2.1 Participants

This study was part of a larger protocol, previously published by Duclos et al. [28]. Patients were recruited from four different intensive care units in Canada and included if they were at least 18 years old, within 14 days of ICU admission, and continuously sedated following a brain injury (Supporting Information: Methods). All patients were unresponsive at the time of recruitment. Therefore, written informed consent for study participation was provided by the patient's legal representative family member, in accordance with the Declaration of Helsinki. Since 2020, a total of 1878 patients were screened; 183 patients were eligible for the study; consent was obtained for 53 patients; and EEG data were acquired from 47 patients. Six patients were excluded due to increased levels of agitation and head movement. The remaining 41 patients (28 male, 13 female), ranging from 21 to 88 years (mean 54.78, SD 19.30) were included in the analysis (18 patients with anoxic brain injury, 12 stroke patients, 8 TBI patients, 3 other etiologies, see Table 1). The time between admission to the ICU and EEG data acquisition ranged from 0 to 11 days (mean = 2.98, SD 2.04). The score on the GCS at the time of data acquisition ranged from 3T to 9T (mean 4.83, SD 2.06, see Tables 1 and 2). The APACHE-II score ranged from 8 to 44 (n = 38, mean 19.21, SD 8.26). All patients were intubated at the time of recording. The study was approved by the Research Ethics Board of the McGill University Health Centre (Project ID 2020-5972) and the Western University Health Science Research Ethics Board (Project ID 114303).

2.2 Electroencephalography Data and Anesthetic Protocol

EEG data were recorded in five phases: (1) sedation baseline (i.e., during continuous sedation); (2) decreasing sedation level (i.e., during the transition after interruption of sedation); (3) sedation interruption; (4) increasing sedation level (i.e., during the transition after re-initiation of sedation) and (5) sedation reinstated (i.e., after resumption of continuous sedation) (see Figure 1 for schematic illustration of the recording protocol). Weaning of sedation was clinically indicated either with the aim to completely stop sedation or to perform a neurophysiological assessment. Decisions about the length of sedation interruption or necessity to reinitiate sedation were made by the attending nurse based on standard-of-care risk assessment and were not affected by study enrollment and EEG data acquisition. Thus, only the length of Phase 1 (sedation baseline) was standardized at 10 min. The total length of the protocol and dose of sedation varied between patients according to standard of care (Supporting Information: Methods). EEG data were recorded using a high-density 128-channel saline net and an Amps 400 amplifier (Electrical Geodesic Inc., USA). Data were recorded at a sampling rate of 1 kHz and were referenced to the vertex; electrode impedances were reduced to below 5 kΩ prior to recording. Preprocessing was performed using MNE python [29] (Supporting Information: Methods).

Patient responsiveness was assessed by the attending nurse using the GCS [1] three times during the protocol: before Phase 1 (sedation baseline); at the end of Phase 3 (sedation interruption); and at the end of Phase 5 (sedation reinstated) (Figure 1). Analysis was performed separately on the interruption of sedation (i.e., Phase 1–3, n = 41) and the re-initialization of propofol (i.e., Phase 3–5, n = 21, see Supporting Information: Methods). In addition to the EEG analysis, the raw EEG traces were evaluated by a trained neurologist according to the ACNS 2021 guidelines [30]. Ratings are provided in the Supporting Information: Methods (Table S2).

2.3 Feature Extraction

The brain response to the interruption of sedation was estimated using three families of EEG markers: (1) spectral power; (2) spatial ratios; and (3) the spectral exponent. Relative power was calculated in five frequency bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), low beta (13–20 Hz) and high beta (20–30 Hz). Total EEG power was estimated from 1 to 45 Hz. Power spectral density between 1 and 45 Hz was estimated using the Welch algorithm with a 2 s window and 50% overlap. Aperiodic component and offset of the power spectral density were estimated using the “fitting oscillation and one over f” algorithm [31] in the 1–45 Hz range (min_peak_height = 0.1, max_n_peaks = 3, aperiodic_mode = “fixed”).

All features were calculated on each 10 s epoch and channel individually. We then summarized features by extracting: (1) the space-averaged feature strength, (2) the standard deviation of the feature over space, and (3) the posterior–anterior ratio [23].

2.4 EEG Dynamic Change

The EEG dynamic change was defined as the amount of change, weighted by a certainty score. The amount of change was defined as the difference between each feature's Phase 1 (sedation baseline) average and Phase 3 (sedation interruption) average (i.e., Phase 1–3; negative values indicating an increase in response to sedation interruption) (see Figure 2, amount of change). The certainty score was defined by the proportion of datapoints post-interruption (i.e., Phase 2 and 3) that were one standard deviation above or below the Phase 1 (sedation baseline) features (see Figure 2, certainty of change). The resulting value indicates whether differences in the EEG feature were induced by the interruption of sedation or by spontaneous, endogenous fluctuations (Supporting Information: Methods).

EEG dynamic change was calculated individually on the mean, standard deviation, and posterior–anterior ratio of each feature (i.e., total power, delta, theta, alpha, low beta, high beta, exponent and offset), yielding a total of 24 features (i.e., eight mean, eight standard deviation, eight posterior–anterior ratio).

2.5 Grouping According to Behavioral Response

Participants' change in behavioral responsiveness (ΔGCS) during the NWT was defined as the difference between GCS score off sedation and pre-interruption of sedation: ΔGCS = GCS_Phase3 − GCS_Phase1 (Figure 1). ΔGCS and GCS_Phase3 were used to categorize patients into three groups: Nine patients regained ability to follow commands during the NWT and were assigned to “Group A: regained responsiveness” (see Figure 3, green markers; see Table 1). Thirteen patients showed increased levels of behavior during interruption of sedation (i.e., ΔGCS ≥ 2), despite not following commands and were categorized as “Group B: Some behavioral change” (see Figure 3, orange markers). Nineteen patients did not show any or only minimal behavioral change during NWT (i.e., ΔGCS < 2) and were assigned to “Group C: No behavioral change” (see Figure 3, red markers).

2.6 Machine Learning Analysis

We first trained a Linear Ridge regression [32] to predict the ΔGCS of Group A participants based on their EEG dynamic change. For the training, participants' age and propofol concentration were combined with the 24 EEG features. Model fit was evaluated using mean absolute error between true and predicted ΔGCS. The optimal regularization parameter was determined using a leave-one participant out cross-validation. A good model performance indicates that the model has learned to identify EEG patterns that are related to behavioral responsiveness during the NWT.

We then used the model to predict the change in GCS (ΔGCS_predicted) during the NWT in participants who showed ambiguous or absent behavioral response (Group B and C). Behavioral responsiveness can be dissociated from consciousness (i.e., covert consciousness) [7-9]. We therefore did not expect to see a high prediction accuracy for Group B and C. Instead, a high deviation between ΔGCS and ΔGCS_predicted could indicate that participants show electrophysiological signs of waking up despite absent or ambiguous behavioral responsiveness.

2.7 Outcome Assessment and Prognostic Analysis

Patient functional outcome was assessed using a follow-up phone assessment at 3, 6, and 12 months post-recording. Functional outcome was quantified as the maximal Glasgow Outcome scale-extended (GOS-E) [33] score and the minimal Disability Rating Scale (DRS) [34] reached within 1 year post-EEG (Supporting Information: Methods). As this study aimed to investigate the value of EEG in patients with absent or ambiguous behavioral responses, the analysis of prognostic value was only performed on participants from Group B and C (n = 32).

Prognostic analysis was performed based on three outcome criteria: (1) survival (n = 23); (2) recovery of responsiveness (n = 24); and (3) functional outcome (i.e., GOS-E and DRS, n = 20). Nine patients had withdrawal of life sustaining treatment and were excluded from this analysis (see Table 2, see Supporting Information: Methods for more detail). To compare our proposed tool to bedside clinical practice, we additionally collected patients' APACHE score [35] and clinicians' prognostication of patient functional recovery (Supporting Information: Methods).

2.8 Statistical Analysis

The prognostic value of the EEG model to predict survival and recovery of responsiveness was performed using a Python implementation of the one-sided Mann–Whitney U test [36], with the expectation of a larger ΔGCS_predicted for patients with a favorable outcome. The performance of the machine learning model was evaluated using the area under the receiver-operating curve (AUC). It is important to note that the machine learning model only predicted an expected ΔGCS, but not patient outcome. The purpose of the AUC is solely to evaluate how clinically informative the EEG-predicted response was, compared to the observed behavioral response. The difference between AUCs was assessed using a bootstrapping approach, using 100 randomly chosen bootstrap samples (i.e., random selection with replacement from the original dataset), followed by a Wilcoxon signed-rank test. The prognostic value for functional outcome was assessed using a Spearman-rank test.

3.1 Changes in Behavior Are Reflected in the Dynamic Change of Participant EEG

We first analyzed the dynamic changes in EEG of only the participants who regained responsiveness during an interruption of sedation (i.e., Group A, see Figure 3).

The best model performance was reached with a regularization parameter of 10, yielding a mean absolute validation error of 1.36. The features that were most important for the prediction of ΔGCS were the change in low beta, alpha, and theta power, and the spectral exponent.

We performed a partial Spearman rank test (using covariates of age and propofol concentration) as a post-hoc analysis to investigate the features' correlation with ΔGCS in Group A. The analysis confirmed that change in responsiveness in Group A was significantly correlated to the change in low beta power (ρ = −0.90, p < 0.01), alpha power (ρ = −0.81, p < 0.05) and the spectral exponent (ρ = 0.89, p < 0.01). A larger increase in responsiveness was accompanied by a stronger increase in alpha and low beta activity, and a stronger flattening of the spectral exponent following interruption of sedation (i.e., negative values indicating an increase in response to sedation interruption).

This demonstrates that behavioral changes following interruption of sedation are accompanied by changes in the participant's EEG response. In addition, the EEG response following interruption of sedation can be used to predict changes in the GCS following interruption of sedation.

3.2 EEG Features Can Be Decoupled From Participant Behavioral Response

Next, we analyzed the dynamic EEG change of participants who showed an ambiguous (i.e., Group B) or absent behavioral response during an interruption of sedation (i.e., Group C, see Figure 3).

Patients' ΔGCS was predicted with a mean absolute error of 2.06 for group B and 2.27 for group C, with predicted values ranging from −0.18 to 7.05. High values of ΔGCS_predicted indicate the presence of electrophysiological patterns of waking up, similar to participants with high ΔGCS in Group A. Negative values indicate EEG dynamic changes that were even weaker than the training example with the lowest ΔGCS.

Figure 4 illustrates the predicted and observed ΔGCS. Although the predicted difference of most Group B participants matched the observed difference, some participants' ΔGCS_predicted were highly under- or over-estimated. We use four case examples of the largest under- and over-estimated ΔGCS_predicted to elucidate the different types of brain responses during the NWT. Cases 1 and 2 (Patient 33 and 11 in Table 1) showed an absent (ΔGCS = 0) or ambiguous (ΔGCS = 2) behavioral response during the NWT. However, based on the EEG reaction to the interruption of sedation, the model generated a ΔGCS_predicted of 6.56 and 7.06, respectively. In contrast, Cases 3 and 4 (Patient 4 and 7 in Table 1) showed an absent (ΔGCS = 0) or ambiguous (ΔGCS = 3) response during the NWT. In both cases, the model generated a ΔGCS_predicted of −0.18 and 0.04, respectively.

The time-resolved spectrogram and topographic maps of spectral exponent for the four highlighted cases reveal the electrophysiological patterns underpinning the prediction (Figure 4). While Cases 1 and 2 show a prominent increase in global power and flattening of the spectral exponent following sedation interruption, no such effect was observed for Cases 3 and 4.

Altogether, these cases demonstrate that dynamic EEG changes reveal heterogeneity in the group of participants whose response to the NWT was ambiguous or absent. Despite remaining completely unresponsive, some participants showed EEG patterns that are associated with an increased behavioral response. Our results indicate that EEG features can be decoupled from behavioral responsiveness during the NWT.

3.3 EEG Response to Interruption of Sedation Has Prognostic Value for Participants With Absent or Ambiguous Behavioral Responsiveness

Of the 23 participants included in the prognostic analysis of survival, only 11 had an observed ΔGCS above zero during the NWT. The observed ΔGCS had low predictive value for survival, with an AUC of 0.64 (Figure 5A). Participants' maximum observed GCS score was also poorly predictive of survival, with an AUC of 0.62 (Figure 5A). In contrast to the observed behavioral changes, the ΔGCS_predicted had higher prognostic value, with an AUC of 0.75 (t-statistic = 1136.5, p < 0.001) (Figure 5B). Observed GCS scores did not significantly differentiate participants according to their outcome; in contrast, the EEG model generated significantly higher ΔGCS_predicted for participants who survived (U = 84, p < 0.05) (Figure 5B). Figure S1 visualizes this effect for each response group individually.

The model-predicted ΔGCS only minimally increased prognostic value for recovery of responsiveness (Figure S2). No significant correlation was identified between observed or predicted ΔGCS and functional outcome (i.e., GOS-E and DRS score). Figure 5C illustrates the improvements offered by a combined neuro-behavioral assessment during the wake-up test.

Altogether, we demonstrated that the EEG response following interruption of sedation has prognostic value for the survival of participants with absent or ambiguous behavioral responsiveness during the NWT. Based on the EEG response, participants who survived showed a higher ΔGCS_predicted compared to participants who did not.

To compare the effect of sedation interruption (i.e., Phase 1–3) to sedation induction (i.e., Phase 3–5), we performed the same analysis on Phase 3–5 and compared the results in the Supporting Information: Methods (Figure S3). We also performed the analysis on Phase 1 (sedation baseline) and Phase 3 (sedation interruption) individually, without considering the contrast between states (Figure S3). Brain activity during Phase 3 alone (sedation interruption) showed a trend towards separating patients according to the recovery of responsiveness (Figure S3). Our results indicate that both the behavioral and EEG responses to propofol interruption should be considered for patient assessment.

3.4 Prognostic Value of the EEG Response to Interruption of Sedation Can Improve Physician Prediction of Outcome

Physicians' predicted GOS-E score at 6 months positively correlated with participants' GCS_Phase3 (ρ(28) = 0.39, p < 0.05, partial correlation corrected by age). As expected, a patient's behavioral responsiveness strongly influenced the physician's prediction of long-term functional outcome, with more responsive patients correlated with predictions of better recovery.

Physicians rated their predictions with a confidence of 1.82 ± 0.71, with 1 indicating “slightly confident” and 2 being “fairly confident”. The lowest average confidence score (1.8 ± 0.35) appeared in Group B, who showed an ambiguous behavioral response. Overall, there was a high level of uncertainty in the functional outcome of unresponsive patients, which was maximal when patients show an ambiguous behavioral response.

The absolute distance between the real and the physician-predicted GOS-E score was 1.80 ± 1.42 (Supporting Information: Methods; Figure S4). However, when binarizing patient outcomes according to survival (GOS-E > 1) or recovery of responsiveness (GOS-E > 2), physician predictions could not separate good from poor outcomes (Supporting Information: Methods; Figure S5). Similarly, the APACHE II score achieved a maximal prognostic AUC of 0.64 for recovery of responsiveness (Supporting Information: Methods; Figure S5). Our results confirm that prognostication after brain injury in the study population is accompanied by an overall high uncertainty, which illustrates the need for additional clinical assessment tools for this population.