Transcutaneous auricular vagus nerve stimulation in drug‐resistant epilepsy: A randomized, sham‐controlled crossover trial with exploratory EEG and MRI response phenotyping

Abstract

Objective

Transcutaneous auricular vagus nerve stimulation (taVNS) is a noninvasive neuromodulation strategy for drug‐resistant epilepsy, but clinical response is heterogeneous. We examined whether electroencephalographic (EEG) and structural magnetic resonance imaging (MRI) phenotypes were associated with active‐specific clinical benefit from taVNS.

Methods

In this single‐center, randomized, double‐blind, sham‐controlled crossover exploratory pilot trial, 14 patients with drug‐resistant focal epilepsy received 8 weeks of active taVNS and 8 weeks of sham stimulation, separated by a 4‐week washout, in randomized sequence. The primary clinical outcome was seizure reduction ratio (SRR) during active versus sham stimulation. Active‐specific clinical benefit was defined as ΔSRR = SRR_active − SRR_sham. Resting‐state EEG was obtained at baseline and at the end of the protocol. Changes in regional relative band power were correlated with ΔSRR, with a focus on the β band based on prior evidence of vagal stimulation‐related desynchronization. Baseline structural MRI findings were explored by active‐phase responder status.

Results

Active taVNS did not significantly outperform sham stimulation. Mean SRR was 46.26 ± 40.96% during active stimulation and 48.62 ± 37.27% during sham stimulation, with a mean within‐subject difference of −2.37 ± 16.64% (p = 0.604). Greater β‐band attenuation was associated with larger active‐specific clinical benefit in the occipital (ρ = −0.723, p = 0.003, q = 0.017) and frontal regions (ρ = −0.657, p = 0.011, q = 0.027). Parietal β‐band attenuation showed a nominal association (ρ = −0.552, p = 0.041, q = 0.068). Bilateral structural abnormalities and limbic involvement were more frequent among nonresponders (both p = 0.021, uncorrected).

Significance

Although primary clinical efficacy was not demonstrated, EEG β‐band modulation and limbic structural network burden may help explain interindividual variability in taVNS benefit. The substantial sham response underscores the need for rigorous sham‐controlled taVNS trials.

Plain Language Summary

About one‐third of people with epilepsy keep having seizures despite medication. taVNS is a noninvasive treatment that stimulates a nerve at the outer ear. In this small, carefully blinded trial, active taVNS did not reduce seizures more than an inactive (sham) treatment, and the sham response was large. However, people whose EEG β brain waves decreased during the study tended to benefit more, and those with more widespread limbic or structural brain changes tended to benefit less. If confirmed in larger studies, these EEG and MRI markers might help identify who is most likely to respond to taVNS.

Key points.

• Active taVNS did not significantly outperform sham stimulation for seizure reduction in adults with drug‐resistant epilepsy.

• Greater β‐band EEG attenuation was associated with greater active‐specific benefit in occipital and frontal regions (FDR‐corrected).

• A substantial sham (placebo) response was observed, underscoring the need for rigorous sham‐controlled taVNS trials.

• Limbic and bilateral structural MRI involvement were more frequent in nonresponders (p = 0.021, uncorrected, exploratory).

• Combined EEG‐MRI phenotyping may help explain interindividual variability in taVNS response and inform future trial design.

1. INTRODUCTION

Epilepsy is a heterogeneous neurological disorder characterized by recurrent seizures arising from abnormal and hypersynchronous neuronal networks. ¹ , ² Despite advances in antiseizure medications (ASMs), approximately one‐third of patients continue to experience seizures despite adequate trials of two or more ASMs and are classified as having drug‐resistant epilepsy (DRE). ³ , ⁴ For many of these patients, curative resective surgery is not feasible because of multifocal seizure onset, involvement of eloquent cortex, poorly localized epileptogenic zones, or the absence of a resectable structural lesion. Consequently, neuromodulation has become an important palliative treatment strategy for DRE. ⁵

Implanted cervical vagus nerve stimulation (VNS) has demonstrated clinically meaningful seizure reduction in a subset of patients with DRE. ⁵ Its antiseizure effects are thought to involve afferent vagal projections to the nucleus tractus solitarius, locus coeruleus–norepinephrine system, and distributed thalamocortical and limbic networks involved in seizure modulation. ⁶ , ⁷ , ⁸ Transcutaneous auricular vagus nerve stimulation (taVNS) has been developed as a noninvasive approach to engage similar central vagal pathways through stimulation of the auricular branch of the vagus nerve. Because taVNS avoids surgical implantation, it may offer a more accessible and lower‐risk neuromodulation strategy for selected patients with epilepsy. ⁹ , ¹⁰ Although taVNS is presumed to engage vagal afferent pathways broadly similar to those recruited by implanted cervical VNS, the two modalities differ in stimulation route, delivered dose, and continuity of stimulation, and the degree to which their central and network‐level effects overlap remains incompletely characterized.

Clinical studies of taVNS in epilepsy have reported mixed results. An early proof‐of‐concept trial published in Epilepsia suggested that transcutaneous vagus nerve stimulation may reduce seizure frequency in a subset of patients with pharmacoresistant epilepsy, supporting the potential feasibility of this noninvasive neuromodulation approach. ¹¹ Subsequent pilot and controlled studies also suggested potential seizure reduction, whereas larger randomized trials have not consistently demonstrated superiority over sham stimulation. ⁹ , ¹⁰ , ¹¹ , ¹² , ¹³ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ These inconsistent findings may reflect heterogeneity in stimulation parameters, stimulation targets, control conditions, adherence, epilepsy phenotype, and outcome definitions. In addition, the design of an adequate sham condition remains challenging because electrical stimulation can produce perceptible somatosensory sensations. These limitations highlight the need for objective biomarkers that can clarify target engagement, identify patients more likely to benefit, and explain interindividual variability in treatment response.

Electroencephalography (EEG) provides a direct and temporally sensitive measure of cortical network activity and may be particularly useful for evaluating neuromodulation‐related changes in epilepsy. ¹⁸ , ¹⁹ Prior studies of VNS have suggested that effective stimulation may be associated with reduced pathological synchronization or band‐specific desynchronization, including changes in β‐band activity. ¹⁸ , ²⁰ , ²¹ , ²² , ²³ , ²⁴ In parallel, structural magnetic resonance imaging (MRI) may provide complementary information regarding the integrity of limbic and cortical networks required for propagation of vagal neuromodulatory effects. ⁵ , ⁷ , ²⁵ Thus, combined EEG and MRI phenotyping may help characterize biological differences between patients who do and do not benefit from taVNS.

In this exploratory randomized, double‐blind, crossover pilot trial, we evaluated the clinical effect of active versus sham taVNS in patients with DRE and examined whether EEG and structural MRI phenotypes were associated with active‐specific clinical benefit. We focused on changes in regional EEG relative band power from baseline to the end of the protocol and explored baseline structural MRI features according to active‐phase responder status. We hypothesized that EEG signatures of reduced pathological synchronization and preserved structural network integrity would be associated with greater clinical benefit from taVNS.

2. MATERIALS AND METHODS

2.1. Patients

This study enrolled patients with focal drug‐resistant epilepsy, defined as seizures refractory to two or more antiseizure medications (ASMs). All patients were Korean‐speaking adults, and each had a caregiver available to assist with seizure documentation when needed. Patients were excluded if they had severe cardiovascular, psychiatric, or infectious disease, malignancy, or a major neurological disorder other than epilepsy. In addition, patients taking antidepressant medications other than ASMs were excluded. Only clinically documented epileptic seizures were counted, and patients with psychogenic nonepileptic seizures were excluded. Antiseizure medications were maintained without change throughout the baseline period and both the active and sham stimulation phases. A history of status epilepticus was not in itself an exclusion criterion; patients judged clinically unstable, including those with recent status epilepticus, could be excluded under the general exclusion criteria at the investigators' discretion, and no episode of status epilepticus occurred during the study period. The study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board of Samsung Medical Center (IRB No. SMC 2023‐11‐121). The clinical investigation of the medical device was also approved by the Ministry of Food and Drug Safety of Korea (MFDS approval No. 1710; November 9, 2023). Patient recruitment was conducted from April 22, 2024 to December 11, 2025.

2.2. Study design

This study was conducted as a single‐center, prospective, randomized, double‐blind, crossover exploratory pilot trial. The trial was designed as an exploratory pilot study and was not statistically powered to confirm the clinical efficacy of taVNS; the group‐level efficacy comparison should therefore be interpreted accordingly. Patients were randomized 1:1 to Group A or Group B using block randomization generated by an independent statistician. The trial comprised two 8‐week stimulation phases separated by a 4‐week washout (latent) period, aligned with the visit schedule: Visit 2 (baseline; Day 0), Visit 3 (Week 4), and Visit 4 (Week 8) occurred during the first stimulation phase; Visit 5 (Week 12) marked the end of the washout period; and Visit 6 (Week 16) and Visit 7 (Week 20) occurred during the second stimulation phase. Group A received active stimulation during the first 8‐week phase (Visits 2–4) and sham stimulation during the second 8‐week phase (Visits 5–7), whereas Group B received sham stimulation during the first 8‐week phase (Visits 2–4) and active stimulation during the second 8‐week phase (Visits 5–7) (Figure 1). During each stimulation phase, patients used the assigned device for ≥5 days/week, totaling 120 min/day (±30 min), with individual sessions lasting 30–60 min and divided into 2–5 sessions/day as needed.

FIGURE 1.

Patient flow and randomized crossover trial design. A total of 21 patients were assessed for eligibility, including 10 males and 11 females. Seven patients were excluded before randomization, and 14 patients were randomized to one of two treatment sequences. Group A (n = 8) received active taVNS during the first 8‐week period followed by sham stimulation during the second treatment period, with a 4‐week washout period between phases. Group B (n = 6) received sham stimulation first and active taVNS second, with the same washout interval.

Active stimulation was delivered as transcutaneous pulsed electrical stimulation at an auricular site corresponding to vagal innervation (e.g., the cymba conchae region) with a stimulation frequency of 25 Hz (allears, TODOC, Seoul, Republic of Korea). The stimulation waveform was a biphasic pulse delivered in fixed‐output mode and applied continuously during each session, with a per‐pulse width of 200–1000 μs across the 23 output levels. Output intensity (Levels 1–23; approximately 0.8–2.0 mA) was individually titrated at the initiation of each stimulation phase. Specifically, the intensity was gradually increased until the patient reported a clearly perceivable sensation at the stimulation site (sensory detection threshold), and the device was then set one level below this threshold so that stimulation was not consciously perceived during routine use. The selected output level was maintained throughout each assigned phase.

Sham stimulation followed the same schedule and device operation as active stimulation; however, no transcutaneous electrical current was delivered. Instead, the electrical stimulation settings were converted to an equivalent sound pressure level (SPL) and presented as an auditory output at the same nominal device level as the active condition, thereby mimicking stimulation while preserving blinding. Because transcutaneous electrical stimulation produces perceptible somatosensory sensations, no sham condition can fully reproduce the experience of active stimulation. ⁹ , ²⁶ This auditory‐mimicry approach follows the sensory‐mimicry principle established for sham conditions in noninvasive brain stimulation ²⁷ and is conceptually similar to recent auricular vagus nerve stimulation studies in which sham devices emitting the same sound as the active device achieved effective participant blinding. ²⁸ , ²⁹ Group allocation and device provisioning were carried out by an independent device administrator who was not involved in clinical care or outcome assessment, so that patients, treating physicians, and outcome assessors remained blinded to allocation; in line with consensus reporting recommendations for transcutaneous VNS trials, ³⁰ blinding was additionally evaluated during the trial.

Patient adherence to taVNS use was monitored using a dedicated smartphone application that controlled the device and automatically logged stimulation usage time. The application was programmed to track whether patients met the prespecified daily dose target of 120 min/day (±30 min) throughout the trial period, enabling automated assessment of adherence.

2.3. Electroencephalogram (EEG)

Resting‐state scalp EEG was obtained at two predefined time points according to the study schedule: pre at baseline (Visit 2, Day 0) prior to initiation of the taVNS clinical protocol and post at the end of the protocol (Visit 7, Week 20 ± 14 days). All recordings were obtained in the afternoon during scheduled clinical visits.

Each recording session consisted of 30 min of resting‐state EEG recorded using a NicoletOne vEEG system (Viasys Neurocare, USA). EEG analyses were conducted using 22 scalp electrodes based on the international 10–20 system, grouped into five anatomical regions of interest (ROIs) for region‐wise profiling: frontal (Fp1, Fp2, F3, F4, F7, F8, Fz), central (C3, C4, Cz), parietal (P3, P4, P7, P8, Pz), occipital (O1, O2, Oz), and temporal (T7, T8, TP7, TP8). For ROI‐based analyses, channel‐level metrics were averaged across electrodes within each ROI to yield region‐level features for each frequency band.

EEG data were processed in MATLAB (MathWorks, USA) using EEGLAB workflows (Delorme and Makeig, 2004). ³¹ Resting‐state EEG segments were visually inspected, and artifact‐contaminated portions were excluded prior to spectral analysis. For each recording, up to 30 min of artifact‐free resting‐state data were analyzed; the signal was detrended to remove any direct‐current offset, and power spectral density was estimated using Welch's method (4‐s Hamming windows with 50% overlap). No time‐domain band‐pass filter was applied, and because the analysis was restricted to the 1–45 Hz range—below the 60‐Hz power‐line frequency—no notch filter was required. Band‐specific oscillatory activity was quantified using relative band power, calculated for each channel as the proportion of power within a given frequency band relative to the total power across 1–45 Hz. Frequency bands were defined as δ (1–4 Hz), θ (4–8 Hz), α (8–13 Hz), β (13–30 Hz), and γ (30–45 Hz). For each patient, pre–post changes in EEG metrics were computed as post–pre (ΔEEG) and were used for subsequent heatmap visualization, correlation analyses with seizure outcomes, and ROI‐based responder/nonresponder comparisons.

2.4. Structural MRI phenotyping

Baseline structural brain MRI was obtained for each patient on a 3.0‐T scanner (Ingenia CX; Philips Healthcare, Best, the Netherlands) using a standard clinical epilepsy protocol in accordance with the HARmonized Neuroimaging of Epilepsy Structural Sequences (HARNESS‐MRI) recommendations, including 3D T1‐weighted, 3D fluid‐attenuated inversion recovery (FLAIR), and high‐resolution coronal T2‐weighted sequences. Images acquired before taVNS were independently reviewed by an epileptologist and a neuroradiologist, and structural phenotypes were assigned by consensus.

For the exploratory analysis of treatment response, each baseline MRI was categorized along predefined structural dimensions. A scan was classified as lesional when a structural abnormality was identified and as nonlesional when reported as normal (no lesion/within normal limits). Lesional status and etiologic classification were treated as distinct dimensions. Etiology was classified according to the ILAE framework as structural (an identifiable causative lesion, such as hippocampal sclerosis, malformation of cortical development, or an infectious/inflammatory sequela) or unknown (no causative lesion identified). Limbic involvement was defined as structural abnormality affecting limbic or paralimbic structures—including the hippocampus, amygdala, mesial temporal region, insular cortex, or orbitofrontal cortex—and was further classified as unilateral or bilateral. A bilateral structural abnormality denoted structural involvement of both hemispheres irrespective of location. Structural phenotyping was applied only to baseline imaging and was used for exploratory, descriptive comparisons by responder status; no volumetric or quantitative MRI analysis was performed.

2.5. Primary efficacy endpoint: Seizure reduction ratio

Patients (and caregivers when applicable) maintained a standardized seizure diary throughout the baseline period and during each intervention period (active and sham). For each event, the diary captured the date/time of occurrence and brief clinical descriptors, including observable seizure type and count; seizure types were classified according to the updated ILAE classification of epileptic seizures as focal preserved consciousness seizures (FPC), focal impaired consciousness seizures (FIC), or focal to bilateral tonic–clonic seizures (FBTC). ³² At each study visit, study staff reviewed the diary with the patient to ensure completeness and to clarify ambiguous entries. Seizure frequency for each period was defined as the total number of diary‐recorded seizures within the corresponding period. The seizure reduction ratio (SRR) was calculated as:

$$\text{Seizure reduction ratio}\left(\%\right)=\frac{\text{Baseline Seizure frequency}-\text{Seizure frequency during the device}-\mathrm{use}\text{period}}{\text{Baseline Seizure frequency}}\times 100.$$

Seizure frequency was defined as the total number of epileptic seizures of any type (FPC, FIC, and FBTC) recorded in the epilepsy diary over the prespecified assessment period. Baseline seizure frequency was defined as the seizure count collected during the 8 weeks prior to Visit 2.

In the crossover protocol, the “device‐use period” seizure frequency corresponded to the seizure count during the assigned 8‐week stimulation phase (Visit 2–Visit 4 or Visit 5–Visit 7), depending on the randomized sequence.

2.6. Statistical analysis

All statistical analyses were performed using IBM SPSS Statistics (IBM Corp., Armonk, NY, USA). All tests were two‐sided, and statistical significance was defined as p < 0.05 unless otherwise specified. Continuous variables are presented as mean ± standard deviation (SD) or median with interquartile range (IQR), as appropriate, and categorical variables are summarized as counts and percentages. Analyses were performed in patients who met the compliance criteria defined in the present study. Seizure outcome analyses included those with available seizure diary data across both crossover phases, and EEG analyses included those with both baseline (V2) and end‐of‐protocol (V7) EEG recordings available.

2.6.1. Seizure outcome metrics

Seizure reduction ratio (SRR, %) was calculated separately for each stimulation phase relative to baseline, yielding SRR_active and SRR_sham. To quantify the active‐specific antiseizure effect within the crossover design, the within‐subject treatment effect was defined as ΔSRR = SRR_active − SRR_sham. Because seizure outcome data were not normally distributed and the sample size was small, within‐subject comparisons between active and sham stimulation were performed using the Wilcoxon signed‐rank test. To further explore treatment‐related effects within the two‐period crossover design, an exploratory linear mixed effects model was additionally fitted with SRR as the dependent variable, treatment, period, and sequence as fixed effects, and patient as a random intercept. Responder status was defined on the basis of the active phase only: patients with SRR_active ≥50% were classified as responders, and those with SRR_active <50% as non‐responders. Because responder status was defined using SRR_active rather than the active‐specific contrast (ΔSRR), responder‐stratified EEG analyses and the ΔSRR‐based correlation analyses should be regarded as addressing related but not identical questions.

2.6.2. EEG measures and association analyses

EEG features were defined as pre‐post changes in relative band power (ΔEEG = post [V7] − pre [V2]) for each frequency band (δ, θ, α, β, γ) and region of interest (frontal, central, parietal, occipital, and temporal). Channel‐level values were averaged within each region to obtain ROI‐level measures. Associations between ΔEEG and active‐specific seizure benefit (ΔSRR = SRR_active − SRR_sham) were assessed using Spearman's rank correlation coefficient for each band × ROI combination. Nominal p‐values were first presented for exploratory visualization. Because the β band was selected as the primary EEG band of interest on the basis of prior reports of vagal stimulation‐related β desynchronization, Benjamini–Hochberg FDR correction was additionally applied across the five β‐band ROIs.

2.6.3. Within‐group EEG changes by responder status

Within‐group pre–post changes in regional EEG measures were assessed separately in responders and nonresponders using the Wilcoxon‐signed rank test because of the small sample size. Between‐group comparisons of regional relative band power at baseline (V2) and at the end of the protocol (V7) were performed using the Mann–Whitney U test. For descriptive effect‐size estimation, Cohen's dz was calculated from paired difference scores for within‐group analyses, and rank‐biserial correlation was reported for between‐group analyses. For multiplicity control, Benjamini–Hochberg FDR correction was applied within each predefined analysis family, and adjusted q‐values are reported alongside raw p‐values. All responder‐based subgroup analyses were considered exploratory.

3. RESULTS

3.1. Patient Disposition and Clinical Profile

A total of 21 patients were assessed for eligibility, including 10 males and 11 females. Seven patients were excluded before randomization because they did not meet the prespecified taVNS adherence requirement (≥5 days/week; 120 ± 30 min/day) during the screening and baseline period or withdrew consent; the remaining 14 patients were randomized, completed the protocol with adequate adherence, and constituted the analyzed cohort. All analyzed patients met the prespecified usage target (≥5 days/week; 120 ± 30 min/day, delivered in 2–5 sessions per day), and adherence did not differ between responders and non‐responders. Among these patients, 8 were assigned to Group A (Active → Sham; 4 males and 4 females), and 6 were assigned to Group B (Sham → Active; 2 males and 4 females). All 14 analyzed patients were included in the diary‐evaluable clinical analysis, as well as in the exploratory EEG and structural MRI analyses. Patient flow and randomized crossover treatment sequences are summarized in Figure 1 in accordance with CONSORT recommendations for randomized crossover trials. ³³

The clinical characteristics of the randomized cohort reflected a heterogeneous population of patients with long‐standing drug‐resistant epilepsy. The mean age at enrollment was 34.1 ± 8.5 years, and the mean age at seizure onset was 13.71 ± 10.60 years. Patients were receiving multiple antiseizure medications (mean ASM number, 4.50 ± 0.94), consistent with a treatment‐refractory population. No patient had undergone prior resective epilepsy surgery. Benzodiazepine (clobazam) use did not differ between responders and nonresponders (5/7 vs. 4/7, p = 1.00). All 14 participants had focal epilepsy, and no participant met criteria for a generalized epilepsy syndrome. Baseline seizure types comprised FIC in 13 participants (92.9%), FBTC in 7 (50.0%), and FPC in 3 (21.4%). Focal to bilateral tonic–clonic seizures were more frequent among non‐responders (6/7 vs. 1/7, p = 0.029), and nonresponders exhibited a greater number of seizure types per patient (2.00 ± 0.58 vs. 1.29 ± 0.49, p = 0.038) (Table 1). Structural MRI abnormalities were observed in a substantial proportion of patients, including mesial temporal sclerosis, malformation of cortical development, infectious or inflammatory etiologies, and encephalomalacia, with both unilateral and bilateral involvement.

TABLE 1.

Clinical and MRI characteristics according to taVNS response status.

Characteristic	Total (n = 14)	Responder (n = 7)	Non‐responder (n = 7)	p‐value
Demographic characteristics
Female sex, n (%)	8 (57.1)	5 (71.4)	3 (42.9)	0.592
Age, years	33.9 ± 8.4	33.3 ± 7.0	34.4 ± 10.3	0.898
Disease‐related characteristics
Age at seizure onset, years	13.79 ± 10.56	11.57 ± 10.06	16.00 ± 11.34	0.482
Number of ASMs	4.50 ± 0.94	4.43 ± 0.98	4.57 ± 0.98	0.789
Etiology: unknown, n (%)	6 (42.9)	5 (71.4)	1 (14.3)	0.103
Etiology: hippocampal sclerosis, n (%)	2 (14.3)	1 (14.3)	1 (14.3)	1.000
Etiology: malformation of cortical development, n (%)	2 (14.3)	0 (0.0)	2 (28.6)	0.462
Etiology: infectious/inflammatory, n (%)	4 (28.6)	1 (14.3)	3 (42.9)	0.559
Baseline seizure characteristics
Seizure frequency (per 8 weeks), median [IQR]	13.0 [8.2–29.5]	28.0 [11.5–39.0]	8.0 [6.0–13.5]	0.097
FPC, n (%)	3 (21.4)	2 (28.6)	1 (14.3)	1.000
FIC, n (%)	13 (92.9)	6 (85.7)	7 (100.0)	1.000
FBTC, n (%)	7 (50.0)	1 (14.3)	6 (85.7)	0.029*
Number of seizure types per patient	1.64 ± 0.63	1.29 ± 0.49	2.00 ± 0.58	0.038*
Structural MRI phenotypes
Lesional MRI, n (%)	9 (64.3)	2 (28.6)	7 (100.0)	0.021*
Bilateral structural abnormality, n (%)	5 (35.7)	0 (0.0)	5 (71.4)	0.021*
Limbic involvement, n (%)	9 (64.3)	2 (28.6)	7 (100.0)	0.021*
Unilateral limbic involvement, n (%)	5 (35.7)	2 (28.6)	3 (42.9)	1.000
Bilateral limbic involvement, n (%)	4 (28.6)	0 (0.0)	4 (57.1)	0.070

Note: Values are mean ± SD or n (%). Mann–Whitney U test for continuous variables; Fisher's exact test for categorical variables. Lesional MRI denotes any structural abnormality on baseline imaging. Bilateral structural abnormality denotes involvement of both hemispheres irrespective of location; bilateral limbic involvement denotes bilateral involvement of limbic or paralimbic structures specifically.

Abbreviations: ASM, antiseizure medication; FBTC, focal to bilateral tonic–clonic seizure; FIC, focal impaired consciousness seizure; FPC, focal preserved consciousness seizure.

^* p < 0.05, uncorrected.

3.2. Primary clinical outcomes and individual response heterogeneity

In the diary‐evaluable set (n = 14), the mean seizure reduction ratio (SRR) was 46.26 ± 40.96% during active stimulation and 48.62 ± 37.27% during sham stimulation. The mean within‐subject difference (Active–Sham) was −2.37 ± 16.64%, indicating no significant difference between the two conditions (p = 0.604). Based on a responder threshold of ≥50% seizure reduction, 7 of 14 patients (50.0%) were classified as responders during active stimulation, whereas 8 of 14 (57.1%) met this criterion during sham stimulation (Table 2). In an exploratory linear mixed effects model including treatment, period, and sequence as fixed effects and patient as a random intercept, neither a period effect (p = 0.707) nor a sequence effect (p = 0.660) was observed, suggesting no appreciable crossover‐related bias in the within‐subject comparison. Individual paired SRR values for sham and active stimulation are shown in Figure 2 and demonstrate substantial within‐subject variability in treatment response. Most patients showed modest Active–Sham differences, whereas a small subset exhibited larger improvements or deteriorations.

TABLE 2.

Seizure reduction ratio during active and sham stimulation.

Outcome	Active	Sham	Active–sham (within‐subject)	p‐Value
Seizure reduction ratio (%), mean ± SD	46.2 ± 40.96	48.62 ± 37.27	−2.37 ± 16.64	0.604
Seizure reduction ratio (%), median (IQR)	48.2 [24.66, 81.24]	55.41 [27.16, 76.75]	2.48 [−9.58, 9.21]	–
Range (min to max)	−48.0 to 100.0	−21.47 to 93.01	−40.84 to 18.25	–
Responders (≥50% reduction), n (%)	7 (50.0%)	8 (57.1%)	–	–

Note: Values are shown as mean ± SD, median (IQR), range, or n (%), as appropriate. SRR was calculated relative to baseline seizure frequency. Active–Sham indicates the within‐subject difference between active and sham stimulation. The p‐value was calculated using the Wilcoxon‐signed rank test. Responders were defined as patients with ≥50% seizure reduction.

Abbreviations: IQR, interquartile range; SD, standard deviation; SRR, seizure reduction ratio.

FIGURE 2.

Seizure reduction ratio during sham and active stimulation (paired within‐subject comparison). Box plots of seizure reduction ratio (%) during sham and active stimulation overlaid with individual paired trajectories. Each dashed line connects sham and active values from the same patient, illustrating within‐subject changes. The box indicates the interquartile range with the median shown as a horizontal line; whiskers represent the data range. (Diary‐evaluable set; n = 14).

3.3. Correlation between regional EEG power modulations and active‐specific benefit

Among all patients (n = 14), we assessed whether changes in regional EEG relative band power (ΔEEG) were associated with the active‐specific clinical treatment effect, defined as the difference in seizure reduction ratio between active and sham stimulation (ΔSRR = SRR_active − SRR_sham). Nominal negative associations were concentrated in the β band, with the strongest effects observed in the occipital (ρ = −0.723, p = 0.003), frontal (ρ = −0.657, p = 0.011), and parietal regions (ρ = −0.552, p = 0.041), indicating that patients with greater β‐band reductions tended to show greater active‐specific clinical benefit (Figure 3). After Benjamini–Hochberg FDR correction restricted to the β‐band family across five ROIs (m = 5), the occipital (q = 0.017) and frontal associations (q = 0.027) remained significant, whereas the parietal association did not (q = 0.068) (Table 3). β‐band correlations in the central and temporal regions did not reach significance (q > 0.05), and no significant associations were detected in the δ, θ, α, or γ bands. Given this overall pattern, we next performed exploratory responder‐stratified analyses to determine whether distinct EEG change patterns were present in responders and non‐responders.

FIGURE 3.

Correlation between active‐specific clinical benefit (ΔSRR) and protocol‐related EEG changes (ΔEEG) across regions and frequency bands. Heatmap of Spearman correlations between active‐specific clinical benefit, defined as ΔSRR (SRR_active − SRR_sham), and protocol‐related EEG changes, defined as ΔEEG (Post[V7] − Pre[V2] relative band power), across scalp regions and frequency bands (n = 14). Cell colors represent Spearman's ρ, and overlaid text indicates uncorrected p values. White asterisks denote nominal uncorrected p < 0.05. No visual outline was used to avoid overemphasizing uncorrected findings. β‐band associations corrected for multiple comparisons within the predefined β‐band ROI family are summarized in Table 3.

TABLE 3.

β‐band Spearman correlations between ΔEEG (post[V7] − pre[V2]) and active‐specific clinical benefit (ΔSRR), with BH‐FDR correction within the β‐band family (m = 5).

Region	Spearman ρ	p‐Value	q (BH‐FDR, m = 5)
Central	−0.459	0.098	0.098
Frontal	−0.657	0.011	0.027*
Occipital	−0.723	0.003	0.017*
Parietal	−0.552	0.041	0.068
Temporal	−0.512	0.061	0.076

^* q < 0.05 after BH‐FDR within β‐band family (m = 5).

3.4. Phenotypic EEG trajectories associated with clinical response status

To further characterize the inter‐individual heterogeneity suggested by the overall association analysis, we examined within‐group changes in regional relative band power from baseline (V2) to the end of the protocol (V7) after stratifying patients by active‐phase responder status (Table S1). Responders showed nominal decreases in occipital α and β power, whereas nonresponders showed nominal central δ/θ shifts, occipital β increase, and frontal γ increase. These subgroup findings were exploratory and did not remain statistically significant after Benjamini–Hochberg FDR correction.

3.5. Divergent EEG signatures differentiating responders and non‐responders

We compared regional relative band power between responders and non‐responders at baseline (V2) and at the end of the protocol (V7) using Mann–Whitney U tests (n = 7 per group; responder status defined during the active phase; Table S2). No significant between‐group differences were observed at baseline across regions and frequency bands (all FDR‐adjusted q = 1.000). At V7, nominal differences were observed selectively in the β and γ bands, including lower occipital/parietal β power and higher central/temporal γ power among responders compared with non‐responders, but these findings did not remain significant after FDR correction (q = 0.109).

3.6. Neuroanatomical determinants of response: Baseline structural MRI phenotypes

Given the marked interindividual variability in clinical response observed in the present cohort, we explored whether baseline structural characteristics on MRI were associated with treatment responsiveness. In this cohort (n = 14; responders = 7, nonresponders = 7), bilateral structural abnormalities and limbic system involvement were more frequent in nonresponders (p = 0.021, respectively).

A lesional MRI was present in 2 of 7 responders compared with all 7 nonresponders (2/7 vs. 7/7, p = 0.021). None of the responders had bilateral structural abnormalities, whereas 5 of 7 nonresponders demonstrated bilaterality (0/7 vs. 5/7, p = 0.021). Lesional status and etiologic classification were not concordant: one participant with an isolated left hippocampal swelling with T2 hyperintensity, considered likely peri‐ictal rather than etiologic, was classified as lesional with limbic involvement while her etiology remained unknown (Table 1 and Table S3).

When limbic involvement was stratified by laterality, all patients with bilateral limbic involvement were nonresponders (4/4, 100%), whereas unilateral limbic involvement was distributed across both groups (2 of 5 responders vs. 3 of 5 nonresponders); given the small subgroup sizes, these laterality‐based patterns should be interpreted descriptively. Bilateral structural abnormality and bilateral limbic involvement were nested but not identical categories: one non‐responder with bilateral focal cortical dysplasia involving the left orbitofrontal cortex together with the right parieto‐occipital cortex had a bilateral structural abnormality whose limbic component was unilateral, the contralateral lesion being entirely extra‐limbic (Table 1 and Table S3).

4. DISCUSSION

4.1. Clinical efficacy trends and response heterogeneity

In this exploratory randomized, double‐blind, crossover trial of taVNS in patients with drug‐resistant focal epilepsy, active stimulation did not significantly outperform sham stimulation for seizure reduction. The mean seizure reduction ratio was numerically similar between active and sham conditions, and the within‐subject Active–Sham difference was not statistically significant. However, substantial interindividual variability was observed, supporting the need to examine biological factors that may explain why some patients appeared to derive greater active‐specific benefit than others.

The main exploratory finding was that greater attenuation of regional β‐band relative power was associated with greater active‐specific clinical benefit, defined as the difference in seizure reduction ratio between active and sham stimulation. This association was most evident in the occipital and frontal regions, where the β‐band correlations remained significant after Benjamini–Hochberg FDR correction restricted to the β‐band ROI family. In contrast, responder‐stratified EEG analyses and between‐group EEG comparisons showed only nominal differences that did not survive FDR correction. Structural MRI analyses further suggested that bilateral structural abnormalities and limbic involvement were more frequent among nonresponders. Taken together, these findings do not establish taVNS efficacy, but they suggest that EEG and MRI phenotyping may provide a useful framework for understanding inter‐individual variability in taVNS response.

4.2. Β‐Band desynchronization as a potential marker of target engagement

The observed association between β‐band attenuation and active‐specific clinical benefit is consistent with the concept that effective vagal neuromodulation may reduce pathological network synchronization in epilepsy. Prior studies of implanted VNS have reported stimulation‐related EEG desynchronization, reduced interictal synchrony, and changes in band‐limited power that correlate with favorable seizure outcomes. ²⁰ , ²¹ , ²² , ²³ , ²⁴ , ³⁴ Similar network‐level reductions in β‐ and γ‐band power have also been reported after ANT DBS in drug‐resistant epilepsy. ³⁵ In this context, reduced β‐band relative power may reflect a shift away from hypersynchronous cortical or thalamocortical activity, rather than a nonspecific change in background EEG activity. Several caveats temper this comparison, however. Most prior evidence derives from implanted cervical VNS, which differs from taVNS in stimulation route, delivered dose, and continuity, so target engagement and downstream network effects may not be equivalent. Moreover, much of the implanted‐VNS EEG evidence was obtained by contrasting stimulation ON versus OFF cycles on an acute timescale, whereas our resting‐state EEG compared baseline with the end of the protocol, so the two designs capture different temporal windows of neuromodulatory effect. Appropriately time‐locked comparisons of implanted VNS and taVNS on EEG, structural substrates, and clinical response will be needed to clarify the extent to which their mechanisms overlap.

The regional pattern observed in the present study may also be relevant. β‐band attenuation was most strongly associated with active‐specific benefit in the occipital and frontal regions, with an additional nominal association in the parietal region. These findings suggest that taVNS‐related clinical benefit may be linked to modulation of distributed cortical networks rather than to a single focal EEG change. The absence of a comparable association at temporal sites may reflect the concentration of structural pathology—such as hippocampal sclerosis and other mesial temporal lesions—in temporal regions in this cohort, which could obscure stimulation‐related spectral modulation; the smaller number of temporal electrodes and their greater susceptibility to myogenic artifact may further reduce sensitivity in this region. This interpretation is compatible with the known anatomy of vagal afferent signaling, which projects through the nucleus tractus solitarius and locus coeruleus–norepinephrine system to widespread thalamic, limbic, and cortical networks involved in seizure modulation. ⁶ , ⁷ , ⁸ , ³⁶

However, these findings should be interpreted cautiously. The β‐band analysis was exploratory, and FDR correction was applied within the β‐band ROI family rather than across all possible band‐by‐region comparisons. Therefore, β‐band attenuation should be regarded as a candidate physiological correlate of active‐specific benefit, not as a validated biomarker of taVNS target engagement. Confirmation in larger datasets with prespecified EEG endpoints will be necessary.

4.3. EEG phenotyping of responders and non‐responders

Responder‐stratified EEG analyses provided additional descriptive support for divergent neurophysiological trajectories. Responders showed nominal decreases in occipital α and β power from baseline to the end of the protocol, whereas nonresponders showed the opposite direction of occipital β change, together with central δ/θ shifts and frontal γ increase.

Although these subgroup findings did not survive FDR correction, their directional consistency with the overall β‐band association is noteworthy. The absence of baseline EEG differences between responders and nonresponders suggests that postprotocol EEG patterns may reflect differential network modulation during the trial rather than preexisting group differences alone. Nevertheless, because responder‐stratified analyses were based on only seven patients per group, these results should be interpreted as descriptive and hypothesis‐generating.

4.4. Structural network burden and reduced taVNS responsiveness

In addition to EEG findings, baseline structural MRI phenotypes appeared to differ between responders and nonresponders. Bilateral structural abnormalities, limbic involvement, and bilateral limbic involvement were more frequent among nonresponders, whereas responders were more often characterized by more focal or lateralized phenotypes. These findings suggest that both the presence and the spatial distribution of structural pathology may inform taVNS responsiveness.

This interpretation is consistent with the concept of structural “network burden.” Vagal neuromodulation is thought to influence seizure networks through brainstem, thalamocortical, and limbic pathways. ⁶ , ⁷ , ⁸ , ³⁶ If these networks are structurally disrupted, particularly in a bilateral or limbic pattern, the capacity of taVNS to modulate epileptic circuits may be reduced. Prior studies of VNS response predictors have reported heterogeneous findings regarding lesional status, but more recent work has emphasized that structural and network‐level features may be more relevant than conventional lesion classification alone. ⁵ , ⁷ , ³⁷ , ³⁸ Notably, some reports indicate that certain lesion‐related or multifocal epilepsies—such as tuberous sclerosis, post‐traumatic epilepsy, or symptomatic generalized epilepsies—may respond favorably to vagal neuromodulation, indicating that the mere presence of a structural lesion does not uniformly predict poor response. Rather, it may be the structural integrity of the specific limbic and paralimbic networks that relay vagal effects that is relevant. Consistent with this idea, in our exploratory cohort limbic involvement was more common among nonresponders than responders (100% vs. 29%; p = 0.021, uncorrected); this pattern supports the hypothesis that disruption of limbic networks reduces taVNS responsiveness, although it should still be regarded as hypothesis‐generating and requires confirmation in adequately powered studies.

Importantly, these MRI findings should not be interpreted as definitive predictors of taVNS outcome. The analysis was exploratory, involved a small number of patients, and was based on categorical structural phenotyping rather than quantitative MRI measures. Future studies should evaluate whether quantitative measures of limbic network integrity, lesion laterality, and structural connectivity improve prediction of taVNS benefit.

4.5. Clinical implications and future directions

The present findings have two potential implications. First, they reinforce the importance of sham‐controlled designs in taVNS trials.

In this study, seizure frequency decreased substantially from baseline under both conditions (mean SRR, 46.3% with active and 48.6% with sham stimulation), yet active stimulation conferred no additional benefit beyond sham. In an uncontrolled or open‐label design, the same improvement could easily have been misattributed to a specific effect of the device. The within‐subject sham comparison instead indicates that most of the observed improvement reflected nonspecific factors—including placebo response, regression to the mean, and natural fluctuation of seizure frequency—rather than a true active‐specific neuromodulatory effect.

Second, the results suggest that future taVNS studies in DRE may benefit from phenotype‐stratified designs. EEG β‐band modulation could be evaluated as a prespecified physiological endpoint, while MRI‐based assessment of limbic structural network burden may, if validated in larger studies, help identify patients less likely to respond. Such an approach may be particularly important for noninvasive neuromodulation, where treatment effects may be modest and heterogeneous across patients.

Several limitations should be acknowledged. The sample size was small, and the study was not powered to establish clinical efficacy or validate predictive biomarkers. EEG was measured only at baseline and at the end of the full protocol, rather than immediately before and after each active and sham phase, which limits phase‐specific interpretation. The sham condition used auditory output without transcutaneous electrical current; although designed to preserve blinding, differences in somatosensory experience cannot be fully excluded. Seizure outcomes were based on diary data, which may be affected by reporting variability. Finally, the MRI analyses were exploratory and should be validated using larger cohorts and quantitative imaging methods.

In addition, formal correlation analyses between MRI phenotypes and EEG changes were not performed given the small sample size; whether the observed EEG modulation is independent of underlying structural pathology therefore remains to be determined in larger cohorts. In addition, all patients had focal epilepsy, so these findings cannot be extrapolated to generalized epilepsy syndromes.

Overall, although active taVNS did not demonstrate superiority over sham stimulation in this cohort, the convergent EEG and MRI findings provide a biologically plausible, hypothesis‐generating signal. These data support further investigation of EEG–MRI phenotyping as a strategy to identify patients who may be more likely to benefit from taVNS and to optimize future clinical trial design.

5. CONCLUSION

In this exploratory randomized, double‐blind, crossover study of taVNS in drug‐resistant focal epilepsy, active stimulation did not significantly outperform sham stimulation for seizure reduction. However, greater attenuation of regional β‐band relative power, particularly in the occipital and frontal regions, was associated with greater active‐specific clinical benefit. In addition, bilateral structural abnormalities and limbic involvement were more frequent among nonresponders, suggesting that limbic structural network burden may contribute to reduced taVNS responsiveness. These findings do not establish clinical efficacy, but they support EEG and MRI phenotyping as a hypothesis‐generating framework for understanding inter‐individual variability in taVNS benefit. Larger phenotype‐stratified studies with prespecified EEG and imaging endpoints are needed to validate these candidate biomarkers and refine patient selection for noninvasive vagal neuromodulation.

AUTHOR CONTRIBUTIONS

S.J.H.: conceptualization, methodology, EEG signal processing and analysis, statistical analysis, visualization, and writing—original draft. S.‐Y.Y.: conceptualization, patient recruitment and clinical assessment, structural MRI phenotyping, and writing—original draft. E.K.B. and J.C.L.: clinical trial coordination, data collection, and data management. Y.K.: data curation and analysis support. K.M.: development and provision of the stimulation device and technical support. Y.‐M.S.: conceptualization, study supervision, clinical oversight, funding acquisition, and writing—review and editing. All authors critically reviewed the manuscript and approved the final version.

FUNDING INFORMATION

This research was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS‐2024‐00357879) and Future Medicine 2030 Project of the Samsung Medical Center (#SMX126001) and National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (2022M3C1A3081294).

CONFLICT OF INTEREST STATEMENT

K.M. is an employee of TODOC Co., Ltd., the manufacturer of the allears device used in this study. The remaining authors declare no conflicts of interest. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.

Supporting information

Table S1.

Table S2.

Table S3.

Hong SJ, Yang S‐Y, Bok EK, Lee JC, Kim Y, Min K, et al. Transcutaneous auricular vagus nerve stimulation in drug‐resistant epilepsy: A randomized, sham‐controlled crossover trial with exploratory EEG and MRI response phenotyping. Epilepsia Open. 2026;00:1–14. 10.1002/epi4.70354

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request. The data are not publicly available because they contain information that could compromise the privacy of research participants.