Vagus Nerve Stimulation Elicits Sleep EEG Desynchronization and Network Changes in Responder Patients in Epilepsy

Abstract

Neural desynchronization was shown as a key mechanism of vagus nerve stimulation (VNS) action in epilepsy, and EEG synchronization measures are explored as possible response biomarkers. Since brain functional organization in sleep shows different synchrony and network properties compared to wakefulness, we aimed to explore the effects of acute VNS on EEG-derived measures in the two different states of vigilance. EEG epochs were retrospectively analyzed from twenty-four VNS-treated epileptic patients (11 responders, 13 non-responders) in calm wakefulness and stage N2 sleep. Weighted Phase Lag Index (wPLI) was computed as connectivity measure of synchronization, for VNS OFF and VNS ON conditions. Global efficiency (GE) was computed as a network measure of integration. Ratios OFF/ON were obtained as desynchronization/de-integration index. Values were compared between responders and non-responders, and between EEG states. ROC curve and area-under-the-curve (AUC) analysis was performed for response classification. In responders, stronger VNS-induced theta desynchronization (p < 0.05) and decreased GE (p < 0.05) were found in sleep, but not in wakefulness. Theta sleep wPLI Ratio OFF/ON yielded an AUC of 0.825, and 79% accuracy as a response biomarker if a cut-off value is set at 1.05. Considering all patients, the VNS-induced GE decrease was significantly more important in sleep compared to awake EEG state (p < 0.01). In conclusion, stronger sleep EEG desynchronization in theta band distinguishes responders to VNS therapy from non-responders. VNS-induced reduction of network integration occurs significantly more in sleep than in wakefulness.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13311-021-01124-4.

Introduction

Vagus nerve stimulation (VNS) has been effectively used for more than 30 years as adjunctive therapy in epileptic patients who are not suitable candidates or failed resective surgery. Responder rate (i.e., > 50% seizure frequency reduction) increases over time, particularly within 1 to 2 years of treatment, in relation with the presumed VNS long-term neuroplastic effects [1]. VNS also exerts an acute antiepileptic action, reflected in the beneficial effects of on-demand VNS activation, such as reduced seizure duration and severity [2, 3]. Neural desynchronization is one of the main presumed mechanisms of the antiepileptic effects of VNS, as shown by pioneering animal studies [4, 5], and later witnessed by human electroencephalography (EEG) connectivity studies [6, 7].

Functional connectivity and network theory have been increasingly used to characterize the brain functional abnormalities of epileptic patients [8–10]. Neural electricity in epilepsy is characterized by elevated levels of synchronization, sign of the tendency of the epileptic brain to hyper-connect in aberrant excitatory circuits [11–14]. Also, the network Global Efficiency (GE), which is a measure of the predisposition to integrate specialized information from distributed brain regions [15], was recently found as a hallmark of pathological epileptic network organization [16]. GE proved to distinguish healthy controls from epileptic patients, who showed a significantly elevated GE in resting-state EEG [17].

In this perspective, VNS has been proposed to act as a network therapy, which modulates brain synchrony towards a less epileptogenic state, through the so-called vagal afferent network [18–20]. Graph measures related to this brain network were successfully applied to define pre-operative predictors of VNS efficacy, such as network modularity and transitivity [21, 22]. However, thus far, no biomarkers of VNS response have been found using EEG synchronization or network measures, which might be able to distinguish the therapeutic effects occurring in responders compared to non-responders.

To the best of our knowledge, it also remains unresolved whether VNS-induced neuromodulation varies depending on the state of vigilance and, consequently, depending on the background condition of neural synchronization. Patterns of higher cortical synchronization characterize NREM sleep compared to wakefulness and to REM sleep [23–25]. A link exists between the higher levels of brain synchronization that can be found in NREM sleep and the tendency to generate interictal epileptiform discharges and seizures, particularly in N2 sleep [26, 27]. Indeed, common neural oscillation patterns have been described between epilepsy and NREM sleep, such as the thalamocortical relay circuits implicated in the generation of sleep spindles, which may be shared for spike-wave and ictal discharges, in both generalized and focal epilepsies [28–31].

On these grounds, we explored VNS-induced effects on EEG during sleep and wakefulness, hypothesizing that a stronger influence might be found in sleep, where closer resemblances with the hyper-synchrony of epileptic circuitry are seen. To this end, we analyzed, and compared between responders and non-responders, the acute VNS-induced changes on (i) EEG band-power, (ii) functional connectivity measures of EEG synchronization, and (iii) network GE.

Materials and Methods

Study Population

The Epilepsy Unit database of Saint-Luc University Hospital, Brussels, Belgium, was retrospectively searched for consecutive adult patients who (1) received a cervical VNS implant (LivaNova, London, UK) between 2008 and 2019; (2) after VNS implant, underwent a video-EEG monitoring (at least 24 h) with concomitant electrocardiogram (ECG); and (3) had active VNS system during the video-EEG monitoring. Patients were classified as responders (R) if showing ≥ 50% seizure frequency reduction as compared to pre-implant condition, and as non-responders (NR) in the case of < 50% reduction, based on clinical records assessment at the moment of the EEG recordings. Local Ethics Committee approval was granted (protocol 2018/07NOV/416). Clinical and medication characteristics of each patient were categorized for group-level analyses (see Supplementary Material).

EEG Recordings and Epoch Selection

(Video-)EEG recordings were performed either at Saint-Luc University Hospital or at William Lennox Centre, Ottignies, Belgium, with Deltamed® system (Natus Europe, Paris, France). Signals were digitized at a sampling rate of 256 Hz. The 19 scalp electrodes of 10–20 system placement were considered for analysis. Additional ECG (Eindhoven lead I-II) and EMG deltoid recordings were available. Due to the anatomical proximity with the pulse generator, ECG traces enable the detection of stimulation artifacts produced by VNS trains, indicating the VNS ON periods [3]. ECG derivations were visually screened to distinguish VNS ON and VNS OFF recording segments. If no clear distinction could be verified (e.g., due to poor quality of ECG recordings), the patient was excluded from analysis.

EEG processing was carried out in MATLAB R2018a (Mathworks, Natick, USA), using both in-house scripts and the Letswave 6 toolbox (UCLouvain, Brussels, Belgium). For each patient, EEG traces were reviewed ensuring that no seizures nor ongoing status epilepticus occurred in the hour preceding the recordings, which might — as themselves or due to postictal state disturbances — significantly alter brain synchronicity. A minimum of 8 to a maximum of 10 EEG epochs during VNS ON condition and 8 to 10 epochs during VNS OFF were collected, in two different states: (i) calm wakefulness, in absence of evident motor activity, for the awake EEG; (ii) stage 2 (N2) NREM sleep, for the sleep EEG. Sleep scoring and epoch selection were performed upon visual analysis, using the presence of sleep spindles and/or K-complexes as determinant for the N2 stage attribution. The epochs retrieved either contained or were located in immediate proximity to distinctive N2 figures. Nonetheless, it cannot be excluded that transitional N1 or N3 periods might have partially been included in the selected EEG traces. Each epoch had 10 s length. VNS ON epochs were searched at least 3 s after the VNS train onset and − 1 s before offset. VNS OFF epochs were selected at least 30 s after the VNS train offset. EEG segments containing deflections of > 500 µV were automatically rejected. Epochs with noticeable eye blinking or major interictal epileptiform discharges were removed by visual analysis, and those with the lowest number of residual artifacts were selected.

EEG pre-processing was completed by re-referencing on the common average of the 19 selected EEG channels. Band-pass filtering (Butterworth 4th-order) was applied for bandwidth-restricted analyses, in the delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and broadband frequency ranges (0.5–30 Hz). A general outline of the EEG analysis is displayed in Fig. 1.

Fig. 1.

General outline of the EEG analysis. Manual marking of the onset and offset of VNS trains was performed. For each patient, 8 to 10 EEG epochs during VNS ON condition and 8 to 10 during VNS OFF condition were collected, in calm wakefulness for the awake EEG, and in stage 2 NREM sleep for the sleep EEG. Each epoch had 10 s length. VNS ON epochs were searched at least 3 s after the VNS train onset and  1 s before offset. VNS OFF epochs were selected at least 30 s after the VNS train offset. EEG pre-processing was completed by re-referencing on the common average of the 19 selected EEG channels. Additional band-pass filtering was applied, for differential bandwidths computation, and following band power, connectivity and network analysis. wPLI, weighted Phase Lag Index. PDC, partial directed coherence

Band-Power Estimation: Hilbert Analysis

We first performed a Hilbert analysis [32, 33] to determine whether acute VNS induces EEG desynchronization (i.e., a decrease of band-power) with respect to ongoing oscillations [34]. Hilbert transform was applied to estimate band-power as a function of time in the different bandwidths. Relative band-power was defined as the resulting envelope amplitude of each epoch. Results were averaged across epochs for each condition (VNS ON and OFF) and state.

Connectivity Analysis: EEG Synchronization Measures

We focused on the connectivity measures quantifying EEG synchronization [35, 36]. Based on previous literature, showing that phase difference–based connectivity measures are sensitive for the acute VNS effects on EEG [6, 37], we addressed the measurement of weighted Phase Lag Index (wPLI). This measure was introduced by Vinck et al. as a development of the more classical Phase Lag Index (PLI) [38]. PLI estimates the asymmetry of the distribution of phase differences between two signals and is partially corrected for volume conduction (or leakage-corrected). wPLI proved to be more sensitive to detect true neural synchronization compared to PLI, due to the weighting of phase difference by the magnitude of the lag, which reduces the risk of bias introduced by small noise perturbations [39, 40]. It is computed through the following equation:

$$\mathrm{wPLI}=\frac{|E\left\{\left|\mathrm{\Delta },\phi \right|,s,g,n,\left(\mathrm{\Delta },\phi \right)\right\}|}{E\{\left|\mathrm{\Delta },\phi \right|\}}$$

where Δϕ is the phase difference between two given signals, E is a function computing the mean value, and sgn is a function that extracts the sign of a real number.

We investigated all possible pairwise combinations of EEG signals. The values obtained for each ON or OFF epoch were then averaged separately per condition, bandwidth, and state. As a desynchronization index, a wPLI Ratio OFF/ON was calculated for each channel pairing (> 1 = desynchronization, < 1 = synchronization), and for each patient. Connectivity matrices were built accordingly and a whole-brain degree of synchronization was estimated by averaging values across all channel pairings. In addition, we performed regional analyses by pooling specific subgroups of electrodes, and computed connectivity values restricted to those regions. Seven channel subgroups were established (Fig. 2).

Fig. 2.

Regional electrode pooling for connectivity analysis. Seven different regions were defined: three per hemisphere plus one central region

To assess wPLI performance compared to previous EEG synchronization measures, we also computed (i) spectrum-based measures, i.e., coherence and imaginary coherence [39, 41]; (ii) phase difference–based measures, i.e., PLI and phase locking value (PLV) [35, 42]. Similar values of Ratio OFF/ON were extracted for each additional connectivity measure. Further details on their computation can be found in Supplementary Material.

Network Analysis: Directed Connectivity and Global Efficiency

A sensor-space network analysis was performed aiming to describe the impact of VNS on Global Efficiency (GE), which may represent the (pathological) tendency of the network to combine information from distributed brain regions [10, 17].

First, we calculated a directed functional connectivity measure, namely partial directed coherence (PDC), to determine the edge values of a whole-brain directed connectome [17, 43–45]. PDC was computed on epochs, conditions, and states as described previously in the “Connectivity Analysis: EEG Synchronization Measures” section. Graph analysis was then performed: the PDC values were related across the different brain nodes (represented by each of the 19 selected EEG electrodes). The network feature of GE was subsequently extracted for each patient (see Supplementary Material for methodological details), and a GE Ratio OFF/ON was also obtained as de-integration index.

Statistical Analysis

Demographic and clinical characteristics of the study population were statistically compared between the R and NR groups, using the Mann–Whitney U test or Fisher’s exact test on scalar values or contingency tables, respectively.

Two types of statistical tests were carried out on EEG results: (i) paired Wilcoxon signed-rank test, between VNS OFF and ON conditions, to detect the EEG feature changes occurring at within-group level; (ii) Mann–Whitney U test, between the different Ratio OFF/ON values, to compare between-groups (R vs. NR) or between-states (awake vs. sleep EEG). In addition, receiver operating characteristic (ROC) curve analysis with area-under-the-curve (AUC) estimation were performed. Different thresholds of Ratio OFF/ON values were tested as classifiers for R vs. NR classification, and for each threshold value, sensitivity, specificity, and accuracy were determined.

Multiple testing in different bandwidths and for each EEG feature was controlled using the false discovery rate (FDR) correction for non-parametric testing, applied to all paired and non-paired tests [46]. The number of hypotheses tested in the FDR correction was 5, i.e., the number of bandwidths. All shown p-values were FDR-corrected, and significance threshold was maintained at p < 0.05.

The correlation of relative band-power with connectivity measures was further explored, as an indicator of sensitivity to volume conduction [47]. Correlation analysis was performed, at patient level, with VNS output current and therapy duration. Ratio OFF/ON values were used for correlation. Pearson’s or Spearman’s correlations were performed depending on the distribution of data.

Results

Study Population: Demographic and Clinical Characteristics

Twenty-six patients were eligible for inclusion after review of our clinical database. Two patients were excluded as the VNS ON/OFF conditions could not be identified correctly. Hence, 24 patients (16 F and 8 M; 11 R and 13 NR) were finally included in the study. A comparative table on demographics and clinical features of the R and NR groups can be found in Table 1. No statistical significance was detected when comparing the characteristics between the R and NR groups. Detailed information can also be found in Supplementary Material Table S1.

Table 1.

Demographic and clinical characteristics of the study population (n = 24). Classification as mixed epilepsy encompasses patients presenting both focal and generalized seizures, such as structural Lennox-Gastaut syndrome. R, responder. NR, non-responder. SD, standard deviation. AEDs, antiepileptic drugs. SSRI, serotonine-selective reuptake inhibitors. SNRI, serotonine-noradrenaline reuptake inhibitors. TCA, tricyclic antidepressants

Variable	R group (n = 11)	NR group (n = 13)	p-values
	Mean (SD)	Mean (SD)
Sex	4 M; 7 F	4 M; 9 F	1
Age	34.1 (13.9)	33.2 (14.6)	0.741
Epilepsy onset age	9.8 (9.6)	6.5 (3.2)	0.649
Duration VNS therapy (in months)	67.2 (61.5)	68.2 (56.6)	0.82
VNS output current (mA)	1.48 (0.51)	1.67 (0.37)	0.424
VNS frequency (Hz)	25 (5)	26.9 (3.8)	0.51
VNS pulse width (µs)	340.9 (126.1)	307.7 (109.6)	0.608
VNS duty cycle (time ON/time OFF %)	18.1% (10.9%)	18.4% (6.7%)	0.424
Epilepsy type:			0.357
Focal epilepsy	7	11
Mixed/generalized epilepsy	4	2
QI range:			1
Normal-mild impairment (IQ > 55)	9	10
Moderate-severe impairment (IQ < 55)	2	3
mean n of AEDs used	2.6 (0.8)	3.2 (0.7)	0.119
n using benzodiazepines	2/11	5/13	0.686
n using antidepressants (SSRI. SNRI. TCA)	5/11	5/13	1

Hilbert Analysis: Band-Power Calculation

No significant differences were detected between the R and NR groups, and no significant differences across OFF vs. ON condition were detected to occur within any groups or states. Detailed results are available in Supplementary Material Table S2.

Connectivity Analysis: EEG Synchronization Measures

wPLI Analysis

Considering the sleep EEG whole-brain synchronization (i.e., average wPLI Ratio OFF/ON), we found a greater desynchronization occurring in the R group, compared to the NR group, in the theta band (mean difference R-NR = +12.8%, p = 0.038; Figs. 3 and 4). Other bandwidths (particularly alpha and broadband) also showed higher mean and median sleep desynchronization occurring in R than NR, but these differences did not reach significance (Fig. 4 and Supplementary Material Table S3). In awake EEG, no significant differences were found between the R and NR groups.

Fig. 3.

Connectivity matrices for wPLI Ratio OFF/ON in sleep EEG, shown at a group-level averaged values. In the right-column matrices, the difference between R and NR groups is shown for each channel pairing: red colors indicate greater desynchronization occurring in the R group, blue colors in the NR group. In the theta band, the degree of whole-brain desynchronization is significantly different between the two groups. A greater theta wPLI Ratio OFF/ON in R group (i.e., red colors in the right graph) can be seen in the vast majority of channel pairings

Fig. 4.

Boxplots showing the between-groups comparison in both sleep and awake EEG for wPLI whole-brain analysis. The grey horizontal line corresponds to the no-effect level (Ratio OFF/ON = 1). Significant higher desynchronization is found in the theta band in sleep EEG, in R group (n = 11) compared to NR group (n = 13), with p = 0.038, FDR-corrected Mann–Whitney U test. Trending higher sleep desynchronization, although non-significant, was also found in alpha and broadband ranges. Despite a general desynchronization is also seen in wakefulness (as shown by median values > 1), R vs. NR differences in the awake EEG are non-significant

We performed ROC curve analysis on the theta wPLI sleep values, for R vs. NR classification. AUC value for theta sleep wPLI Ratio OFF/ON was 0.825 (CI 0.62–0.96). The best classifier threshold was identified at 1.05: with a > 1.05 response cut-off, sensitivity was found to be 73% (CI 39–94%), specificity = 85% (CI 66–98%), and overall accuracy 79% = (CI 58–93%) (Fig. 5, Table 2).

Fig. 5.

ROC curve analysis of theta wPLI Ratio OFF/ON in sleep, for responders (R) vs. non-responders (NR) classification. Different thresholds of the sleep theta wPLI Ratio OFF/ON are tested as classifiers for clinical response. The best performing threshold was found to be 1.05, i.e., the furthest point of the ROC curve from the random allocation of depicted by the diagonal red line. This means that, with a cut-off set at > 1.05 to be classified as responder, the wPLI Ratio OFF/ON has a 73% sensitivity, 85% specificity, and 79% overall accuracy as response biomarker. AUC, area-under-the-curve

Table 2.

ROC curve analysis for responders (R) vs. non-responders (NR) classification, which tests different threshold values of theta sleep wPLI Ratio OFF/ON (having as condition that R are > than threshold). The best performing cut-off was found at 1.05 (which equals to 5% acute desynchronization)

Theta sleep wPLI Ratio OFF/ON
Threshold		Sensitivity	Specificity	Overall accuracy
> 1		91% (59–100%)	54% (25–81%)	71% (49–87%)
> 1.02		82% (48–98%)	62% (32–86%)	71% (49–87%)
> 1.04		73% (39–94%)	69% (39–91%)	71% (49–87%)
> 1.05		73% (39–94%)	85% (55–98%)	79% (58–93%)
> 1.08		54% (23–83%)	85% (55–98%)	71% (49–87%)
> 1.1		45% (17–77%)	92% (64–100%)	71% (49–87%)

Considering the regional subanalysis, we found significantly higher values of desynchronization in R compared to the NR group in the sleep EEG theta band in the left frontal (mean difference R-NR = +25%, p = 0.037) and central region (mean difference R-NR = +17.3%, p = 0.042), as shown in Fig. 6.

Fig. 6.

Boxplots showing the between-groups comparison in the theta band in sleep EEG, for wPLI regional analysis. Significant higher desynchronization was found in the left frontal and central region, in R group compared to NR group (p < 0.05, Mann–Whitney U test)

Correlation Analysis: Band-Power and Clinical Characteristics

In sleep EEG, theta, alpha, and beta bands showed the lowest coefficients, indicating that results in those bands might be less prone to be influenced by volume conduction issues (Fig. 7). Particularly, in the theta band, only PLV and COH (both leakage-uncorrected measures) showed significant correlations with band-power. For the connectivity measures, both wPLI and iCOH show lower correlation coefficients with band-power. We refer to Supplementary Material Fig. S1 for detailed results of other connectivity measures.

Fig. 7.

Correlation analysis between relative band-power and connectivity measures in sleep EEG. Theta, alpha, and beta bands show lower coefficients, indicating that measures in those bands are less prone to be influenced by volume conduction issues. Likewise, wPLI shows, as well as iCOH, lower correlation coefficients compared to other measures that are non-corrected for volume conduction (such as PLV and COH). Delta and broadband measurements appear reliable only for iCOH. Coefficients are expressed in ρ values. Only significant (p > 0.05) values are reported

No significant correlations were found between wPLI Ratio OFF/ON values and VNS output current or VNS therapy duration, performed at single-patient level. Particularly, no significant association was found by linear regression for theta wPLI, neither in sleep (Fig. 8) nor awake EEG.

Fig. 8.

Correlation analysis between theta sleep wPLI and GE values with VNS output current and therapy duration, and linear regression fit. Each point in the scatter plot corresponds to values of a single patient (n = 24). No significant associations (i.e., low R₂ values, p-value was > 0.05 for all correlation analyses) were found between higher stimulation intensities or longer durations of therapy and greater indexes of desynchronization or de-integration

Network Analysis: Directed Connectivity and Global Efficiency

At a within-group level, in sleep EEG, as described in Fig. 9, in the R group were detected significant lower values of GE in the VNS ON condition were detected in the R group, in all bandwidths (p = 0.042) except for the delta band (p = 0.1). No significant changes were detected in the NR group in sleep EEG. By contrast, for awake EEG, no differences could be found in any of the two groups. The between-group (R vs. NR) analysis did not yield significant differences, by directly comparing the GE Ratio OFF/ON values; hence, GE was not found to perform accurately as a response classifier.

Fig. 9.

Paired boxplots for within-group Wilcoxon signed-rank test on Global Efficiency values, as computed from PDC values. In each bandwidth, for both sleep and awake EEG, VNS OFF and VNS ON paired results are shown linked for each patient. Significant VNS ON decrease of Global Efficiency was found in sleep EEG, in theta, alpha, beta, and broadband ranges, to occur only in responder (R) group but not in non-responders (NR)

For the between-states analysis, after pooling all patients (R and NR) together, the values of GE Ratio OFF/ON were significantly higher in sleep compared to awake state, in all bandwidths (p < 0.01 in delta, theta, and beta bands, p < 0.05 in alpha and broadband, Fig. 10, Table 3). No significant correlation was detected between GE and either VNS output current or therapy duration (Fig. 8).

Fig. 10.

Between-state comparison — awake (AW) vs. sleep (SL) — for global efficiency (GE) values, in the five analyzed bandwidths, after pooling all patients of the study population (n = 24). Consistently higher values of GE Ratio OFF/ON (i.e., decrease of network integration during VNS ON) can be seen in sleep, throughout all frequency bands

Table 3.

Between-state analysis for global efficiency (GE) results. Higher de-integration is found in the sleep state compared to awake state, in all bandwidths. Mean values and standard deviation are reported after averaging all patients for the values of one state. Between-state difference (mean SL values minus mean AW values) and FDR-corrected p-values for Mann–Whitney U test. SL, sleep. AW, awake. SD, standard deviation. FDR, false discovery rate

GE Ratio OFF/ON, all patients (n = 24)
(Band)	SL mean	(SD)	AW mean	(SD)	Diff. SL-AW%	p-value (FDR)
Delta	1.203	(0.198)	1.041	(0.111)	+ 16.2%	0.008**
Theta	1.260	(0.240)	1.062	(0.148)	+ 19.7%	0.005**
Alpha	1.229	(0.225)	1.074	(0.144)	+ 15.5%	0.014*
Beta	1.291	(0.264)	1.056	(0.141)	+ 23.5%	0.007**
Broadband	1.290	(0.250)	1.094	(0.167)	+ 19.6%	0.015*

Discussion

Despite three decades of use in epilepsy, a reliable VNS response biomarker has not entered clinical practice yet. Such biomarkers could indicate that treatment is being biologically active, and could help uncover the electrophysiological mechanisms related with VNS effectiveness. Furthermore, to the best of our knowledge, no study had addressed so far how the state of vigilance might influence VNS cortical modulation. With the present study, we provide evidence that (i) sleep EEG desynchronization in the theta band distinguishes responders from non-responders; (ii) wPLI Ratio OFF/ON values may be used successfully as a classifier for VNS response; (iii) responders show a significant VNS-induced decrease of sleep network integration; and (iv) acute VNS induces a significantly stronger network reorganization in sleep compared to wakefulness.

EEG Synchronization Measures

We proved that acute VNS-induced sleep desynchronization in the theta band can distinguish responders from non-responders. Our results suggest that acute sleep neural desynchronization might reflect, at least partly, the mechanism of action of the antiepileptic effect of VNS. In this context, we propose the sleep wPLI desynchronization index for theta band (i.e., wPLI Ratio OFF/ON) as a possible VNS response biomarker. This marker could correctly classify responders and non-responders with considerably high AUC values (0.825): with a cut-off set at > 1.05, accuracy values reached the best performance (sensitivity = 73%, specificity = 85%, accuracy 79%). These accuracy values stand close to those described by de Taeye et al. for P3b amplitude (values > 1 µV, sensitivity = 70%, specificity = 90%), so far the sole acute VNS response biomarker previously described in the literature, in a study on 19 epileptic patients [48].

Our results on connectivity EEG synchronization measures are in line with previous literature, which proved the usefulness of phase difference–based measures to detect VNS impact on brain connectivity. In a resting-state wakefulness study, Bodin et al. showed that lower interictal PLI values can distinguish responders from non-responders to VNS, suggesting that cumulative desynchronization could occur in R and not in NR group [37]. However, in relation with acute stimulation, no differences were found between the two groups. More recently, Sangare et al., in another study on awake EEG, proved that significant PLI reduction occurs in responders, and that PLI Ratio OFF/ON (i.e., acute desynchronization) correlates with seizure frequency reduction [6]. Nonetheless, no classifiers could be defined as a potential response biomarker based on EEG synchronization measures so far. Indeed, in the study of Sangare et al., only within-group differences could be demonstrated in wakefulness. We presume that R vs. NR differences were not pronounced enough in awake EEG to define a response biomarker based on between-group analysis, as awake EEG desynchronization might reflect less closely the therapeutic effect of VNS compared to sleep.

wPLI has been recently applied in EEG synchronization research as potentially more sensitive to detect true connectivity changes than older measures [49, 50]. We could demonstrate that wPLI was the most sensitive index in detecting true acute VNS-induced desynchronization. Indeed, lower correlation values with relative band-power suggest that wPLI shows less proneness to volume conduction than PLI or PLV (previously used in [6, 37]), and further supports the genuineness of the wPLI-based connectivity. Nonetheless, although non-significantly, other leakage-corrected measures (PLI, iCOH) showed similar trends of greater desynchronization occurring in responders compared to non-responders, which adds robustness to our findings (see Supplementary Material Table S3). Additionally, wPLI values in theta, alpha, and beta bands did not show significant band-power correlation (Fig. 7), supporting the interpretation that results in these bandwidths might present higher reliability [47].

Topographic Analysis

Our wPLI analysis proved that greater levels of theta desynchronization occur in responders over central and left frontal areas, compared to non-responders. We interpret them as possibly related to a stronger VNS action in responders on key regions of the vagal afferent network.

Previous EEG studies on low-frequency (10 Hz) acute VNS in Crohn’s disease patients showed stronger effects in the theta band in the fronto-temporal left regions, supporting the hypothesis of a prevalent ipsilateral acute cortical activation [51]. The stronger effects over the frontal region observed in responders in our study could mirror a stronger modulation in prefrontal, anterior cingular, and anterior insular cortices, for which a role in VNS responsiveness in epilepsy was already proposed by fMRI, EEG, and MEG studies [22, 52, 53]. These regions might be directly modulated via the thalamus, as indirectly supported by previous SPECT studies which showed prominent acute left thalamic activation [54].

However, it is renowned that VNS exerts its therapeutic action regardless of epilepsy lateralization, and that despite a left-sided implant, it elicits bilateral brain effects due to bilateral projections from the nucleus tractus solitarius [55]. In this context, higher desynchronization in the central region might correspond to stronger bilateral VNS effects on somatosensory primary areas, also recently identified as important targets of the afferent VNS-induced modulation [56].

Prudence is recommended in considering the mechanistic value of our topographic findings: while it cannot be disputed that VNS exerts a widespread action, the present results suggest that left frontal and central desynchronizations show potential to be investigated as more promising markers of efficacy compared to whole-brain analyses. Strengths of our topographic analysis were the fact of being carried out on a regional channel pooling rather than a channel-by-channel comparison (which would have increased drastically the number of multiple comparisons) and that survived a rigorous FDR-correction. Due to the limited size of our cohort, we could not perform a stratified subgroup analysis based on the presumed localization of the epileptogenic focus. Further studies, addressing specific groups of focal epilepsy, should investigate whether greater desynchronization over areas corresponding to the epileptogenic focus might correlate with clinical response.

Network Analysis

We proved for the first time a significant EEG network reorganization induced by acute VNS. Previous studies attempted to identify network changes in awake EEG, by means of minimum spanning tree (MST) computation, but failed to prove significant differences between responders and non-responders [6, 57]. In particular, we demonstrated that network efficiency (i.e., GE), whose elevated values were previously associated with epileptic brain dysfunction [17], was significantly decreased in sleep by acute VNS in responders but not in non-responders. This suggests that a decrease of network integration might reflect a modulation of epileptogenic networks towards a less pathologic organization, and that this phenomenon, occurring selectively in sleep, might be associated with the therapeutic effect of VNS.

Moreover, a tendency of VNS to modulate brain networks in sleep is supported by the between-state analysis, after pooling all patients regardless of clinical response: a consistently greater decrease of GE was detected in sleep when directly compared with wakefulness (Fig. 10). It should be noted that a trending decrease of GE in sleep, although non-significant (p = 0.1), was also detected among non-responders. We may hypothesize that the decrease of integration could correspond to a network signature of the modulating VNS effect on the brain, and that the higher this decrease, the higher the impact of VNS at an antiepileptic level. The duration of VNS therapy did not appear to influence the acute VNS-induced GE changes (Fig. 8). However, further exploration is justified to explore potential long-term resting-state effects of VNS on network integration, as compared to baseline measurements.

VNS Action on Epileptic and Sleep Oscillatory Patterns

Despite several works addressed sleep architecture and breathing disorders in the light of VNS side-effects (see [58, 59] for more extensive reviews), we report for the first time an EEG connectivity and network analysis performed in sleep upon acute VNS. Only one small study had addressed VNS effects on EEG power distribution in sleep: Rizzo et al. found chronic EEG power increase (comparing pre-to post-implantation), but reported no significant VNS-induced acute changes. These negative results were confirmed by our Hilbert analysis findings [60].

In light of our desynchronization and network analysis results, a putative link emerges between VNS neuromodulation in sleep — and subsequently on the cortical and subcortical generators of its oscillatory patterns — and its antiepileptic action. Proving a state-dependency of VNS effects might pave the way to developing new circadian stimulation strategies.

Different explanations may be given for the observed results. Neural networks typical of epilepsy, such as the thalamo-cortical circuits studied in generalized seizures and spike-wave discharge origination [29], were shown to mirror neural circuits of synchronization that are typical of sleep (e.g., those generating spindles in N2 sleep) [61, 62]. It may be proposed that the more pronounced VNS-induced neuromodulation in sleep relates to a peculiar effect on thalamo-cortical connections in this state, as sleep oscillatory patterns tend to a more pro-epileptic organization. Indirect evidence to support this hypothesis is given by higher rates of clinical efficacy of VNS in generalized and spike-wave epilepsies [1, 63] and VNS-induced thalamic and cortical activation, as seen through PET and SPECT studies [54, 64].

Physiological variations of vagal function might also contribute to enhanced VNS effects during sleep. It is renowned that higher vagal tone characterizes NREM sleep (especially stages N2 and N3), as shown by decreased heart rate and blood pressure, as well as increased heart rate variability (HRV), an indirect marker of parasympathetic activity [65, 66]. An elevated tonic neural activity in vagal fibers might result in stronger afferent effects in sleep as compared with wakefulness, and thus in more prominent neuromodulation during this state. Further studies are needed to understand the relationship between peripheral biomarkers of vagal tone (such as HRV measures [67]), efferent VNS-induced activation (such as laryngeal evoked potentials [68]) and the observed afferent effects on brain connectivity and networks.

Theta Band Effects: Functional and Therapeutic Implications

We successfully identified an EEG-based classifier for VNS response in the theta bandwidth. Several factors support the interpretation of theta activity modulation as a hallmark of the therapeutic action of VNS on brain electricity.

In past literature, theta band activity was consistently pointed out as the brain oscillation range where important VNS-induced modulation occurs: Marrosu et al. found in an early work on EEG coherence a long-term desynchronization induced after 1 year of therapy in the theta band only [69]. Sangare et al. performed a PLI and PLV brain synchrony analysis, and theta activity was the sole bandwidth where acute desynchronization was reliably found regardless of the used metric [6]. Ernst et al. recently demonstrated, through an electrocorticography (ECoG) study, that significant acute power spectrum reduction (i.e., single-channel desynchronization) is elicited by VNS only in the theta band [7].

Interestingly, theta band modulation might be a responsiveness feature shared with other neurostimulation approaches, indicating its possible role as a marker of effective epileptic network disruption [70]. A recent deep brain stimulation (DBS) study from Scherer et al. proved that acute theta desynchronization induced by anterior thalamic DBS in the temporal lobe is a physiological hallmark of response [71].

In support of this hypothesis, previous evidence in the literature highlighted the prominent role of hypersynchronous theta oscillations in the pathophysiology of generalized epilepsy, particularly of oscillations reflecting thalamo-cortical circuits [72]. Another study also demonstrated the important disruption of theta connectivity as a hallmark of focal epilepsy networks [73].

Considering the pattern of brain oscillations in sleep, where theta waves are dominant in stages N1 and N2 [74], a stronger theta modulation might signify that responders to VNS undergo a prominent modulation of the dominant cortical waves in the given analyzed stage. Sleep spindles, which were included — although non-systematically — in the analyzed epochs, stretch across the upper theta/alpha band (7–14 Hz) [75, 76], and it could be hypothesized that the observed effect reflects at least partly an impact on the thalamocortical spindle-generating networks, which are more closely associated with circuits of epilepsy pathophysiology. Unsurprisingly, the stronger sleep desynchronizing effect seen besides the theta (12.8%) was found in the alpha band (8%) (Fig. 3, Supplementary Material Table S3). However, further studies could determine whether in a stage where other oscillations express the main thalamocortical synchronicity, such as the delta band in slow-wave N3 sleep, a different bandwidth selectivity would be found or not.

In conclusion, our results further suggest the modulation of pathological theta rhythms might represent a key feature of the therapeutic action of neurostimulation in epilepsy, and more specifically of VNS, and adds consistency to the search for VNS electrophysiological biomarkers in this bandwidth. This study casts new light on how VNS exerts its effects during sleep, where this modulation might mirror the effect on the thalamocortical circuits shared by sleep and epileptic hypersynchrony phenomena.

Study Limitations and Future Perspectives

Our study was conducted on a relatively small and heterogeneous study population, especially with respect to the duration of VNS therapy (min. 11, max. 220 months), which could not be standardized due to the retrospective nature of the study. Stratification for duration of VNS was not possible due to the limited number of participants. Nonetheless, we could demonstrate that no significant correlation existed between EEG changes and VNS therapy duration or output current (Fig. 8), meaning that these variables, although uncontrolled, do not affect the entity of acute VNS-induced neuromodulation at an inter-subject level, whereas they still might have a role in chronic effects. Future prospective studies might be needed to evaluate the evolution of acute VNS afferent effects at serial time points after therapy start. Furthermore, in a future experimental setting, it would be of interest to analyze the dose-dependency of the observed EEG desynchronization. An intra-subject comparison of the induced modifications, at increasing VNS intensities, might help to define threshold and saturation levels of stimulation for obtaining the therapeutic afferent effects.

The analyzed EEG data did not include the whole range of sleep stages, but was only restricted to stage N2 sleep, which is the most abundant (40–50% of total sleep time) sleep stage [77]. A shortcoming of the present study is therefore the lack of comparison between VNS-induced desynchronization and network reorganization across different NREM or REM sleep stages. Furthermore, it was not possible to entirely tease out a contamination of the selected epochs by epileptiform activity: in the absence of concomitant intracranial recordings, we cannot exclude that deep epileptic discharges, despite not being seen at scalp, might have altered background EEG rhythms and synchronization.

Conclusion

We demonstrated that VNS-induced EEG desynchronization in sleep distinguishes responders from non-responders to VNS for epilepsy. Theta band wPLI showed 79% accuracy as a response biomarker, with a high performance as a classifier (AUC = 0.825). Theta desynchronization emerges as a candidate hallmark for the therapeutic effect of VNS. Moreover, responders showed a VNS-induced decrease of EEG network efficiency, which might reflect the VNS action at a network scale. Greater decrease in GE was observed in sleep compared to awake EEG, where neuronal organization reflects more closely the epileptic organization. These insights might open new avenues to stimulation strategies taking into account the sleep/wake cycle.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contribution

Simone Vespa: conceptualization, methodology, formal analysis, investigation, data interpretation, writing — original draft. Jolan Heyse: methodology, software, formal analysis, writing — review and editing. Lars Stumpp: methodology, software, writing — review and editing. Giulia Liberati: data interpretation, writing — review and editing. Susana Ferrao Santos: data interpretation, writing — review and editing. Herbert Rooijakkers: data interpretation, writing — review and editing. Antoine Nonclercq: conceptualization, data interpretation, writing — review and editing. André Mouraux: methodology, data interpretation, writing — review and editing. Pieter van Mierlo: methodology, data interpretation, writing — review and editing, supervision. Riëm El Tahry: conceptualization, investigation, data interpretation, writing — review and editing, supervision.

Funding

SV is supported by a grant from F.R.S-FNRS (Fonds National de la Recherche Scientifique) as a Research Fellow “Aspirant.” RET is supported by a “Fonds de Recherche Clinique” grant from Saint-Luc University Hospital. No other specific funding has supported the present publication.

Declarations

Conflict of Interest

PvM is co-founder and stakeholder of Epilog NV. The authors declare no other conflict of interest.