed in entire regions between the UWS group and the other groups of patients. Beta power is related to object recognition (Sehatpour et al., 2008), perception (Donner et al., 2007), and working memory (Tallon‐Baudry et al., 2001). A similar structure was observed in the gamma band. Compared with existing literature, there were a few studies that demonstrated there were no differences in the power of beta (Lechinger et al., 2013; Lehembre et al., 2012; Naro et al., 2018; Sebastiano et al., 2015; Sitt et al., 2014) and gamma (Lechinger et al., 2013; Lehembre et al., 2012; Naro et al., 2018; Sitt et al., 2014) between UWS and MCS patients. While in our study, the power of beta and gamma exhibited a statistically significant increase in MCS patients. There were also inconsistent conclusions about theta power. In our findings, there were no prominently significant differences in relative theta power between UWS and MCS patients. The conclusions (Lechinger et al., 2013; Naro et al., 2018; Sebastiano et al., 2015) were consistent with ours in the theta power. However, a few studies presented that the power of the theta band was lower in UWS patients than that in MCS patients (Piarulli et al., 2016; Sitt et al., 2014). The (delta + theta)/(alpha + beta) as well as theta/alpha power ratio in subjects decreased gradually with a rise in their behavioral diagnosis. Likewise, some studies (Coleman et al., 2005; Lechinger et al., 2013) found ratios between fast‐wave activity (alpha and beta frequency bands) to slow‐wave activity (delta and theta frequency bands) had a positive correlation with CRS‐R scores. Lechinger et al. (2013) also noted that there was a stronger correlation between the alpha/theta power ratio and CRS‐R scores in DOC patients. It indicated that spectral power was relatively concentrated in the low‐frequency component in unconscious states.

In agreement with Sitt et al. (2014), the fluctuation of alpha power was stronger with a higher state of consciousness. The result demonstrated that MCS patients had instability of behavioral responses and awareness of themselves and their surroundings (Giacino et al., 2002; Majerus et al., 2009). Parietal and occipital regions showed increased values. The structure was similar in the EMCS and the CTRL group. Significant differences were observed in entire regions between groups of patients. Conversely, a stable state of the (delta + theta)/(alpha + beta) power ratio was noted in conscious patients. A similar pattern was observed in the theta/alpha power ratio.

3.2. Nonlinear metrics show larger values with MCS compared with UWS

As illustrated in Tables S3 and S4, most nonlinear measures, except for PE in the beta and gamma bands, were capable of distinguishing between the UWS and the MCS. Nonlinear indices rose with the reemergence of behavioral consciousness (such as ApEn in Figure 4). The UWS group exhibited the lowest values, followed by the MCS and EMCS groups, while the control group had the highest values.

In consistency with Sitt et al. (2014), the enhanced mean and diminished time variability of PE in the theta and alpha band were found as patients became more behaviorally conscious. Piarulli et al. (2016) suggested that spectral entropy time variability was greater in the MCS when compared to the UWS in a single midline channel (Fz, Cz, and Pz). In our study, the standard deviation over the 63 channels of spectral entropy time variability was significantly increased in the MCS in comparison with the UWS.

3.3. Functional connectivity is weak in UWS compared with MCS

Overall, both local and distant synchronization were impaired in UWS patients in comparison with MCS patients (Table S3). Weak long‐range connections, such as frontal‐PO information flows, were found to be signs of unconsciousness (Cavinato et al., 2015; King et al., 1914; Lehembre et al., 2012).

Notably, most of the connectivity measures of alpha oscillations were remarkably strong in EMCS patients. For example, WPLI in the alpha band was shown in Figure 5a. The findings are analogous to strikingly high relative alpha power in the EMCS, demonstrating activated alpha activity at rest.

FIGURE 5.

Group‐wise connectivity‐theoretic and graph‐theoretic topology metrics. (a) Weighted phase lag index (WPLI) connectivity in the alpha band. Intra‐area and inter‐area connectivity measures of alpha oscillations were remarkably strong in emergence from the minimally conscious state (EMCS) patients. Statistical analysis was performed using the Kruskal–Wallis nonparametric test with Dunn's posttest to evaluate the statistical difference between the groups (*p < .05, **p < .01, ***p < .001). C: central; F: frontal; PO: parieto‐occipital. (b) Modularity of uPLI‐based function networks. The panel plots group‐wise modularity averaged across all connection densities for each band. Growing loss of consciousness was accompanied by larger delta modularity. The pattern was reversed in beta and gamma bands. The Kruskal–Wallis test with Dunn's multiple comparisons test was applied (*p < .05, **p < .01, ***p < .001).

3.4. Graph metrics have the discriminative power of states of consciousness

As shown in Figure 4, for the alpha frequency range, the hubs were primarily located in the parietal regions as quantified by the strength of dwPLI‐based connectivity topographs. These regions exhibited increased importance as patients became more behaviorally aware. Parietal regions were crucially engaged in the processing of the functional connectivity network, interacting with many other nodes in the network. As reported by Schiff et al. (2005) and Vogt and Laureys (2005), the precuneus and posterior cingulate, central nodes in the default mode network, served as hubs in a long‐range cortical network. They were involved in awareness and conscious information processing (Kjaer et al., 2001; Long et al., 2016), working memory (Kozlovskiy et al., 2012), self‐processing (Northoff & Bermpohl, 2004), and self‐representation (Lieberz et al., 2021). Reappearance of hub nodes along with conscious states was also found by some other centrality features in PO, fronto‐central and midline areas, suggesting their function in global integration and network resilience (Figure S3).

As shown in Figure 5b, for the delta frequency range, when the synchronization values were estimated by the unbiased estimator of the squared PLI, growing loss of consciousness was accompanied by larger modularity. Low‐frequency band networks were more modularized in unconscious patients, indicating that modules were more segregated in communication and specialized neural processing usually occurred within the same module. The pattern was reversed in high‐frequency ranges: the modularity gradually increased with elevated consciousness levels in beta and gamma bands. Using thresholds did not lead to distortion of graph metric differences. For example, delta band modularity was higher in the UWS than that in the MCS across a wide range of connection densities (Figure S4).

3.5. Evaluation of the discriminative ability of the metrics

To test the stability of significant (p‐value <.05) features over the whole set of the population against the change of sampled subjects, we randomly selected 80% of the datasets without replacement by stratified sampling in pairwise comparisons and calculated the p‐value for these features from the subgroup. We repeated the process for 100 times. In the random‐subset‐division strategy, a significant feature was considered to be reliable if the mean of these p‐values was less than .05 and the proportion of p‐values smaller than .05 occurred at a minimum frequency of 80% over 100 trials (Table S4).

Results showed that, some features, although significantly different between MCS and UWS patients, failed to meet the criteria of reliability (such as relative gamma power). Relative alpha power, PLI‐related connectivity measures and relative delta power were the most reliable signatures of the conscious state. Figure S5a plots the top 20 metrics measured by AUC in pairwise comparisons. All of them were both significantly different and robust. Their correlations with CRS‐R scores are plotted in Figure S5b (see Table S5 for details). For spectral and nonlinear measures, the AUC of their mean of epochs was found to have a significantly positive correlation with AUC of their standard deviation of epochs (Figure S5c).

Model‐free AUC in comparison (Table S6) was lower than model‐based AUC in discrimination (Table S7). Models can deal with nonlinear classification problems with high‐dimensional input (e.g., relative power from every channel rather than a summarized single value‐mean or standard deviation across channels in model‐free comparison), giving more details such as interactions between different brain areas and thus more information about overall brain activities. Model‐based pairwise discrimination between UWS, MCS, and EMCS patients is shown in Figures 6, 7, 8.

FIGURE 6.

Top features for distinguishing unresponsive wakefulness states (UWS) versus minimally conscious state (MCS) patients. (a) Panel plots features ranked in the top 20 of their area under the curve (AUC). (b) Panel plots the distribution of AUC of top 10 metrics of four categories. (c–e) Panels plot top 10 features in spectrum, nonlinear measures, and connectivity. (f) Panel plots top 15 features in graph theory. Error bars illustrate the standard deviation of the mean.

FIGURE 7.

Top features for distinguishing unresponsive wakefulness states (UWS) versus emergence from the minimally conscious state (EMCS) patients. (a) Panel plots features ranked in the top 20 of their area under the curve (AUC). (b) Panel plots the distribution of AUC of top 10 metrics of four categories. (c–e) Panels plot top 10 features in spectrum, nonlinear measures, and connectivity. (f) Panel plots top 15 features in graph theory. Error bars illustrate the standard deviation of the mean.

FIGURE 8.

Top features for distinguishing emergence from the minimally conscious state (EMCS) versus minimally conscious state (MCS) patients. (a) Panel plots features ranked in the top 20 of their area under the curve (AUC). (b) Panel plots the distribution of AUC of top 10 metrics of four categories. (c–e) Panels plot top 10 features in spectrum, nonlinear measures, and connectivity. (f) Panel plots top 15 features in graph theory. Error bars illustrate the standard deviation of the mean.

3.5.1. MCS versus UWS

Figure 6 presents the top features for distinguishing UWS versus MCS patients. Functional connectivity did the best in discriminating UWS from MCS patients. PLV in the beta frequency performed best at distinguishing between UWS and MCS patients. The most discriminative measures in spectrum, nonlinear measures and graph theory were relative alpha power, spectral entropy and the clustering coefficient of PLV‐based gamma connectivity topographs. In the spectrum, relative delta power yielded good results as well. In nonlinear measures, PE in theta and alpha bands, LZC and ApEn also performed well. In connectivity and graph theory, the best features were mainly in high‐frequency bands (alpha‐gamma).

3.5.2. UWS versus EMCS

Figure 7 plots the top features for distinguishing UWS versus EMCS patients. Functional connectivity and graph theory did well in distinguishing the UWS versus the EMCS. The most effective measures in spectrum, nonlinear measures, connectivity and graph theory were the theta/alpha power ratio, PE in the theta band, PLV in the delta band and the strength of COH‐based delta connectivity topographs. In spectrum, relative power in other bands successfully classified patients into UWS/EMCS as well. In nonlinear measures, LZC, ApEn, PE in the alpha band, and spectral entropy were also good at discriminating UWS from EMCS patients. In connectivity, PLV, PLI, and COH were discriminative. In graph theory, centrality measures of PLV‐based and COH‐based connectivity topographs could separate the UWS from the EMCS, with the most informative features predominantly found in low‐frequency bands (delta and theta).

3.5.3. EMCS versus MCS

Figure 8 plots the top features for distinguishing EMCS versus MCS patients. Functional connectivity and graph theory were good at discriminating EMCS from MCS patients. The best‐performing measures in spectrum, nonlinear measures, connectivity, and graph theory were the fluctuation of relative alpha power, PE in the alpha band, PLV in the theta band and the eigenvector centrality of PLV‐based beta connectivity topographs. In the spectrum, relative high frequency power (alpha‐gamma) could segregate the EMCS from the MCS. In nonlinear measures, PE in other bands also had the ability to discriminate consciousness. In connectivity, the top metrics for discriminating awareness were in low‐frequency bands (delta and theta). In graph theory, centrality measures of PLV‐based and COH‐based connectivity topographs were distinguishable between the EMCS and the MCS.

3.6. DOC discrimination performance with feature combination

In this section, we examined the optimal subset of metrics. To handle the high dimensionality of combined features, channel‐wise measures were aggregated by computing the mean, excluding the degree in graph theory. SFS was assessed in comparison to the top metrics method. The top metrics approach evaluated the performance of each metric individually, measured by AUC, and selected the optimal subset consisting of the top N features.

As depicted in Figure 9, the SFS method yielded superior outcomes compared to the top metrics method. In the latter approach, features were selected independently, allowing for the selection of correlated features. In contrast, SFS took into account feature interdependencies and consequently selected more informative sets of features.

FIGURE 9.

Area under the curve (AUC) of feature selection methods using varying number of features.

The average absolute value of Spearman's correlation coefficient was calculated both between‐group (spectrum measures, nonlinear measures, functional connectivity, and graph‐based measures) and within‐group. The correlation matrix of top 10 metrics for distinguishing UWS versus MCS patients measured by AUC in each group is shown in Figure S6. Overall, weaker correlations were shown between‐group as compared to within‐group, highlighting that features of different groups may provide non‐redundant and complementary information.

Multivariate SFS classification offered superior discrimination between UWS and MCS patients. Initially, the classification performance improved with an increasing number of features. SFS with seven features was found to be the most optimal (AUC = 0.74). Performance declined when selecting redundant features, which may induce over‐fitting. Seven features were comprised of the fluctuation of the (delta + theta)/(alpha + beta) power ratio, the delta betweenness centrality of PLV‐based, the gamma modularity of uPLI‐based, the assortativity coefficient of COH‐based, the assortativity coefficient of PLV‐based, the beta participation coefficient of dwPLI‐based, and the beta betweenness centrality of PLV‐based connectivity topographs. Most of them were graph theory features. Using the optimal feature subset, we also constructed two other classifiers: linear discriminant analysis (LDA) and support vector machines (SVMs) with radial basis function (RBF) kernels. Figure S7 shows performance of three classifiers. XGBoost reached an AUC = 0.75 after hyperparameter (the number of trees and learning rate) tuning using five‐fold cross‐validation. LDA and SVM classifiers achieved similar AUC and both of them were lower than XGBoost. One‐way ANOVA with Tukey's multiple comparison test was carried out. No significant difference was shown among three classifiers. XGBoost was preferred since it had the advantage of low hardware complexity.

4. DISCUSSION

We systematically evaluate EEG‐based neural signatures for consciousness states with a large clinical dataset with resting‐EEG signals. By using the large dataset and the random‐subset‐division strategy, more robust conclusions can be reached. Our study find that, relative alpha power, PLI‐related connectivity measures and relative delta power are the most reliable signatures of the conscious state, and the combination of the features facilitates an automatic detector of conscious states. Here, we further discuss the consistency and inconsistency between our findings and the literature.

4.1. Spectrum

A decrease was observed with the degree of consciousness for delta power, while alpha showed an increase with the degree of consciousness. The same results were reported in (Lehembre et al., 2012; Naro et al., 2018; Piarulli et al., 2016; Sebastiano et al., 2015; Sitt et al., 2014; Stefan et al., 2018). A positive correlation between relative alpha power and CRS‐R scores and a negative correlation between relative delta power and CRS‐R scores were found (Lechinger et al., 2013; Naro et al., 2018; Sebastiano et al., 2015). Patients with locked‐in syndrome (LIS) preserve most of their cognitive functions, but their motor output is near null (except for vertical gaze) (Halan et al., 2021). Babiloni et al. (1816) found that LIS patients had increased delta but decreased alpha power in comparison with MCS patients. An increased amount of delta activity was frequently detected in anesthesia (Murphy et al., 2011; Supp et al., 1988) and slow wave sleep (Brown et al., 2010; Franks, 2008). Consequently, the presence of delta activity is related to a lack of consciousness. Alpha waves tended to diminish during sleep (Moini & Piran, 2020). These waves have also been associated with the regulation of inhibitory control. They played a role in governing cognitive processes (von Stein & Sarnthein, 2000). Interestingly, we found relative alpha power to be the highest in EMCS patients.

Beta activity is commonly linked to alertness, active thinking, and cognitive processes (Kirstein, 2007; Sammler et al., 2007). The lack of beta oscillation was observed in cortical damage (Kozelka & Pedley, 1990). In this study, relative beta power was found to be significantly lower in UWS patients compared to MCS patients. Nevertheless, several other studies have reported no differences in beta power between these two patient groups (Lechinger et al., 2013; Lehembre et al., 2012; Naro et al., 2018; Sebastiano et al., 2015; Sitt et al., 2014). It is worth noting that the paradigms used in these studies (Sebastiano et al., 2015; Sitt et al., 2014) were different from ours, as they analyzed evoked potentials, whereas we analyzed resting‐state recordings. Additionally, the sample sizes (17–32 patients) in the studies (Lechinger et al., 2013; Lehembre et al., 2012; Naro et al., 2018) that used resting‐state data were smaller than ours. Figure S8 plots p‐values obtained by modifying the sizes of the selected subgroups in pairwise comparisons (UWS/MCS). There were smaller p‐values and confidence intervals along with larger sample sizes, which may help to explain conflicting conclusions observed in datasets with varying sample sizes. For relative beta power, significant differences were not observed until the sample size was sufficiently large. In contrast, for relative theta power, which showed significant difference between UWS and MCS patients in some papers (Piarulli et al., 2016; Sitt et al., 2014) but not in in our study and (Lechinger et al., 2013; Naro et al., 2018; Sebastiano et al., 2015), confidence intervals remained quite large even when a fair number of patients were included. By utilizing a large dataset and employing a random‐subset‐division strategy, more robust and reliable conclusions can be drawn.

4.2. Nonlinear measures

Our findings are consistent with previous researches that have demonstrated a correlation between elevated nonlinear measures and the state of consciousness. Stefan et al. (2018), Wu et al. (2011), Sarà and Pistoia (2010), and Sarà et al. (2011) described that UWS patients had smaller ApEn and LZC indices than MCS patients, and the healthy controls had the highest. The spectral entropy was increased in MCS as compared to UWS patients. Sitt et al. (2014), Piarulli et al. (2016), and Didier Ledoux et al. (2011) and correlated with the CRS‐R (Didier Ledoux et al., 2011). Besides, greater PE was associated with a higher state of consciousness and had ability to discriminate between an unconscious state and a conscious one in theta and alpha bands (Sitt et al., 2014; Stefan et al., 2018; Thul et al., 2016). Variability of neural activities in normal cognition reflected interactions between dispersed networks of neurons (Guevara Erra et al., 2016; Pakhomov & Sudin, 2013). Hence, nonlinear measures were supposed to be indicative of healthy cognition, while reduced ones may indicate pathologies (Dimitriadis et al., 2015; Garrett et al., 2013).

4.3. Connectivity

Our findings were in agreement with researches that suggested connectivity measures were mainly lower in UWS patients than MCS patients. Particularly, comparable to the startlingly excessive relative alpha power, EMCS patients had high levels of alpha connection in the majority of cases.

A variety of connectivity methods are introduced in DOC researches. Sitt et al. (2014) reported the delta PLI, which decreased with state of consciousness, significantly differentiated the UWS and the MCS. Frontal‐to‐posterior connectivity values of the PLI and the imaginary part of coherence were larger in MCS participants than UWS participants in the theta frequency range (Lehembre et al., 2012). Moreover, the PLI between hemispheres in the alpha band was also greater. When it comes to the debiased weighted PLI, it increased with the consciousness state in the alpha range, particularly between frontoparietal regions (Chennu et al., 2014; Chennu et al., 2017). Another study showed that mean dwPLI in the alpha band was significantly correlated with the CRS‐R (Bareham et al., 2020).

The validity of coherence is controversial. Notably, several studies have reported no significant differences in coherence between UWS and MCS patients (Lehembre et al., 2012; Schorr et al., 2016; Stefan et al., 2018). On the other hand, Cavinato et al. (2015) addressed that alpha coherence was increased in the posterior area in MCS patients and healthy subjects as compared to UWS patients, while theta coherence in frontal and fronto‐parietal areas was the opposite in resting condition (Cavinato et al., 2015). The latter one was in stark contrast to our study, which suggested UWS patients exhibit lower theta coherence than MCS patients. It is worth mentioning that (Cavinato et al., 2015) employed a different region division and a smaller dataset. In order to investigate whether the way of division of electrodes into regions caused inconsistent results, we used the same regions with the same electrode subset as (Cavinato et al., 2015) and theta coherence in fronto‐parietal and frontal areas are depicted in Figure S9. Our results indicated that UWS patients exhibited decreased coherence, irrespective of the way of region division. We noted that (Cavinato et al., 2015) had only 12 UWS and 14 MCS patients in their study, which may lead to bias and different conclusions due to the small sample size.

4.4. Graph theory

We discovered that centrality measures, such as participation coefficients, were elevated in healthy individuals as compared to patients, with hubs mainly found in parietal areas. The parietal cortex, playing a pivotal role in detecting alterations in consciousness, was considered as the hot zone associated with consciousness (Giacino et al., 2014; Naro et al., 2018). In conscious state, the parietal cortex was involved in cortico‐striatal‐thalamic structures (Afrasiabi et al., 2021). In Chennu et al. (2014) and Chennu et al. (2017), the debiased WPLI‐based graph measures were used to distinguish UWS patients from others. There was a significantly positive trend of standard deviations of participation coefficients in the alpha band as CRS‐R scores rose. The standard deviations of participation coefficients in the alpha frequency band were significantly greater in healthy subjects as compared to patients (Chennu et al., 2014). The reemergence of hub regions (consisting of nodes with high participation coefficients) in frontal and parietal areas in the alpha range was observed as states of consciousness improved (Chennu et al., 2017). Moreover, fronto‐parietal connection exhibited a strong correlation with the level of consciousness in DOC patients (Giacino et al., 2014; Naro et al., 2018). It was found to be increased in MCS patients compared to UWS patients (Chennu et al., 2017).

4.5. Assessment strategy for EEG metrics

Lots of studies have employed the p‐value to assess the efficacy of metrics, without taking reliability into account (Cavinato et al., 2015; Lehembre et al., 2012; Naro et al., 2018; Piarulli et al., 2016; Schorr et al., 2016; Sebastiano et al., 2015). The significance of difference of a given metric may change as the samples vary. In pairwise comparisons, to test the robustness of our metrics, we randomly sampled subsets of subjects from a large population. The results demonstrated that while some features, like relative gamma power, showed significant differences between MCS and UWS patients, they did not satisfy the standards for reliability. This indicated that bootstrap sampling aided in identifying both effective and reliable features.

4.6. About improving the performance of detecting DOC

We achieved an AUC of 0.75 with seven features. Recordings with a greater number of channels may yield more precise and valuable information, thereby enhancing classification performance. EEG data in Sitt et al. (2014), Chennu et al. (2017), and Engemann et al. (2018) were recorded with high‐density 256‐channels, four times the number of channels used in this study. Participation coefficient of dwPLI‐based networks achieved the highest discriminative performance through SVM with RBF with an AUC of 0.83 (Chennu et al., 2017). Multivariate classifiers led to better results by comparison with univariate classifiers through a linear kernel SVM (Sitt et al., 2014) and ensembles of decision trees with an AUC of 0.78 (Engemann et al., 2018). Additionally, it is possible to further enhance the discrimination capability with multimodal neuroimaging technology since MRI and PET are reported to perform well at diagnosis and prognosis of DOC patients (Monti et al., 2010; Owen et al., 2006; Stender et al., 2014).

4.7. The evaluation of the impact of time since injury and etiologies

We tested discrimination performance of the most effective indicators of consciousness, the top 10 metrics in the pairwise comparison between UWS and MCS patients measured by AUC (depicted in Figure S5a), across time since injury and etiologies.

Regarding time from injury, we categorized patients into short‐term (up to 3 months) and long‐term (longer than 3 months) subgroups according to Casarotto et al. (2016) and King (2013). The contrast between UWS and MCS patients was more pronounced in the long‐term subgroup compared to that in the short‐term subgroup (Figure S10).

Concerning etiologies, patients were classified into anoxic and non‐anoxic subgroups. The difference between UWS and MCS patients exhibited greater significance in the not‐anoxic subgroup than that in the anoxic subgroup (Figure S11). A recent study implied that alpha power, a well‐established discriminator of conscious states in several studies (Lehembre et al., 2012; Naro et al., 2018; Piarulli et al., 2016; Sebastiano et al., 2015; Sitt et al., 2014; Stefan et al., 2018), distinguished consciousness only among anoxic patients but failed to discern consciousness among not‐anoxic patients (Colombo et al., 2023). However, the difference in indicators of consciousness, such as alpha power, remained significant in not‐anoxic patients in this study.

4.8. Clinical implications and potential applications

Reliable signatures of consciousness suggest residual cognitive functions that may not be evident in behavior. The European Academy of Neurology recommends that a composite reference standard should include not only standardized clinical rating scales, but also EEG and neuroimaging measures (Kondziella et al., 2020). The highest level of consciousness detected by any of these methods confirms the diagnosis. The data‐driven recognition of signatures of consciousness can be applied in large‐scale first screening (Comanducci et al., 2020). Patients who show signs of covert awareness can be chosen for specific care‐taking and rehabilitation strategies (Comanducci et al., 2020; Monti & Schnakers, 2022). For instance, they can engage in brain–computer interfaces to enhance interaction with the environment and restore communication. Additionally, these signatures are valuable for treatment evaluation. For example, alpha power was higher in patients with improved behavior following theta burst stimulation (Wu et al., 2018) and could be beneficial in setting parameters for neuromodulation (Pellicciari et al., 2018). What is more, these measures could be employed in prognosis (Schnakers et al., 2018) and have an effect on goals of care, including life‐sustaining options.

4.9. Limitations

The data for our study was obtained from a single center. In order to mitigate potential biases stemming from demographic imbalances, it is essential to expand the clinical cohorts, ensuring a well‐distributed and balanced representation across demographic subgroups. This can be effectively achieved through collaborative efforts across multiple centers.

5. CONCLUSION

We found EMCS patients demonstrated relatively high alpha power and connectivity. In addition to delta and alpha power, we discovered that beta power could distinguish conscious states. Our findings confirmed other studies' findings that MCS patients exhibited stronger nonlinear measures and connectivity than UWS patients. We also observed increased centrality in healthy individuals, predominantly locating hubs within the parietal regions. The network displayed increased modularity during unconscious states in the low‐frequency range. However, in the high‐frequency range, the modularity progressively increased with higher levels of consciousness. The results proposed that relative alpha power, PLI‐relate connectivity metrics, and relative delta power demonstrated the most reliable signatures of consciousness.

The integration of multiple features leaded to advancement in discerning different states of consciousness. EEG analysis has the potential to supplement behavioral assessment, helpful in diagnosis and prognosis, as well as assessment of particular treatment.

CONFLICT OF INTEREST STATEMENT

The authors report no competing interests.

Supporting information

DATA S1. Supporting Information.

Ma, X. , Qi, Y. , Xu, C. , Weng, Y. , Yu, J. , Sun, X. , Yu, Y. , Wu, Y. , Gao, J. , Li, J. , Shu, Y. , Duan, S. , Luo, B. , & Pan, G. (2024). How well do neural signatures of resting‐state EEG detect consciousness? A large‐scale clinical study. Human Brain Mapping, 45(4), e26586. 10.1002/hbm.26586

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request.