Characterizing electroencephalogram dynamics during sedation withdrawal: Insights into cortical recovery following status epilepticus

Abstract

Objective

Characterize electroencephalogram (EEG) dynamics during propofol withdrawal in patients with generalized convulsive status epilepticus (GCSE) and explore their association with functional outcomes.

Methods

We conducted a retrospective cohort study of adult patients with GCSE who received continuous EEG monitoring and propofol sedation. EEG data were analyzed during two predefined periods: during sedation (0.5–1.5 mg/kg/h) and 4 h after complete sedation withdrawal. Preprocessed EEG signals underwent spectral and complexity analyses, including relative spectral power across frequency bands, alpha‐delta ratio, median frequency (MF), spectral edge frequency 85% (SEF85), and sample entropy. Principal component analysis was applied to EEG trajectory vectors for dimensionality reduction; hierarchical clustering was used to identify distinct evolution patterns during sedation withdrawal. Associations between EEG patterns, SE duration, and clinical outcomes were assessed.

Results

Twenty‐one patients were included. At the population level, EEG features shifted toward faster and more complex activity after sedation withdrawal, with decreased delta power, increased theta, alpha, and beta power, and greater signal complexity. Clustering analysis identified two distinct EEG trajectories: one subgroup (24%) showed marked recovery of cortical dynamics with significant increases in frequency metrics and entropy, while the other (76%) demonstrated minimal changes, with persistent delta‐band dominance and low complexity. Longer SE duration was significantly associated with the minimal‐change group (p = 0.007). No significant differences in functional outcomes were observed between groups.

Significance

EEG dynamics during propofol withdrawal in patients with GCSE evolve along distinct trajectories. Two distinct EEG trajectories emerged: one toward normalization and another with persistently slow, low‐complexity patterns. The majority of patients did not normalize their EEG after sedation withdrawal. Notably, patients with prolonged GCSE were overrepresented in the group lacking EEG recovery. However, these trajectories could not be correlated with functional prognosis. These findings underscore the potential of early EEG trajectories as markers of cerebral recovery.

Plain Language Summary

We studied how brain activity changes after stopping the sedative propofol in adults treated for severe, long‐lasting seizures. Most patients showed little improvement in their brain activity after sedation was stopped, while a smaller group showed signs of recovery. Patients whose seizures lasted longer were more likely to have little or no improvement in their brain activity patterns. These early brain activity patterns did not predict long‐term functional outcomes but may help identify how the brain is recovering after severe seizures.

Key points.

• Propofol withdrawal revealed two distinct EEG trajectories in GCSE patients.

• Few patients showed clear EEG recovery; most remained slow and low complexity.

• Prolonged status epilepticus was associated with the nonrecovery EEG pattern.

1. INTRODUCTION

Generalized convulsive status epilepticus (GCSE) is considered the most severe form of status epilepticus (SE), due to its profound and diffuse disruption of cerebral function and high association with adverse neurological outcomes and increased mortality.^{1, 2, 3, 4} The condition is characterized by sustained, generalized tonic–clonic seizures without recovery of consciousness between ictal events, posing a serious threat to brain integrity.^{5, 6} Several studies have demonstrated that patients who experience GCSE are at substantial risk of irreversible neuronal damage and persistent cognitive impairment, especially when seizures are not rapidly and effectively controlled.^{1, 2, 3, 4}

Timely and effective management of GCSE is therefore critical. Current treatment guidelines emphasize the early termination of seizures through a stepwise pharmacological approach beginning with benzodiazepines followed by second‐line antiseizure medications (ASMs). When first‐ and second‐line ASM fail to achieve seizure control, escalation to anesthetic agents is recommended.^{2, 7, 8} Among the anesthetics recommended in refractory GCSE are propofol, midazolam, and thiopental, each with distinct neurophysiological profile.^{7, 8}

Propofol has been extensively studied in both anesthesiology and neuroscience for its effects on brain activity as measured by the electroencephalogram (EEG). In non‐neurological populations and healthy subjects, propofol induces characteristic EEG signatures that correlate with sedation depth.^{9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19} These changes typically include a shift from baseline activity to diffuse, noncoherent delta rhythms, emergence of coherent frontal alpha oscillations, and suppression of coherent posterior alpha activity. At deeper levels of sedation, EEG patterns progress to high‐amplitude slow‐wave activity and, eventually, to burst‐suppression. Importantly, these EEG dynamics have been shown to be reversible upon withdrawal of the anesthetic agent.^{16, 20}

Despite this well‐documented EEG behavior in healthy and non‐neurological individuals, the evolution of EEG activity during propofol withdrawal in patients with GCSE has not been systematically investigated. Given the altered neurophysiological substrate in SE including neuronal injury, structural and functional alteration of neuronal network, particularly after prolonged seizure activity, it is plausible that EEG recovery patterns may vary significantly in this population.^{5, 6} Understanding how EEG patterns evolve during propofol withdrawal in this context could have both pathophysiological and prognostic implications. It might help identify distinct EEG trajectories that correlate with functional outcomes, offering potential biomarkers for prognosis and therapeutic decision‐making.

The primary objective of this preliminary study is to assess whether distinct EEG dynamics can be identified during the withdrawal of propofol sedation in patients with GCSE requiring sedation. A secondary objective is to evaluate whether these EEG profiles are associated with functional outcomes or the duration of SE.

Our main hypothesis is that EEG evolution during sedation withdrawal is heterogeneous, allowing patients to be stratified into two groups: those whose EEG activity evolves toward normalization after sedation withdrawal, and those whose EEG patterns remain abnormal. We further hypothesize that the persistence of abnormal EEG activity reflects altered cortical function related to the consequences of SE, and may have prognostic implications.

2. MATERIALS AND METHODS

2.1. Design, setting, and population

This retrospective cohort study was conducted at Lille University Hospital between November 2020 and October 2023. We screened all adult patients (≥18 years) admitted to the intensive care unit (ICU) with GCSE who received propofol and underwent continuous EEG (cEEG) monitoring. GCSE was defined using the international league against epilepsy (ILAE) operational definition as an epileptic seizure abnormally prolonged (≥5 min) or repeated at sufficiently brief intervals such that recovery of between episodes does not occur. ⁵ We focused on two distinct time periods during the sedation withdrawal phase:

1. The period during which the propofol dosage was between 0.5 and 1.5 mg/kg/h.

2. The period beginning 4 h after the complete discontinuation of sedation.

Patients were excluded if their cEEG recordings did not contain at least 30 min of data during both of these two defined periods.

Additional exclusion criteria included patients under 18 years of age and those who declined to participate. Nonepileptic dissociative seizures were also excluded and defined as clinical paroxysmal events without any epileptic discharges on EEG. ²¹ Finally, patients whose EEG recordings were rendered unanalyzable due to excessive artifacts were excluded (see flowchart in Figure 1).

FIGURE 1.

Flowchart of the study population. GCSE, generalized convulsive status epilepticus.

2.2. Data collection

2.2.1. Demographic and clinical data

Demographic variables, relevant medical history, and preadmission functional status—assessed using the modified Rankin scale (mRS)—were extracted from medical records. ²² Severity on admission was evaluated using the simplified acute physiology score II (SAPS II), with higher scores indicating a higher risk of poor outcomes. ²³

SE characteristics were recorded. Refractory SE (RSE) was defined according to the previously published definition as persistence of seizures despite administration of first‐ and second‐line ASM.^{24, 25} The status epilepticus severity score (STESS) was used to assess SE severity. ²⁶ Etiologies were classified according to the ILAE framework and grouped as follows ⁵ :

1. Acute cerebrovascular: ischemic or hemorrhagic stroke, cerebral venous thrombosis, posterior reversible encephalopathy syndrome.

2. Acute toxic/metabolic: drug or alcohol intoxication/withdrawal, metabolic disturbance.

3. Acute central nervous system infections.

4. Progressive brain tumors.

5. Progressive autoimmune disorders.

6. Remote stroke.

7. Remote epilepsy of unknown etiology.

8. Other remote etiologies.

9. Unknown etiology.

ASM regimens and dosing were reviewed. All patients received continuous propofol; coadministration of other anesthetic agents was recorded. Sedation parameters included duration, time from induction to withdrawal, initial and maximal infusion rates, and cumulative dose. End of sedation was defined as the last discontinuation of the propofol. Withdrawal interruption was defined as temporary re‐escalation of sedation during weaning.

2.2.2. Electroencephalographic acquisition and preprocessing

cEEG recordings were obtained using a 19‐electrode setup following the international 10–20 system. Scalp electrodes were applied by trained EEG technicians. Signals were sampled at 256 Hz with electrode impedances maintained below 20 kΩ using System Plus Evolution Software® (Micromed, S.p.A., Mogliano Veneto, Treviso, Italy).

Signal preprocessing was performed using the EEGLAB 2025.0 toolbox in MATLAB. ²⁷ A high‐pass filter at 0.5 Hz and a low‐pass filter at 35 Hz (Hamming windowed sinc finite impulse response filters) were applied. The Clean_rawdata EEGLAB plugin was then used to automatically remove artifacted channels, defined as channels with less than 75% correlation with their spatial reconstruction and flatline channels (with a maximum tolerated flatline duration of 5 s). Remaining noisy electrodes were identified and rejected via visual inspection. Rejected channels were spherically interpolated with the condition that the interpolation did not exceed four electrodes. To further suppress nonstationary, large‐amplitude artifacts, the artifact subspace reconstruction method was applied with a parameter of 20.^{28, 29} Signal was then re‐referenced to the average reference. Independent component analysis was performed using the Infomax algorithm to isolate and reject stationary artifacts.^{30, 31} Artifactual components were identified through visual inspection and subsequently rejected. A final visual inspection was conducted to remove any remaining artifacts and epileptiform discharges, ensuring that at least 10 min of clean EEG signals were retained for analysis. No seizures were recorded on the EEG segments included in the analysis.

2.2.3. Electroencephalographic data analysis

We conducted both spectral and complexity analyses on the preprocessed EEG data to gain a comprehensive understanding of brain rhythms.

Spectral analysis was performed using fast Fourier transform with a frequency resolution of 0.1 Hz and 50% overlap, implemented via the EEGLAB toolbox in MATLAB. ²⁷ Absolute spectral power was computed for the standard EEG frequency bands: delta (≥1 and <4 Hz), theta (≥4 and <8 Hz), alpha (≥8 and <13 Hz), and beta (≥13 and <30 Hz). Relative spectral power (RSP) was calculated by normalizing the absolute power in each band by the sum of the absolute spectral power across all bands. The alpha‐delta ratio was defined as the ratio of alpha‐band to delta‐band relative power. In addition, we extracted two spectral features: the median frequency (MF)—the frequency below which 50% of the total spectral power is contained—and the spectral edge frequency at 85% power (SEF85)—the frequency below which 85% of the total spectral power is contained.

EEG complexity was assessed using sample entropy, which quantifies the regularity and unpredictability of a time series: lower values indicate more predictable (i.e., less complex) signals.^{32, 33} Sample entropy was calculated on the final 60 s of artifact‐free data from each patient using the following parameters: embedding dimension m = 2, time lag τ = 1, and similarity threshold r = 0.15. ³⁴

All measures were computed at the electrode level and then averaged to yield both global (mean across all electrodes) and regional values. Regional groupings were defined as follows:

1. Frontocentral: Fp1, Fp2, Fz, F3, F4, F7, F8, Cz, C3, C4.

2. Temporal: T3, T4, T5, T6.

3. Parieto‐occipital: Pz, P3, P4, O1, O2.

2.2.4. Outcomes

The clinical outcomes were:

1. Functional disability at discharge and at 3‐month follow‐up defined as mRS score >1 and higher than baseline.

2. Mortality at discharge and at 3 months.

3. Late seizure recurrence defined as seizure occurring within 3 months following hospital discharge.

4. Duration of sedation and mechanical ventilation.

5. Time to awakening, defined as the delay between sedation withdrawal and clinical awakening, which was characterized by the patient's response to simple commands.

6. Length of stay in the ICU and in the hospital

Functional status, mortality, and seizure recurrence were assessed after discharge by contacting patients, their relatives, or their general practitioner. Seizure recurrence was defined as any seizure requiring emergency care or hospitalization.

2.3. Statistical analysis

Qualitative variables were expressed as counts and percentages. Quantitative variables were expressed as median (first quartile; third quartile). We analyzed population‐level changes in EEG features using Wilcoxon signed‐rank tests to assess independence between periods (before vs. after sedation withdrawal) and EEG features.

For each patient, 32 EEG‐derived feature trajectories were calculated as the difference between presedation and postsedation values: global and regional relative power in the four frequency bands, alpha‐delta ratio, MF, SEF85, and sample entropy. Results from all the patients were considered in a data driven approach to identify independent clusters. We applied the HCPC (Hierarchical Clustering on Principal Components) approach. This method allows us to combine the two standard methods used in multivariate data analyses: principal component methods and clustering. Dimensionality reduction was performed using principal component analysis (PCA) applied to EEG‐derived feature trajectories defined as the difference between postsedation and presedation values of global and regional EEG measures (spectral power, alpha–delta ratio, MF, SEF85, and sample entropy). This approach ensured that PCA captured interpatient variability in EEG changes while preserving the one‐to‐one mapping between presedation and postsedation features. Principal components explaining at least 90% of the total variance were retained. Then a hierarchical clustering was applied to these components. We used the agglomerative hierarchical clustering algorithm for its suitability in combination with the outcomes of a PCA. It has two main advantages: it works from the dissimilarities between the objects to be grouped together and allows a visualization of the results as a dendrogram which shows the progressive grouping of the data.

The algorithm proceeds in an iterative way. The process starts by calculating the dissimilarity between the N objects to be clustered, then two objects which, when clustered together, minimize a given agglomeration criterion are clustered together, leading to the creation of a class with these two objects. Lastly, the dissimilarity between this class and the N‐2 other objects is calculated using the agglomeration criterion. The two objects or classes of objects whose clustering together minimizes the agglomeration criterion are then clustered together. The process continues until all the objects have been clustered. Here we used Euclidean distance as a measure of dissimilarity and Ward's method as the aggregation method. The number of clusters was fixed at two in line with our hypothesis and to reduce the risk of overfitting given the limited sample size.

To validate whether the two identified clusters reflected distinct EEG evolutions—precisely, one group evolving toward EEG normalization after sedation withdrawal and another with persistently abnormal EEG patterns—we compared changes in EEG features between clusters using the Wilcoxon rank‐sum test. This approach assessed the statistical independence of EEG‐derived feature trajectories with respect to cluster membership. To account for multiple comparisons, p‐values were adjusted using the false discovery rate correction.

Separately, associations between cluster membership and clinical outcomes or SE duration were evaluated using the Wilcoxon rank‐sum test for continuous variables and Fisher's exact test for categorical variables.

All statistical analyses were conducted in MATLAB version R2025a ³⁵ with the exception of hierarchical clustering, which was performed using XLStat software (Addinsoft. XLSTAT 2019: Data Analysis and Statistical Solution for Microsoft Excel. Paris, France). All tests were two‐sided. The threshold for statistical significance level was set at p < 0.05.

3. RESULTS

3.1. Population characteristics

From November 2020 to October 2023, 74 patients were hospitalized for GCSE and required sedation. cEEG monitoring was performed in 51 patients, of whom 21 (41%) met the inclusion criteria for this study (see flowchart in Figure 1).

The population characteristics are summarized in Table 1. Initial treatment included benzodiazepines in 18 patients (86%), of whom 16 received clonazepam and two received diazepam, and second‐line ASM in six patients (29%), most frequently fosphenytoin (four patients, 19%). All patients received propofol; seven (33%) were treated with a combination of propofol and midazolam. None of the patients received thiopental or ketamine. The etiologies of SE are detailed in Table 2, with brain tumors, acute cerebrovascular events, and acute toxic or metabolic disturbances each accounting for the most common etiologies affecting four patients each (19%).

TABLE 1.

Population characteristics.

Variable	All patients (n = 21)
Demographics and medical history
Age, median [IQR]	54 [33; 69]
Male, n (%)	15 (71.4)
mRS score 0–1, n (%)	15 (71.4)
Previous epilepsy, n (%)	9 (42.9)
Previous status epilepticus, n (%)	4 (19.0)
Cerebral lesion, n (%)	11 (52.4)
Status epilepticus characteristics
SAPS II, median [IQR]	52 [44; 66]
Focal onset, n (%)	2 (9.5)
Out‐of‐hospital onset, n (%)	13 (61.9)
Intermittent seizures, n (%)	14 (66.7)
Duration <30 min a n (%)	7 (46.7)
Unknown onset time, n (%)	6 (28.6)
Refractory status epilepticus, n (%)	2 (9.5)
STESS, median [IQR]	3 [2; 4]
Sedation characteristics
Propofol dose at initiation (mg/kg/h), median [IQR]	2.9 [1.7; 3.0]
Propofol maximum (mg/kg/h), median [IQR]	3.8 [3.4; 5.3]
Sedation duration (hours), median [IQR]	44 [35; 87]
Time before sedation withdrawal (hours), median [IQR]	21 [12; 30]
Sedation withdrawal duration b (hours), median [IQR]	16 [10; 23]
Withdrawal interruption, n (%)	5 (23.8)
Time to awakening (hours), median [IQR]	3 [1; 31]

Note: Sedation withdrawal duration was defined as the time from the beginning of withdrawal to complete cessation. Withdrawal interruption was defined as temporary re‐escalation of sedation during weaning.

Abbreviations: IQR, interquartile range; mRS, modified Rankin scale; STESS, status epilepticus severity score.

^a Analysis conducted only in patients with known onset time (n = 15).

^b Analysis conducted only in patients without withdrawal interruption (n = 16).

TABLE 2.

Etiologies of status epilepticus.

Etiology	All patients (n = 21)
Acute, n (%)	9 (42.9)
Acute cerebrovascular	4 (19.0)
Ischemic stroke	2 (9.5)
Hemorrhagic stroke	1 (4.8)
PRES	1 (4.8)
Acute toxic or metabolic	4 (19.0)
Alcohol intoxication or withdrawal	1 (4.8)
Metabolic disturbance	1 (4.8)
Drug intoxication or withdrawal	2 (9.5)
CNS infection	1 (4.8)
Progressive, n (%)	5 (23.8)
Brain tumor	4 (19.0)
Autoimmune disorder	1 (4.8)
Remote, n (%)	7 (33.3)
Stroke	2 (9.5)
Epilepsy of unknown etiology	2 (9.5)
Others	3 (14.3)
Unknown, n (%)	0 (0.0)

Note: Etiology was classified according to the international league against epilepsy classification.

Abbreviations: CNS, central nervous system; PRES, posterior reversible encephalopathy syndrome.

3.2. EEG dynamics during sedation withdrawal

3.2.1. Evolution of EEG features at the population level

The detailed evolution of quantitative EEG features during sedation withdrawal at the population level is available in supplementary materials (Tables S1). A significant decrease in RSP in the delta frequency band was observed between the period before sedation withdrawal (propofol dosage between 0.5 and 1.5 mg/kg/h) and 4 h following complete withdrawal. Conversely, RSP in the theta and alpha frequency bands showed an overall significant increase following sedation withdrawal. This enhancement was observed both globally and regionally, with the alpha band increase being particularly notable in the parieto‐occipital region (2.6% vs. 5.1%, p = 0.012). The alpha‐delta ratio also increased after sedation withdrawal, both globally and across individual regions, suggesting a shift toward faster cortical rhythms following sedation cessation.

The SEF85 increased globally after sedation withdrawal versus before (4.1 vs. 5.0 Hz, p = 0.011). Similarly, the MF was significantly higher after sedation withdrawal across the entire scalp (1.4 vs. 1.9 Hz, p = 0.002), indicating a global shift toward faster frequency components.

Finally, signal complexity, as assessed by sample entropy, increased globally after sedation withdrawal (2.2 × 10⁻² vs. 3.6 × 10⁻²), although this change did not reach statistical significance.

3.2.2. Clustering analysis of EEG changes during sedation withdrawal

To explore EEG dynamics during propofol withdrawal we performed dimensionality reduction using PCA on EEG‐derived feature trajectories, followed by HCPC. The evolution of quantitative EEG features by cluster is summarized in Figure 2, while a detailed comparison of feature changes during sedation withdrawal across clusters is provided in Table 3. EEG dynamics during sedation withdrawal differed markedly between the two clusters.

FIGURE 2.

Evolution of quantitative electroencephalogram features by cluster. Median values of each quantitative EEG feature are shown for the periods before sedation withdrawal (propofol dosage between 0.5 and 1.5 mg/kg/h) and 4 h after sedation cessation, stratified by cluster (Panel A: Cluster 1; Panel B: Cluster 2).

TABLE 3.

Compared evolution of quantitative EEG features during sedation withdrawal.

Variable	Cluster 1	Cluster 2	p
	(n = 5)	(n = 16)
Delta‐band relative spectral power (%), median [IQR]
Global	−12.1 [−15.1; −11.0]	−3.9 [−5.6; −1.1]	0.007
Frontocentral	−11.7 [−13.8; −10.3]	−4.0 [−6.0; −0.7]	0.007
Temporal	−12.0 [−16.2; −10.5]	−3.3 [−5.0; −0.7]	0.007
Parieto‐occipital	−12.7 [−19.0; −11.4]	−4.3 [−7.3; −2.0]	0.007
Theta‐band relative spectral power (%), median [IQR]
Global	8.9 [2.9; 11.8]	2.9 [0.6; 3.9]	0.065
Frontocentral	8.8 [2.2; 10.8]	3.1 [0.2; 3.6]	0.091
Temporal	9.2 [4.1; 11.0]	1.9 [0.9; 3.9]	0.046
Parieto‐occipital	8.7 [3.7; 13.5]	3.0 [1.3; 5.2]	0.039
Alpha‐band relative spectral power (%), median [IQR]
Global	6.1 [2.2; 6.9]	0.7 [−0.4; 1.3]	0.009
Frontocentral	6.2 [2.3; 6.9]	0.7 [−0.5; 1.4]	0.009
Temporal	5.1 [1.9; 7.3]	0.6 [0.0; 1.5]	0.009
Parieto‐occipital	5.8 [2.5; 6.5]	0.6 [−0.2; 1.6]	0.008
Beta‐band relative spectral power (%), median [IQR]
Global	0.5 [0.5; 1.6]	0.1 [−0.1; 0.2]	0.010
Frontocentral	0.6 [0.3; 1.6]	0.1 [−0.2; 0.2]	0.011
Temporal	0.5 [0.3; 1.9]	0.1 [−0.1; 0.2]	0.011
Parieto‐occipital	1.0 [0.3; 1.4]	0.1 [−0.1; 0.3]	0.021
Alpha‐delta ratio (×10−2), median [IQR]
Global	9.5 [3.0; 13.7]	0.9 [−0.3; 2.3]	0.006
Frontocentral	8.6 [3.0; 15.3]	1.0 [−0.7; 2.3]	0.011
Temporal	7.9 [2.5; 13.8]	0.8 [0.0; 2.6]	0.014
Parieto‐occipital	9.0 [4.4; 13.1]	0.9 [0.1; 2.7]	0.008
Spectral edge frequency 85% (Hz), median [IQR]
Global	2.8 [2.6; 4.3]	0.7 [−0.2; 1.5]	0.009
Frontocentral	2.9 [2.5; 4.4]	0.7 [−0.2; 1.2]	0.007
Temporal	3.2 [2.5; 4.3]	0.7 [−0.1; 1.4]	0.007
Parieto‐occipital	2.9 [2.7; 4.2]	0.7 [0.2; 1.4]	0.007
Median frequency (Hz), median [IQR]
Global	1.0 [0.7; 1.4]	0.3 [0.1; 0.5]	0.007
Frontocentral	0.9 [0.7; 1.4]	0.3 [0.1; 0.4]	0.007
Temporal	0.8 [0.7; 1.6]	0.3 [0.0; 0.4]	0.007
Parieto‐occipital	1.2 [0.7; 1.7]	0.2 [0.1; 0.5]	0.007
Sample entropy (×10−2), median [IQR]
Global	5.5 [4.4; 11.3]	0.1 [−0.9; 1.5]	0.007
Frontocentral	4.8 [3.0; 11.4]	0.1 [−1.4; 1.7]	0.010
Temporal	6.0 [4.0; 12.8]	0.2 [−0.5; 1.2]	0.007
Parieto‐occipital	8.4 [4.7; 11.6]	0.3 [−1.0; 0.9]	0.007

Note: p‐values highlighted in bold are significant at the 0·05 level.

Abbreviations: EEG, electroencephalogram; IQR, interquartile range.

Cluster 1 included five patients (24%) and exhibited pronounced EEG changes indicative of a shift toward faster, more complex, and physiologically organized cortical activity following sedation cessation. In contrast, Cluster 2 comprised 16 patients (76%) and was characterized by persistent slow, low‐complexity EEG activity, suggesting limited recovery of underlying cortical network dynamics following sedation withdrawal.

Across all cortical regions, the decrease in delta‐band relative power was significantly greater in the Cluster 1 than in Cluster 2 (global: −12.1% vs. −3.9%, p = 0.007), with the most pronounced decreases in the parieto‐occipital region. Conversely, RSP in the alpha, and beta bands increased significantly more in Cluster 1 than in Cluster 2 (global alpha: 6.1% vs. 0.7%, p = 0.009; global beta: 0.5% vs. 0.1%, p = 0.010). Changes in RSP in the theta band followed a similar pattern, reaching significance in temporal and parieto‐occipital regions. The alpha‐delta ratio also increased more in Cluster 1 across all regions (global: 9.5 × 10⁻² vs. 0.9 × 10⁻², p = 0.006). Measures of spectral central tendency further distinguished the clusters. Both MF and SEF increased significantly more in Cluster 1 compared to Cluster 2 (global SEF: 2.8 vs. 0.7 Hz, p = 0.009; global MF: 1.0 vs. 0.3 Hz, p = 0.007). Increase in signal complexity, as quantified by sample entropy, was also significantly greater in the Cluster 1, particularly in the parieto‐occipital region (parieto‐occipital: 8.4 × 10⁻² vs. 0.3 × 10⁻², p = 0.007).

3.3. Relationship between outcomes and EEG dynamics

Clinical outcome data across clusters are summarized in Table 4. There were no statistically significant differences between the two groups with respect to most clinical characteristics. Exceptions included prolonged SE duration, history of epilepsy (cluster 1: 0 patients (0%), cluster 2: nine patients (56%), p = 0.045), out‐of‐hospital onset (cluster 1: one patient (20%); cluster 2: 12 patients (75%), p = 0.048) and acute toxic or metabolic etiology (cluster 1: three patients (60%); cluster 2: one patient (6%), p = 0.028). The time between the two studied periods did not differ significantly between clusters (Cluster 1: 4.5 h [4.5; 5.2]; Cluster 2: 4.5 h [4.5; 4.5], p = 0.857). Overall, no statistically significant associations were found between EEG trajectory clusters and clinical outcomes; however, prolonged SE episode lasting more than 30 min was significantly more common in the second cluster (eight patients) compared to the first cluster (no patient; p = 0.007).

TABLE 4.

Clinical outcomes by cluster.

Variable	All patients	Cluster 1	Cluster 2	p
	(n = 21)	(n = 5)	(n = 16)
Functional disability at discharge, n (%)	10 (47.6)	2 (40.0)	8 (50.0)	1.000
90‐day functional disability a , b n (%)	6 (31.6)	2 (40.0)	4 (28.6)	1.000
90‐day mortality b n (%)	1 (5.0)	0 (0.0)	1 (6.7)	1.000
90‐day seizure recurrence a , b n (%)	2 (10.5)	0 (0.0)	2 (14.3)	1.000
Sedation duration (hours), median [IQR]	44 [35; 87]	44 [39; 114]	49 [31; 89]	0.780
Mechanical ventilation duration (days), median [IQR]	4 [2; 10]	4 [3; 9]	5 [2; 10]	0.780
Time to awakening (hours), median [IQR]	3 [1; 31]	11 [2; 30]	3 [1; 37]	0.858
ICU length of stay (days), median [IQR]	9 [6; 15]	10 [6; 21]	9 [6; 15]	0.955
Hospital length of stay (days), median [IQR]	20 [12; 34]	20 [16; 72]	20 [10; 33]	0.495

Note: Functional disability was defined as a modified Rankin scale score >1 and higher than baseline. Seizure recurrence was defined as any seizure requiring emergency care or hospitalization within 90 days after discharge.

Abbreviations: ICU, intensive care unit; IQR, interquartile range.

^a Patients who were deceased at follow‐up were excluded from the analysis.

^b One missing value.

4. DISCUSSION

In this study, we observed that the evolution of EEG background activity following propofol withdrawal in patients with GCSE is heterogeneous. Through clustering analysis of dynamic quantitative EEG parameters, we identified two distinct patient profiles, suggesting divergent neurophysiological trajectories during sedation withdrawal.

The first profile (n = 5, 24% of patients) demonstrated a trajectory consistent with a progression toward EEG background normalization. This was characterized by an increase in the RSP in the theta, alpha, and beta frequency bands, a decrease in the RSP in the delta band, as well as increases in the alpha‐delta ratio, MF, SEF85, and sample entropy. Together, these trends indicate a shift from slow, synchronous activity toward faster and more complex EEG patterns, which is expected after sedation withdrawal.

In contrast, the second profile (n = 16, 76% of patients) did not show any clear trajectory toward EEG normalization after sedation withdrawal. These patients displayed persistently high RSP in the delta band and low RSP in faster frequency bands (theta, alpha, beta), along with low alpha‐delta ratio, SEF85 and sample entropy. Their EEG patterns remained dominated by slow and low‐complexity signals. This absence of evolution toward normalization is particularly striking given that sedative effects should no longer be pharmacologically active 4 h after discontinuation, suggesting a persistent influence of SE‐related cerebral dysfunction.^{16, 20}

To our knowledge, this is the first study to investigate the dynamic evolution of EEG following propofol withdrawal specifically in the context of GCSE. Prior studies in healthy individuals have shown that propofol induces stereotyped EEG patterns, including diffuse delta oscillations and highly coherent frontal alpha activity.^{16, 19, 20} This phenomenon, often referred to as anteriorization, is believed to result from altered thalamocortical dynamics. ³⁶ Propofol's potentiation of GABAergic activity within both the cortex and thalamus leads to an enhancement of the thalamus's intrinsic alpha oscillations.^{36, 37} Concurrently, propofol interferes with the generation of the posterior dominant rhythm by disrupting the activity of high‐threshold thalamocortical neurons.^{36, 38, 39} These effects are typically reversible: following cessation of sedation, EEG activity in healthy individuals progressively evolves toward the reappearance of the posterior dominant rhythm.^{16, 19, 20}

In our cohort, this expected evolution was seen only in the first group. We hypothesize that, beyond the neurophysiological effects of sedation, the cerebral networks in the second group remain profoundly altered due to the consequences of prolonged seizure activity. These may include excitotoxic neuronal injury, synaptic remodeling, and large‐scale network disruptions.^{4, 5, 6} In essence, the EEG dynamic observed after sedation withdrawal likely reflects the interplay between reversible, sedation‐induced neurophysiological changes and more persistent, seizure‐related brain dysfunction. This interpretation is consistent with current pathophysiological models of SE, which a cascade of neurobiological changes involving dysregulated inhibitory and excitatory signaling, blood–brain barrier disruption, neuroinflammation, cellular injury, and connectivity alterations.^{4, 5, 6} These processes may impair the capacity for normal oscillatory behavior to re‐emerge after sedation is withdrawn, resulting in persistently abnormal, low‐complexity EEG activity in some patients.

Importantly, our findings may have potential prognostic significance. The presence or absence of a clear EEG trajectory toward normalization could reflect differences in the degree of underlying cerebral recovery. Patients whose EEG demonstrates early improvements may represent a subgroup with more preserved neural network integrity and, consequently, better cognitive and functional outcomes. In contrast, the absence of such evolution may indicate persistent disruption in cortical dynamics, with a need for more intensive rehabilitation and long‐term monitoring. Interestingly, we found that a prolonged duration of SE, specifically episodes lasting more than 30 min, was significantly more common in the group that failed to show EEG normalization after sedation withdrawal. This 30‐min threshold corresponds to the conceptual time point T2 as defined by the ILAE, which marks the duration beyond which seizure activity is more likely to cause long‐term neuronal injury, synaptic reorganization, and disruption of large‐scale brain networks.^{4, 5, 6} The overrepresentation of patients with prolonged seizures in the cluster that failed to show EEG normalization reinforces the idea that sustained epileptic activity may lead to persistent impairments in cortical function. These impairments are likely to interfere with the brain's ability to restore normal oscillatory patterns, even after the sedative effects have dissipated.

Other clinical factors, including acute toxic etiology, pre‐existing epilepsy, and out‐of‐hospital SE onset, showed weaker associations with EEG non‐normalization. Acute toxic SE may be less prone to prolonged seizure duration, potentially accounting for its lower representation in this cluster. In contrast, epilepsy is associated with interictal network disorganization and may confer increased vulnerability to the deleterious effects of SE on cortical dynamics. Similarly, out‐of‐hospital onset may predispose to longer seizure durations, indirectly contributing to impaired EEG recovery. However, comparisons between clusters were exploratory in nature; definitive statistical inference will require validation in larger cohorts.

Although we did not identify statistically significant differences in clinical outcomes between the two clusters, the interpretation is limited by the relatively small sample size and residual heterogeneity within groups. Further studies with larger cohorts are necessary to validate these findings, explore the possibility of additional subgroups, and clarify the relationship between EEG recovery trajectories and long‐term neurological outcomes. Moreover, the outcome measures employed may lack the sensitivity required to detect subtle impairments in higher‐order cognitive functions, which are often affected in patients with SE.^{3, 4} Alpha and beta oscillations are known to play key roles in cognitive and perceptual processing.^{38, 40, 41, 42} Therefore, the spectral differences observed between clusters may reflect differential impacts of SE on these higher cortical functions.^{3, 4} This hypothesis warrants further investigation, ideally incorporating more nuanced neurocognitive assessments and longitudinal follow‐up.

More broadly, the identification of reproducible EEG biomarkers of cerebral recovery could have implications beyond SE, including for other acute encephalopathies and disorders of consciousness. As quantitative EEG methods become increasingly accessible, they offer a powerful, noninvasive window into brain function.

The strengths of our study include the homogeneity of sedative agents used across patients, and the integration of several complementary quantitative EEG metrics to capture cerebral dynamics. Moreover, the use of an unsupervised clustering approach enabled the identification of distinct patient subgroups based on EEG trajectories, rather than relying on predefined thresholds or arbitrary classifications. To mitigate the risk of overfitting in the clustering analysis, we adhered to the rule proposed by Formann requiring at least 2ⁿ subjects per n clustering variables and fixed the number of clusters to 2. ⁴³

However, several limitations must be acknowledged. This was a single‐center, observational study with a relatively small sample size, which limits the generalizability of our findings. Some patients were excluded due to incomplete EEG data over the relevant timeframe, which may have introduced selection bias. The included patients also exhibited considerable heterogeneity in baseline characteristics such as age, comorbidities, SE etiology, pre‐existing neurological status, and timing of interventions, all of which can influence both the clinical course of SE and the trajectory of cerebral recovery. This heterogeneity is not unique to our cohort but represents a well‐recognized challenge in SE research, complicating efforts to identify reliable biomarkers and develop prognostic tools. ⁴⁴

Despite its exploratory nature, this study contributes to a growing body of literature advocating for more nuanced, data‐driven approaches to understanding SE. The clustering of patients based on EEG dynamics rather than clinical phenotype alone may offer a more granular view of recovery trajectories. Future studies should aim to integrate multimodal data—including imaging, serum biomarkers, and neuropsychological testing—with EEG findings to better characterize patient subtypes. Additionally, serial EEG recordings over longer timeframes could further clarify the temporal dynamics of recovery and help distinguish transient sedation‐related changes from persistent network dysfunction.

Despite these limitations, this preliminary study provides new insights into EEG dynamics during propofol withdrawal in patients with GCSE. It sheds light on the underlying pathophysiological processes of SE. Importantly, this study underscores the value of data‐driven classification approaches that go beyond conventional clinical categorizations of SE. EEG‐based phenotyping may offer a more sensitive framework for assessing cerebral function, particularly in heterogeneous patient populations. This heterogeneity presents a significant challenge for biomarker development and outcome prediction and highlights the need for personalized, multimodal strategies.

In addition to its clinical implications, this study contributes a methodological framework for future work, including the use of dynamic EEG metrics to study brain recovery in critically ill patients. Subsequent studies should aim to replicate these findings in larger cohorts, integrate longitudinal EEG recordings, and combine electrophysiological data with imaging, biomarkers, and neuropsychological assessments. Such integrative approaches will be essential to evaluate the prognostic and therapeutic utility of EEG biomarkers.

5. CONCLUSION

EEG dynamics following propofol withdrawal in patients with GCSE is heterogeneous, with two distinct recovery trajectories emerging. While a subset of patients exhibited EEG patterns consistent with normalization, the majority showed persistently abnormal activity. These findings suggest that EEG normalization after sedation withdrawal cannot be considered a realistic expectation. Early EEG trajectories could serve as dynamic biomarkers of recovery. Although limited by small sample size and cohort heterogeneity, this study underscores the value of quantitative EEG and clustering methods for identifying clinically meaningful patient subgroups.

AUTHOR CONTRIBUTIONS

Mathieu Dhoisne: conceptualization (equal); methodology (equal); data curation (lead); formal analysis (lead); writing—original draft preparation (lead); writing—review and editing (equal). Philippe Derambure: conceptualization (equal); methodology (supporting); writing—review and editing (equal). Arnaud Delval: conceptualization (supporting); methodology (supporting); formal analysis (supporting); writing—review and editing (equal). Saad Nseir: writing—review and editing (equal). Nacim Betrouni: conceptualization (supporting); methodology (equal); formal analysis (supporting); writing—review and editing (equal). Romain Tortuyaux: conceptualization (equal); methodology (supporting); writing—review and editing (equal).

FUNDING INFORMATION

This research received no specific grant from any funding agency in the public, commercial, or not‐for‐profit sectors.

CONFLICT OF INTEREST STATEMENT

None of the authors have any conflict of interest to disclose. We confirm that we have read the Journal's position on issues involved in ethics publication and affirm that this report is consistent with those guidelines.

ETHICS STATEMENT

The study was approved by the institutional data protection board (DEC206354) and the ethics committee of the French society of intensive care (CE SRLF 21‐38) in accordance with French regulations. Written informed consent was obtained from all participants, or from their legal representatives when patients were unable to provide consent themselves. The study adhered to the ethics principles outlined in the Declaration of Helsinki. The authors have full access to all data and have the right to publish any and all data, separate and apart from the guidance of any sponsor.

Supporting information

Table S1: Evolution of quantitative EEG features during sedation withdrawal (population level).

DATA AVAILABILITY STATEMENT

Anonymized data supporting the findings of this study are not publicly available but are available from the corresponding author upon reasonable requests.