耳迷走神经调控治疗重性抑郁障碍的脑网络影像证据系统综述

Abstract

Background:

Auricular vagus neuromodulation, encompassing transcutaneous auricular vagus nerve stimulation (taVNS) and transcutaneous electrical cranial–auricular acupoint stimulation (TECAS), is an emerging noninvasive intervention for major depressive disorder (MDD). Neuroimaging studies of its central effects have expanded rapidly, yet findings remain dispersed across heterogeneous designs and have not been integrated at the cohort level.

Methods:

We followed PRISMA 2020 guidelines and searched five databases (PubMed/MEDLINE, Embase, Cochrane CENTRAL, Web of Science Core Collection, and PsycINFO) from inception to June 30, 2026; only English-language reports were included. Two reviewers independently screened records, extracted effect-level data, and appraised methodological quality using the 2024 JBI tool for quasi-experimental studies. Reports from the same cohort were identified via trial registrations, and denominators were counted once per cohort. Findings were stratified by five imaging paradigms and synthesized via cohort-level vote-counting, with null effects reported alongside positive findings.

Results:

Seventeen neuroimaging reports arose from nine independent cohorts (321 enrolled patients with MDD, 293 in imaging analyses, 144 healthy controls). The most reproducible finding was bidirectional normalization in striatal–cortical reward-circuit regions across four cohorts, qualified by null nucleus accumbens and putamen seed findings in one cohort. Default-mode-network regions were implicated in seven of nine cohorts, with heterogeneous directions. The only anterior insula effect surviving between-group correction was confined to an acute stimulation-state paradigm, and stimulation-state connectivity changes were frequency-band-specific. Twenty-seven of the coded effects were explicitly null. The sole machine-learning model (AUC = 0.856) was evaluated only by internal cross-validation. Quality appraisal rated two reports high, fourteen moderate, and one low.

Conclusion:

The evidence remains strictly hypothesis-generating: auricular vagus neuromodulation is accompanied by distributed treatment-associated brain changes, but establishes neither causal pathways nor clinically actionable imaging markers. Confirmatory multi-center, sham-controlled trials with preregistered imaging endpoints, harmonized stimulation parameters, explicit null-result reporting, and independent external validation are needed.

Systematic review registration:

https://www.crd.york.ac.uk/PROSPERO/view/CRD420261421762, identifier CRD420261421762.

1 Introduction

Major depressive disorder (MDD) affects more than 280 million people worldwide and is a leading cause of disability (). Its core features—persistent low mood, anhedonia, cognitive impairment, and functional decline—contribute to excess mortality, reduced quality of life, and substantial socioeconomic costs (). Available treatments include antidepressant medications (), psychotherapy, electroconvulsive therapy (ECT) (), and repetitive transcranial magnetic stimulation (rTMS) (). Nevertheless, many patients respond inadequately to initial treatment, others develop treatment-related adverse effects, and some fail to maintain remission after initial improvement (). Safe, effective, biologically grounded treatment strategies remain urgently needed.

Conventional transcutaneous auricular vagus nerve stimulation (taVNS) and transcutaneous electrical cranial–auricular acupoint stimulation (TECAS) are noninvasive auricular vagus neuromodulation approaches under investigation for MDD (, ). Both stimulate the auricular branch of the vagus nerve but differ in technique. Conventional taVNS targets the cymba conchae or tragus, where vagal afferent fibers are dense, and typically uses low-frequency stimulation (20 Hz or alternating 4/20 Hz). TECAS, rooted in traditional Chinese medicine, stimulates defined auricular acupoints, sometimes combined with scalp acupoints such as Baihui (GV20) and Yintang (EX-HN3), and uses different parameters (e.g., the 4/20 Hz dense-dispersed wave). These differences in montage, frequency, and concurrent acupoint engagement may recruit partially distinct afferent pathways and central circuits. Neither approach requires surgical implantation; both engage central autonomic and limbic pathways, carry fewer procedural risks than invasive VNS—whose implantation can cause infection, hoarseness, and voice alteration ()—and are convenient for repeated, long-term use (). Clinical evidence suggests that auricular vagus neuromodulation can improve depressive symptoms with favorable safety and tolerability (), yet its neurobiological basis remains poorly understood, limiting protocol refinement and patient selection.

Neuroimaging offers a way to examine the central effects of neuromodulation (). Contemporary models frame MDD as disruption of distributed large-scale networks rather than of discrete regions; these networks support cognitive control, salience detection, self-referential processing, and reward processing (). Studies have therefore used resting-state functional magnetic resonance imaging (rs-fMRI), structural MRI, diffusion tensor imaging (DTI), and other techniques to characterize the neural response to auricular vagus neuromodulation. Findings implicate several depression-relevant networks, notably the default mode network (DMN) (), salience network (SN) (), frontoparietal/cognitive control network (FPN/CCN), and limbic–corticostriatal reward circuit (). These networks interact rather than operate independently: the DMN and SN show anticorrelated dynamics, with engagement of one typically accompanied by suppression of the other; the FPN couples flexibly with the SN during cognitively demanding tasks; and the reward circuit shares nodes with the DMN, including the ventromedial prefrontal and posterior cingulate cortices (, –). How auricular vagus neuromodulation reshapes these between-network interactions may be central to its antidepressant effects.

Previous reviews and meta-analyses have summarized the clinical efficacy of taVNS for depression () and its mechanisms, parameters, and applications across neurological and psychiatric conditions, but none has focused on neuroimaging evidence in MDD or integrated findings across analytical approaches—functional connectivity, independent component analysis (ICA), ALFF, ReHo, FCD, and machine learning—into a network-level framework. This literature is fragmented: stimulation protocols, imaging modalities, analyses, treatment duration, and outcome measures vary widely across studies (). Direct comparison is further complicated by the use of distinct imaging metrics, including functional connectivity (FC), amplitude of low-frequency fluctuations (ALFF) (), regional homogeneity (ReHo) (), functional connectivity density (FCD) (), and machine-learning models (). Individual studies thus report numerous imaging alterations, but convergent changes across large-scale networks have not been systematically integrated.

To address these gaps, we systematically reviewed neuroimaging studies of auricular vagus neuromodulation in MDD, with four objectives: (1) to consolidate the available neuroimaging evidence; (2) to identify the large-scale networks and principal brain regions repeatedly modulated by treatment; (3) to integrate findings linked to clinical improvement and candidate imaging predictors of treatment response; and (4) to formulate a network-level account of the central effects underlying antidepressant action.

2 Methods

2.1 Search strategy

The review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) statement (). Five databases—PubMed/MEDLINE, Embase (Elsevier), the Cochrane Central Register of Controlled Trials (CENTRAL), Web of Science Core Collection, and PsycINFO (Ovid)—were searched from inception through June 30, 2026. Search strategies combined controlled vocabulary (Medical Subject Headings, Emtree, and APA Thesaurus terms) with free-text terms in three conceptual blocks: (1) major depression; (2) auricular vagus neuromodulation, including transcutaneous auricular vagus nerve stimulation (taVNS) and transcutaneous electrical cranial–auricular acupoint stimulation (TECAS); and (3) neuroimaging, encompassing functional and structural MRI, functional connectivity, network/connectome analyses, and electroencephalography (EEG) terminology, the last included to maximize retrieval sensitivity. No language or publication-type restrictions were applied at the search stage; reports using EEG as the sole neuroimaging modality and reports in languages other than English were excluded at screening under the pre-specified eligibility criteria. Full database-specific strategies, including the Embase strategy rebuilt with Emtree explosion terms and the complete auricular-vagus/TECAS synonym set, are provided in Supplementary Table S1. Reference lists of eligible articles and relevant reviews were hand-searched for additional publications. The protocol was registered with PROSPERO (CRD420261421762). Three deviations from the registered protocol were made: expansion from three to five databases, addition of a linked-report identification procedure (§2.3), and replacement of the 2020 JBI critical appraisal tool by the 2024 revised version (§2.5). Reports using EEG as the sole modality were excluded because, unlike MRI, EEG cannot localize treatment-related changes to the large-scale distributed networks and MNI-coordinate-based regions that frame this synthesis, and source-localized EEG findings cannot be integrated with the MRI evidence base within a common anatomical framework.

2.2 Eligibility criteria

Eligibility followed the PICOS framework. Population: adults (≥18 years) with major depressive disorder diagnosed by DSM or ICD criteria, including first-episode drug-naïve, recurrent, and treatment-resistant presentations; studies enrolling minors or participants without a primary MDD diagnosis were excluded. Intervention: auricular vagus neuromodulation—taVNS (including transauricular VNS) or TECAS targeting the auricular branch of the vagus nerve; invasive VNS, manual acupuncture, and non-auricular body electro-acupuncture were excluded. Comparator: sham stimulation, active pharmacological comparators (e.g., escitalopram), active within-subject stimulation-condition comparators (e.g., a second active stimulation parameter setting), healthy comparison participants, or within-subject pre-treatment baselines; no comparator was required, and single-arm pre–post studies were eligible. Outcomes: treatment-related magnetic resonance neuroimaging outcomes, including resting-state fMRI indices (ALFF/fALFF, ReHo, seed-based and whole-brain functional connectivity, independent component analysis/functional network connectivity), task-based fMRI activation, graph-theoretical/connectome topology and network-based statistics, machine-learning neuroimaging features, and structural or diffusion MRI measures. Study design: original peer-reviewed pre–post, controlled (sham- or active-controlled, randomized or non-randomized), or single-arm neuroimaging studies, including acute single-session and longitudinal (4–8-week) paradigms; cross-sectional studies without a treatment-related imaging endpoint, reviews, meta-analyses, editorials, letters, conference abstracts, protocols without results, case reports, and animal or in vitro studies were excluded. Eligible reports were in English and reported neuroimaging findings adequate for qualitative synthesis.

2.3 Study selection

Retrieved citations were imported into EndNote X9.1 (Clarivate, Philadelphia, PA, USA), and duplicates were removed before screening. Two reviewers independently screened titles and abstracts and then assessed full texts against the pre-specified eligibility criteria; disagreements were resolved by discussion, with unresolved decisions referred to a third reviewer.

Multiple reports from the same parent trial were identified through trial registration numbers, ethics approvals, author groups, and participant overlap. Following PRISMA 2020 guidance on companion reports, linked reports were treated as a single study unit: each report contributed its imaging metric or stimulation timepoint, whereas participant denominators were counted once at the cohort level.

Database searches yielded 1,131 records across five databases; after removal of 293 duplicates, 838 unique database records remained, and together with 4 reports identified by hand-searching reference lists, 842 records underwent title and abstract screening. Full texts were sought for 39 reports; 22 did not meet the criteria, and 17 neuroimaging reports from 9 independent participant cohorts entered the qualitative synthesis. Reports excluded after full-text assessment are listed with their reasons in Supplementary Table S2, and the complete selection pathway is presented in Figure 1. Inter-reviewer agreement supported reliable study selection at both screening stages; percentage agreement at each stage is reported in the Supplementary Methods.

Figure 1

2.4 Data extraction

Two reviewers independently extracted data using a standardized form. Recorded variables included study design and participant characteristics, diagnostic criteria, stimulation protocols, neuroimaging modalities, analytical methods, affected brain regions and functional networks, clinical outcome measures, and associations between imaging findings and treatment response. Resting-state fMRI measures were recorded as reported: amplitude of low-frequency fluctuations (ALFF/fALFF) indexes the intensity of local spontaneous BOLD signal; regional homogeneity (ReHo), the local synchrony of neighboring voxels; functional connectivity density (FCD), the number of functional connections of each voxel at whole-brain level; seed-based functional connectivity (FC), the correlation of a seed region’s timecourse with all other voxels; independent component analysis (ICA) and functional network connectivity (FNC), spatiotemporal networks and their between-network coupling; graph-theoretical metrics and network-based statistics (NBS), connectome topology and connected subcomponents; and machine-learning analyses, multifeature predictive models. ALFF, ReHo, and FCD describe local activity, whereas seed FC, ICA/FNC, and NBS describe connectivity; the two classes can change in different directions and are not interchangeable. Task-fMRI BOLD activation and during-scan on/off stimulation-state FC were classified separately. Extraction differences were resolved by discussion and consensus.

Reports drawing on the same participant cohort were linked as described in §2.3, with participant denominators counted once per cohort. Inter-reviewer agreement for data extraction was high; the percentage agreement is reported in the Supplementary Methods. The complete effect-level coding matrix—recording for each effect its region, network attribution, direction relative to the stated comparator, comparator type, seed or whole-brain scope, multiplicity correction, and reported statistics—is provided as Supplementary Table S3 (126 coded effects).

2.5 Quality assessment

Methodological quality (risk of bias) was appraised independently by two reviewers using the revised Joanna Briggs Institute (JBI) Critical Appraisal Tool for Quasi-Experimental Studies (non-randomized experimental studies) (). No included study allocated participants to study groups at random: the eligible reports comprised sham-controlled sequential cohorts, non-randomized active-comparator trials, single-arm pre–post studies, and acute within-session designs, as well as one within-subject cross-over study in which only the order of the two active stimulation conditions was randomized. Because that randomization concerned condition order rather than participant allocation, the cross-over study does not constitute a randomized controlled trial; the authors themselves stated that their study was “not designed as a clinical trial.” The JBI randomized-controlled-trial checklist presupposes random allocation, and the prevalence and cross-sectional checklists are intended for non-interventional designs, so the quasi-experimental tool was the appropriate instrument. The revised tool comprises nine signaling questions spanning internal validity (temporal precedence; selection and allocation; confounding; intervention administration; outcome assessment, detection, and measurement; and participant retention) and statistical conclusion validity. Each item was answered “yes,” “no,” “unclear,” or “not applicable,” and appraisal was performed at the study level for each report’s primary neuroimaging outcome(s). “Not applicable” was used only where an item was logically inapplicable, such as differential co-interventions in single-group pre–post designs; the reliable-outcome-measurement item applies to every design, including single-arm studies, because standardized acquisition and processing are required regardless of whether a comparison group exists. For single-group pre–post comparisons, the participant-similarity item was answered “yes” per the JBI manual, as participants serve as their own controls; an independent control group was credited only where a separate group (sham arm, active-comparator arm, or healthy controls undergoing identical pre- and post-intervention measurements) was available, and not where healthy controls were measured at baseline only. Disagreements were resolved by consensus; inter-rater agreement for the item-by-item quality judgments is reported in the Supplementary Methods. No study was excluded based on quality; appraisal findings were used to weight and interpret the evidence. Item-level judgments and justifications for all 17 reports are provided in Supplementary Table S4.

The JBI prescribes no numerical cut-off scores for its appraisal tools and cautions against relying on aggregate scores alone; we therefore defined our grading rule a priori and report it transparently. For each report, the proportion of “yes” judgments among applicable items (excluding “not applicable”) was calculated, following the convention used in published JBI-based reviews in which ≥70% “yes” denotes high, 50–69% moderate, and <50% low methodological quality. This proportion was interpreted alongside five items judged critical for causal inference in this evidence base—control group (Q2), participant comparability/confounding (Q3), reliable outcome measurement (Q7), complete follow-up/retention (Q8), and appropriate statistical analysis including multiplicity control (Q9): reports with ≥70% “yes” and no “no” or “unclear” judgment on any critical item were graded high quality (low risk of bias); reports with 50–69% “yes,” or ≥70% “yes” with any “no”/”unclear” critical-item judgment, were graded moderate quality, with single-arm/within-subject studies lacking an independent control group (Q2 = “no” by design) capped at moderate; and reports with <50% “yes,” a “no” judgment on statistical conclusion validity (Q9), or “no” judgments on two or more critical items were graded low quality (high risk of bias).

2.6 Data synthesis

Given heterogeneity across studies in stimulation protocols, imaging modalities, analytical methods, and outcome measures, quantitative meta-analysis was not feasible and findings were synthesized qualitatively. Synthesis was performed by cohort and stratified across five non-interchangeable imaging paradigms, kept separate throughout: (1) longitudinal resting-state fMRI before and after 4–8 weeks of treatment; (2) acute single-session resting-state fMRI immediately before and after one stimulation session; (3) block-design stimulation fMRI during the first session, with stimulation-on blocks modeled against implicit baseline; (4) concurrent stimulation-state connectivity with on/off blocks during scanning; and (5) acute single-session task-based fMRI with stimulation delivered around a stress task (one cross-over respiratory-gated RAVANS study, in which both compared conditions were active parameter settings and all findings were acute/state-dependent). Key synthesis decisions were made a priori as follows.

2.6.1 Network classification

Network attribution followed two rules, recorded separately for every effect. First, labels assigned by the original authors were retained and explicitly reported as the authors’ interpretation, introduced with the phrase “the authors interpreted.” Second, networks were attributed anatomically from peak coordinates using fixed a priori definitions: the default mode network comprised the precuneus/posterior cingulate cortex, medial prefrontal cortex, angular gyrus, middle temporal gyrus, and hippocampus; the salience network, the anterior insula and anterior/mid-cingulate cortex; the striatal–cortical reward circuit, the nucleus accumbens, caudate, putamen, globus pallidus, and ventral striatum with their medial, orbitofrontal, and ventrolateral prefrontal projections; and the frontoparietal/control network, the dorsolateral prefrontal cortex, middle frontal gyrus, and inferior parietal lobule. Regional signal change was never equated with network-level psychological function. Effects outside these four labels—including sensorimotor and visual networks, the hypothalamus, and the cerebellum—were retained under their original anatomical system rather than reassigned. Regional membership of the cortical network labels followed the Automated Anatomical Labeling atlas (AAL3) (), chosen because it is the parcellation most frequently used by the included studies for region definition and seed placement, which maximizes correspondence with the primary reports; subcortical and brainstem labels (ventral striatum, nucleus accumbens, hypothalamus, and brainstem relay nuclei) followed the Harvard–Oxford subcortical structural atlas (), which provides probabilistic anatomical boundaries in standard MNI space. All coordinates were expressed in MNI space before attribution, allowing consistent atlas-based assignment regardless of the parcellation used in the primary report.

2.6.2 Directionality

Null findings (seeds, group contrasts, and post-hoc tests without significant treatment-related change) were recorded and are reported in parallel with positive findings. Because studies differed in seed regions, baseline comparisons, and statistical thresholds, the direction of change relative to the stated comparator was recorded for every effect; effects in opposite directions were never combined into a single “convergent” count, vote counts were direction-specific, and conflicts within or across reports are reported as inconsistencies with candidate methodological explanations. “Convergent” is used only for same-direction, multiplicity-corrected effects across independent cohorts.

2.6.3 Acute versus long-term effects

Acute single-session and longitudinal studies were analyzed separately to distinguish immediate stimulation effects from sustained treatment-related changes; when a study reported both, each timepoint was described independently.

Overlapping cohorts. Linked reports from the same parent trial (§2.3) were treated as a single study unit: participant denominators and cohort-level vote counts were counted once per independent cohort, and overlapping results were not double-counted.

2.6.4 Multiple imaging outcomes and analysis scope

All reported findings were included without selective extraction; when one study contributed results to more than one network category, each was recorded under the corresponding network, as some regional effects may reflect overlapping neural systems rather than independent observations. Studies using a priori seed regions or small-volume correction were recorded as ROI-restricted: their findings cannot speak to regions outside the search space, and vote-counting was performed only within the anatomical systems each study actually examined.

3 Results

3.1 Evidence base: linked reports, independent cohorts, and imaging paradigms

Seventeen neuroimaging reports met the eligibility criteria (Figure 1; Table 1). Audit of trial registrations, ethics approvals, and participant overlap showed that these reports arose from nine independent participant cohorts rather than seventeen independent samples. Three cohorts were organized under registered parent trials and each contributed linked analyses applying different metrics or stimulation timepoints to overlapping participants: Cohort C1 (parent trial ChiCTR-TRC-11001201; the parent clinical report, Rong et al., 2016 (), enrolled 160 MDD patients, of whom the imaging subcohort enrolled 49), reported across five linked analyses (Fang et al., 2016 (); Liu et al., 2016 (); Fang et al., 2017 (); Tu et al., 2018 (); Wang et al., 2018 ()), a single-blind, non-randomized sequential cohort with sham stimulation at the non-vagal ear margin, no healthy controls, analyzed subsets N = 34–38; Cohort C2 (ChiCTR-1800014277; ethics 2017-021-SQ), reported across three linked analyses (Ma et al., 2022 (); Sun et al., 2023 (); Sun et al., 2024 ()), maximum 86 MDD and 54 controls; and Cohort C6 (ChiCTR2000029109; TECAS; ethics 2019-250-KY/YJKY2019052), reported across three linked analyses (Liao et al., 2023 (); Ma et al., 2023 (); Ma et al., 2024 ()), maximum 51 MDD and 51 controls. The remaining cohorts were single-report studies: C3 (Chen et al., 2022 (); 30 enrolled, 22 analyzed), C4 (He et al., 2023 (); 22 MDD, 23 controls), C5 (Zhang et al., 2022 (); NCT03607331; 15 MDD, 16 controls), C7 (Yi et al., 2022 (); 20 analyzed), and C8 (Guo et al., 2024 (); 19 analyzed), and C9 (Garcia et al., 2021 (); 20 enrolled, 18 analyzed). Counting each cohort once, the independent evidence base comprised 321 enrolled MDD participants, 293 with analyzed imaging, and 144 healthy controls.

Table 1

Characteristics of the 17 included neuroimaging reports arising from nine independent participant cohorts (C1–C9).

1. Reports are ordered by cohort (C1–C9). Reports sharing a parent trial analyze overlapping participants with different imaging metrics or stimulation timepoints and are counted once; participant denominators are counted only once at the cohort level (321 enrolled MDD, 293 with analyzed imaging, 144 healthy controls).

2. N columns report participants included in the imaging analysis; enrolled/recruited numbers are given in parentheses where they differ (e.g., Fang et al., 2016 (): 34 analyzed from 49 enrolled; Zhang et al., 2022 (): 15 analyzed from 16 enrolled; Chen et al., 2022 (): 22 analyzed from 30 enrolled).

3. Three interventions are distinguished: taVNS = transcutaneous auricular vagus nerve stimulation (auricular concha only); TECAS, transcutaneous electrical cranial–auricular acupoint stimulation, combining auricular taVNS with transcranial electrical stimulation at GV20 (Baihui) and EX-HN3 (Yintang), in trigeminal as well as vagal territories; TECAS was used by only one cohort (C6). RAVANS, respiratory-gated auricular vagal afferent nerve stimulation; e-RAVANS, expiratory-gated; i-RAVANS, inspiratory-gated.

4. Five non-interchangeable imaging paradigms: Tier 1 = longitudinal resting-state fMRI (pre/post 4–8 weeks); Tier 2 = acute single-session resting-state fMRI (immediate pre/post one 30-min stimulation); Tier 3 = block-design stimulation fMRI during the first session (stimulation-on blocks vs implicit baseline; Fang et al., 2017 () only); Tier 4 = concurrent stimulation-state connectivity with on/off blocks during scanning (Tu et al., 2018 (); Wang et al., 2018 ()); Tier 5 = acute single-session task-based fMRI with stimulation delivered around a stress task (Garcia et al., 2021 (); C9 only; cross-over of two active RAVANS parameter settings, not a treatment trial).

5. Analysis scope: WB, whole-brain voxelwise exploratory analysis; ROI, a priori seed/region of interest with small-volume or cluster correction; ICA, independent component analysis network-level; graph/NBS, graph-theory/network-based connectome-level analysis.

6. GRF, Gaussian random field cluster-level correction; FWE, family-wise error correction; FDR, false discovery rate; SVC, small-volume correction. Correction details are as reported in the primary papers. Detailed findings including null results are in Table 3 (Parts A and B); item-level JBI ratings are in Supplementary Table S4. For Garcia et al., 2021 () (C9), the methods text states cluster-level FDR P <.05 whereas the results tables are labeled FWE-corrected; we report as published.

Three stimulation approaches were distinguished: conventional taVNS (auricular concha only); TECAS, which combines auricular taVNS with transcranial electrical stimulation at cranial acupoints GV20 (Baihui) and EX-HN3 (Yintang), in trigeminal as well as vagal territories (only Cohort C6); and respiratory-gated auricular vagal afferent nerve stimulation (RAVANS), a respiratory-gated taVNS variant delivered to the cymba concha, in which expiratory- versus inspiratory-gated timing serves as an active parameter comparator (only Cohort C9; Garcia et al., 2021 (); a cross-over experimental study, not a treatment trial). The imaging experiments fell into five non-interchangeable paradigms (tiers): (1) longitudinal resting-state fMRI before and after 4–8 weeks of treatment (seven cohorts); (2) acute single-session resting-state fMRI immediately before and after one 30-min stimulation (C2, C7); (3) block-design stimulation fMRI during the first session, with stimulation-on blocks modeled against implicit baseline (Fang et al., 2017 () only); (4) concurrent stimulation-state connectivity with on/off blocks during scanning (Tu et al., 2018 (); Wang et al., 2018 ()); and (5) acute single-session task-based fMRI with stimulation delivered around a stress task (Garcia et al., 2021 (), C9 only; a cross-over comparison of expiratory- versus inspiratory-gated RAVANS, both active parameter settings; the authors explicitly state that the study was not designed as a clinical trial). Tiers 3, 4 and 5 are not pre/post-treatment resting-state studies and are synthesized separately; tier-5 evidence is acute and state-dependent. A given cohort could contribute reports to more than one tier, but its participants were counted once in all denominators. The strongest evidence came from group × time or between-group contrasts surviving FWE or GRF cluster-level correction; the between-network FNC analyses of Liao et al. (2023) () survived only uncorrected p <.05 (none survived FDR) and were treated as exploratory, as were most brain–symptom correlations. Methodological quality (JBI critical appraisal) is summarized in Table 2; grades indicate risk of bias within this non-randomized evidence base and do not constitute a formal evidence-certainty hierarchy.

Table 2

Methodological quality (risk of bias) of the 17 included reports.

1. Item-by-item ratings (yes/no/unclear/not applicable) with study-level justifications are provided in Supplementary Table S4. Grades follow the a priori rule defined in Methods (§2.5): high = ≥70% of applicable items met with no “no” or “unclear” judgment on the five critical items (control group, baseline comparability, reliable outcome measurement, complete follow-up, appropriate statistical analysis); moderate = 50–69% of items met, or ≥70% with a deficient critical item—single-group and within-subject studies without an independent control group are capped at moderate; low = <50% of items met, a “no” on statistical analysis, or “no” on two or more critical items. Denominators are the number of applicable items: 9 for studies with an independent control group (sham arm, active comparator, or healthy controls measured pre- and post-treatment) or a cross-over design, and 8 for single-group pre–post studies, for which the control-group item is not applicable.

2. No included study used random allocation of participants to study groups: all reports were non-randomized intervention studies (sequential cohort enrollment, non-randomized assignment to TECAS versus escitalopram, single-arm designs, or within-subject parameter cross-over). The single low-grade report also reported its between-network functional connectivity contrasts at uncorrected p <.05 only; those contrasts are treated as exploratory throughout the synthesis.

3.2 Methodological quality

Risk of bias was appraised for all 17 included reports (arising from 9 independent participant cohorts) with the revised JBI critical appraisal tool for quasi-experimental studies (Supplementary Table S4). Two reports (11.8%) were rated high quality (low risk of bias)—Ma et al., 2022 () (9/9 applicable items met) and Zhang et al., 2022 () (9/9)—both of which combined an independent comparison group measured at both time points, complete follow-up with explicit accounting of exclusions, and whole-brain inferences under Gaussian random field correction with prespecified covariates. Fourteen reports (82.4%) were rated moderate quality, and one report (5.9%)—Liao et al., 2023 ()—was rated low quality (high risk of bias). At the cohort level, one cohort (C5) was high quality, six cohorts (C1, C3, C4, C7–C9) were moderate, C2—although it contained the high-rated Ma et al., 2022 () report—was classified moderate because its other two linked reports (Sun et al., 2023 (); Sun et al., 2024 ()) were moderate, and one cohort (C6) spanned moderate to low because one of its three linked reports (Liao et al., 2023 ()) received a low rating; the two other reports from that cohort (Ma et al., 2023 (); Ma et al., 2024 ()) were moderate.

Several weaknesses recurred across the evidence base and temper the strength of causal inference. First, no included study used random allocation of participants to study groups: reports were non-randomized (sequential cohort enrollment, non-randomized assignment to TECAS versus escitalopram, or one within-subject cross-over in which only the order of two active stimulation conditions was randomized, or single-arm designs), so selection and allocation bias cannot be excluded; baseline between-group comparability could not be confirmed from the published texts for the five C1 reports and for the two active-comparator C6 reports (Liao et al., 2023 (); Ma et al., 2023 ()). Second, seven reports (Chen et al., 2022 (); Sun et al., 2023 (); Sun et al., 2024 (); Ma et al., 2024 (); Yi et al., 2022 (); Guo et al., 2024 () and Garcia et al., 2021 ()) lacked any independent control group for the pre–post treatment contrast (healthy controls, where present, were measured at baseline only), leaving natural fluctuation, regression to the mean, and expectancy effects uncontrolled; notably, the five C1 reports did include a sham-controlled arm, and Ma et al., 2022 () He et al., 2023 (), and Zhang et al., 2022 () included healthy controls undergoing identical pre- and post-intervention measurements. Third, reporting of attrition was incomplete or unclear in eight reports (the five C1 reports, He et al., 2023 (), Sun et al., 2023 (), and Liao et al., 2023 ()), where participant flow was described in aggregate but group-wise losses and their potential impact were not analyzed. Fourth, statistical rigor varied: whereas the primary imaging contrasts in 16 of the 17 reports were protected by cluster-level correction (GRF/Gaussian random field, FWE, or FDR; e.g., voxel P<.005–.01 with cluster P<.05, or cluster FWE/FDR), the headline between-network functional connectivity treatment effects in Liao et al., 2023 () were reported at an uncorrected threshold of P<.05, and the corresponding FDR-corrected between-network analyses were entirely null; that report was graded low quality accordingly. In addition, most brain–clinical correlation analyses across studies were not multiplicity-corrected and should be regarded as exploratory. Fifth, samples were small (MDD participants analyzed per report ranged from 15 to 77; Sun et al., 2024 () enrolled 86, of whom 77 were modeled after motion exclusions and 57 had paired pre–post scans); in total, 293 MDD and 144 healthy controls were analyzed across the 9 independent cohorts, precluding formal assessment of publication bias and limiting precision. Against these limitations, temporal precedence of intervention over outcome was unambiguous in every report, outcomes were measured with identical protocols and standardized processing pipelines within each study (all 17 reports met the “same way” and “reliable measurement” items), and null findings were increasingly reported in the more recent literature. Overall, the neuroimaging evidence is methodologically coherent in measurement but uniformly observational/quasi-experimental in design; network-level findings should therefore be interpreted as evidence of treatment-associated brain changes warranting confirmation in randomized, adequately powered, sham-controlled trials with pre-registered analysis plans.

3.3 Longitudinal resting-state fMRI (4–8 weeks)

Sham-controlled taVNS (C1). Fang et al. (2016) (), using ICA dual-regression of the DMN (Z > 2.6, cluster p <.05), found group × time interactions in both directions: DMN coupling with regions including right anterior insula and parahippocampal gyrus decreased more after real than sham, whereas DMN coupling with left precuneus and lateral orbitofrontal cortex increased more after real; additional opercular, occipital, and medial-frontal clusters showed the reverse pattern; cluster-level DMN-connectivity correlations with HAMD change ran in both directions (uncorrected, exploratory). Liu et al. (2016) (), using amygdala-seed FC (voxel p <.005, cluster FWE), found right amygdala–left DLPFC connectivity to increase after real (p = .031), decrease after sham (p = .012), and differ between groups (p = .011); the left amygdala seed yielded no significant cluster. Larger connectivity increases correlated with greater reductions in HAMD (r = −0.42), anxiety (r = −0.55), and retardation (r = −0.46) (exploratory).

Single-arm and control-comparator taVNS cohorts. Chen et al. (2022) () (C3; 22 patients with recurrent MDD, no controls) examined eight basal-ganglia seeds under GRF correction: bilateral nucleus accumbens seeds showed no significant clusters (P = .07) and bilateral putamen seeds were null (P >.05); globus pallidus seeds showed decreased FC with sensorimotor, inferior-parietal/precuneus, and visual regions, and the right caudate seed decreased FC with cerebellar/visual regions, with larger decreases correlating with greater symptom improvement (r up to −0.60; exploratory); normalization toward controls could not be tested. He et al. (2023) () (C4; 22 first-episode MDD, 23 twice-scanned untreated controls) found baseline lower cortical ALFF (dlPFC, OFC) and higher subcortical ALFF (ACC, dmPFC, ventral striatum, anterior thalamus, hippocampus, amygdala). Group × time interactions (GRF voxel p <.001) showed right ventral-striatal ALFF, ReHo, and FCD—hyperactive at baseline—decreasing toward control levels after treatment (F = 21.1/18.8/22.3), whereas precuneus/PCC clusters showed the opposite pattern, ALFF and ReHo increasing; controls showed no significant retest changes. Ventral-striatum–precuneus FC decreased (p = .048; exploratory) whereas an analogous control ventral-striatum–medial-OFC connection was unchanged (p = .934); larger ventral-striatal decreases correlated with higher remission rates (r = −0.763; r = −0.869; exploratory). Zhang et al. (2022) () (C5; 15 MDD, 16 controls), using six bilateral striatal subregion seeds with GRF-corrected interactions, found bidirectional normalization: hypoconnected striatal–prefrontal links (ventral caudate–ventrolateral PFC, NAc–dmPFC/vlPFC) increased after treatment, whereas hyperconnected striatal–visual/temporal links (to superior occipital, cuneus, middle temporal cortex) decreased; intra-striatal dorsal-caudate–putamen connectivity increased; correlations with HAMD change were exploratory (r = −0.586, −0.584, 0.629).

Guo et al. (2024) () (C8; 19 MDD, no controls) applied graph theory and NBS (FDR; 5,000-permutation FWE). Global efficiency increased and characteristic path length decreased (both p <.001), indicating greater global integration; gamma, lambda, sigma, and local efficiency showed no significant changes. Left angular gyrus degree centrality and nodal efficiency increased (FDR); all other nodes were null. An NBS component of 16 nodes and 10 edges—between-network connections the authors assigned to DMN–FPN, DMN–CON, and FPN–CON—decreased after treatment; topological changes correlated with HAMD-24 reduction (r = 0.616; r = −0.703; exploratory). Sun et al. (2024) () (C2; 86 enrolled, 77 modeled after motion exclusions, 57 with paired scans) used baseline whole-brain FC in an elastic-net model with nested leave-one-out cross-validation: internal cross-validated prediction of symptom reduction gave R = 0.66 (R² = 0.44) and responder classification AUC = 0.856 (accuracy 76.6%), with no external validation, sham arm, or controls; features concentrated in cortico-striato-pallido-thalamic, hippocampal/parahippocampal, and cerebellar connections; two vermis–hippocampal features were nonsignificant.

Active-controlled TECAS cohort (C6). Liao et al. (2023) () (50 MDD [29 TECAS, 21 escitalopram], 49 controls), using group-ICA, found lower DMN within-network connectivity in MDD at baseline (FWE-corrected). An insula-network group × time interaction and several time/group main effects reached FWE, but FDR-corrected post-hoc pre/post comparisons were null for all these networks, and between-network FNC showed no significant differences after FDR; at uncorrected p <.05 (presented as exploratory), differences involved DMN–DAN, DMN–RFPN, and visual–RFPN pairs, with post-TECAS and post-escitalopram directions largely opposite. Baseline medial-frontal/ACC connectivity correlated with baseline symptom severity (exploratory), while network changes showed no association with clinical change in either group. Ma et al. (2023) () (51 MDD [34 TECAS, 17 escitalopram], 51 controls) found lower baseline ALFF and ReHo in right medial superior frontal cortex (t = −5.07/−4.76; anatomically consistent with the DMN); medial-SFG-seed FC with occipital/temporal regions was elevated at baseline and decreased after treatment in both treatment groups (interactions P = .011–.017, GRF), whereas ALFF/ReHo group × time interactions yielded no significant clusters and no imaging–clinical correlations were reported. Ma et al. (2024) () (the 34 TECAS patients, 34 matched controls) found lower baseline left-insula ALFF (t = −4.30) and lower insula–prefrontal/ACC FC (discussed by the authors within an insula–DMN framework; the insula is anatomically consistent with a salience-network region); insula–ACC/orbitofrontal and insula–SFG/SMA connectivity increased after 8 weeks, and patients no longer differed significantly from controls post-treatment, though full normalization was not demonstrated (p = .075; p = .187; null); insula–ACC FC increase correlated with HAMD reduction (r = −0.34; exploratory). Clinically, response and remission rates did not differ significantly between TECAS and escitalopram in this small sample (HAMD response 62.9% vs 57.1%, P = .55; remission 29.4% vs 35.3%, P = .67); these null differences do not support equivalence or non-inferiority.

3.4 Acute single-session resting-state fMRI (immediate pre/post)

Ma et al. (2022) () (C2; 40 patients with treatment-resistant depression, 40 controls; one 30-min 4/20-Hz session) found a group × time interaction for ReHo in right medial orbitofrontal cortex (F = 18.06, GRF-corrected)—a region that the authors interpreted as part of the reward network: baseline ReHo was higher in patients and decreased toward control levels after stimulation, with no significant control change. Across-group time effects comprised ReHo decreases in cerebellar, inferior-temporal, orbitofrontal, and superior-frontal clusters and a precuneus increase; group main effects were null for both ReHo and seed FC. Right-medial-OFC seed FC with left inferior parietal cortex increased in patients after stimulation (F = 11.66), whereas a second OFC–supramarginal cluster showed a null patient post-hoc change (P = .113) despite a significant interaction; imaging changes showed no association with HAMD or HAMA change. Sun et al. (2023) () (C2; 55 MDD, 54 controls; one 20-Hz session) found baseline lower right-precuneus and higher right-angular-gyrus ALFF in MDD (DMN-anatomical regions); immediately after stimulation, right-precuneus ALFF decreased (per the abstract/results text; the sign printed in the results table is internally inconsistent) and precuneus-seed FC with left middle frontal cortex, PCC, and angular gyrus decreased (all GRF-corrected); baseline precuneus–PCC FC correlated with baseline HAMD (r = −0.288; exploratory). Yi et al. (2022) () (C7; 20 first-episode drug-naïve MDD, no controls) found ReHo to decrease immediately after the first session in seven clusters—bilateral mid-cingulate, left precentral/postcentral, right calcarine/lingual, left SMA, and paracentral lobule (cluster FWE)—with no cluster showing increased ReHo; MCC and SMA ReHo changes correlated with 4-week HAMD improvement (r = 0.62; r = 0.74; uncorrected, exploratory).

3.5 Block-design stimulation fMRI during the first session

Only Fang et al. (2017) ()(C1; N = 38, real 17/sham 21) used this design (stimulation-on blocks vs implicit baseline; FWE). Real stimulation activated bilateral cerebellar crus I and left opercular/insular/precentral cortex and deactivated right medial frontal cortex; sham activated right middle/inferior frontal cortex, and sham-specific deactivations encompassed PCC, precuneus, hippocampus, and medial frontal cortex—DMN-anatomical deactivations observed under sham and not specific to taVNS. The real > sham contrast yielded a single surviving region, the left anterior insula (FWE small-volume correction); the reverse contrast produced no significant cluster and no brainstem/nucleus-of-the-solitary-tract activation was observed. Left-anterior-insula activation correlated with 4-week HAMD improvement in real (r = −0.65, p = .01) but not in sham (p = .6; exploratory).

3.6 Stimulation-state connectivity during scanning (on/off blocks)

Tu et al. (2018) () (C1; voxel p <.001, cluster FDR), using hypothalamic seeds during on/off blocks, found medial-hypothalamus connectivity with left rACC and right MFG to be lower, and with cerebellum higher, in real than sham during stimulation; the medial-hypothalamus–rACC on-vs-off change was significant in real (p = .004) but not sham (p = .18). This connection correlated with HAMD improvement in real during stimulation (r = 0.58, p = .02) but not in sham (p = .15), and showed no significant correlation in the resting (off) state; hypothalamus–MFG/cerebellar connections showing changes did not correlate clinically. Lateral-hypothalamus connectivity with MCC and putamen was lower in real than sham but did not correlate with outcome, and four preset non-hypothalamic control seeds yielded no significant clusters. Wang et al. (2018) () (C1; N = 37), using NAc seeds with frequency stratification (voxel p <.005, cluster FWE), found frequency-specific effects: in slow-5 (0.01–0.027 Hz), left-NAc connectivity with bilateral MPFC/rACC increased during stimulation in real (on vs off) and was higher in real than sham (Z = 5.07); slow-4/typical-band on/off effects were null in real, sham showed no change in any band, and slow-4 between-group effects were null. In the typical band, right-NAc connectivity with occipital, lingual/fusiform, and insular regions was higher in real than sham, while the reverse contrast was null; slow-5 NAc–prefrontal connectivity correlated with HAMD improvement in real only. The authors framed NAc-seed connections within a reward-circuit model; the MPFC/rACC targets are anatomically consistent with DMN/limbic regions.

3.7 Acute single-session task-based fMRI (stimulation around a stress task)

Garcia et al. (2021) () reported a cross-over experimental study in 20 enrolled premenopausal women (18 with complete paired sessions; NCT04467164) in which respiratory-gated transcutaneous auricular vagus nerve stimulation (RAVANS; cymba concha) was delivered in two active conditions in randomized order during block-design task fMRI. Unlike the resting-state paradigms synthesized in Sections 3.3–3.6, this study modeled stimulation-on blocks against an implicit baseline within the same scanning session, so all findings were state-dependent rather than reflecting pre–post treatment changes. Whole-brain analyses revealed connectivity changes that varied significantly across stimulation conditions, with the most consistent effects observed in regions receiving vagal afferent projections. Seed-based functional connectivity used a priori seed regions in the nucleus tractus solitarius (NTS) and dorsal raphe nucleus (DRN), the key vagal-afferent relay stations, with an additional hypothalamus (HYPO) seed; all were defined from the authors’ prior RAVANS studies rather than from an atlas. The reported patterns were anatomically consistent with the expected vagal afferent pathway, although direct activation of these brainstem nuclei was not measurable at the available imaging resolution. The methods text states cluster-level P <.05 corrected for false discovery rate, whereas the results tables are labeled family-wise error corrected; we report the findings as published.

3.8 Cross-paradigm synthesis and evidence grading

Vote-counting across independent cohorts (a qualitative tally, not a meta-analysis; Table 3) yielded four patterns. First, longitudinal striatal/corticostriatal changes—examined with ventral-striatal, NAc, caudate, globus-pallidus, and putamen seeds—appeared in four cohorts (C2 at feature level; C3; C4; C5), predominantly as bidirectional normalization (baseline hyperactivity/hyperconnectivity decreasing and baseline hypoconnectivity increasing toward control levels), although the null NAc/putamen seeds in C3 (P = .07) indicate seed/subregion dependence. Second, DMN-anatomical regions (precuneus, PCC, medial frontal cortex, angular gyrus, hippocampus) were implicated in seven cohorts, but direction was heterogeneous: local-activity metrics (ALFF, ReHo, degree centrality) mostly increased or normalized after treatment whereas connectivity metrics (seed FC, FNC, NBS) mostly decreased—no unidirectional DMN enhancement or suppression is supported. Third, the strongest between-group insula/salience evidence came from the first-session block design (left anterior insula, FWE-SVC), with supporting longitudinal seed-FC normalization; acute resting-state studies did not consistently detect insula effects. Fourth, acute resting-state changes were predominantly local-activity decreases, whereas stimulation-state effects were state-dependent and frequency-specific (slow-5 only) and cannot be pooled with longitudinal findings. Null results were pervasive and are reported in parallel throughout—null seeds and control connections, null group main effects, null FDR-corrected post-hoc and between-network ICA findings, null ALFF/ReHo interactions, and null imaging–clinical correlations and global graph metrics. Table 3 summarizes the paradigm-stratified evidence, and Supplementary Figure S1 presents the corresponding paradigm-stratified evidence map; mechanistic interpretation is developed as hypothesis in the Discussion.

Table 3

Synthesis of neuroimaging evidence for auricular vagus neuromodulation in major depressive disorder.

1. The 17 included reports derive from 9 independent participant cohorts; reports sharing a parent trial (C1: ChiCTR-TRC-11001201, five linked analyses; C2: ChiCTR-1800014277, three linked analyses; C6: ChiCTR2000029109 [TECAS], three linked analyses) analyze overlapping participants with different imaging metrics or stimulation timepoints and are counted once; participant denominators are counted only once (321 enrolled MDD, 293 with analyzed imaging, 144 healthy controls).

2. Five non-interchangeable imaging paradigms are kept separate: Tier 1 longitudinal resting-state fMRI (pre/post 4–8 weeks); Tier 2 acute single-session resting-state fMRI (immediate pre/post one 30-min stimulation); Tier 3 block-design stimulation fMRI during the first session (stimulation-on blocks vs implicit baseline; Fang et al., 2017 () only); Tier 4 concurrent stimulation-state connectivity with on/off blocks during scanning (Tu et al., 2018 (); Wang et al., 2018 ()); Tier 5 = acute single-session task-based fMRI with stimulation delivered around a stress task (Garcia et al., 2021 ()).

3. Network labels are attributed using two conventions: those explicitly stated by the original authors are reported verbatim; those inferred from peak-coordinate anatomy are described as “anatomically consistent with” the corresponding network. Regional activation is not equated with network-level mechanism.

4. Correction symbols: GRF, Gaussian random field cluster-level correction; FWE, family-wise error correction (SVC, small-volume correction over a priori regions); FDR, false discovery rate; cluster-level p values indicate corrected significance; CAN, central autonomic network; “uncorrected/exploratory” denotes effects reported only at uncorrected p<.05 (the Liao et al., 2023 () between-network FNC findings) or brain–symptom correlations without multiple-comparison correction.

5. null = the analysis was performed but no significant cluster/association was detected; null results are reported alongside positive effects and were not excluded from the synthesis.

6. Direction is reported relative to the stated comparator (group×time interaction, between-group, on-vs-off, or within-group pre/post); normalization denotes movement of baseline-deviant values toward healthy-control levels and is only asserted where a control comparison was made.

7. The AUC of 0.856 (Sun et al., 2024 (); R² = 0.44) reflects internal nested leave-one-out cross-validation within a single-arm cohort without external validation and is an internal cross-validated exploratory predictive signature, not a clinical marker.

8. TECAS (combined transcranial cranial-acupoint + auricular stimulation, targeting trigeminal and vagal territories) is reported separately from auricular-only taVNS; response and remission rates did not differ significantly between TECAS and escitalopram (HAMD response 62.9% vs 57.1%, P = .55; remission 29.4% vs 35.3%, P = .67), and these null differences do not support equivalence or non-inferiority.

9. All coordinates are MNI; coordinates marked unverifiable in primary reports (e.g., Wang et al., 2018 () typical-band insular cluster) are flagged in Supplementary Table S3. Directional conflicts internal to individual primary reports (e.g., the Sun et al., 2023 () precuneus ALFF sign) are resolved toward the abstract/results-text direction and flagged for author clarification.

1. Full study characteristics (intervention parameters, design, imaging metrics, analysis scope, multiplicity correction) are in Table 1; JBI item-level ratings are in Supplementary Table S4.

2. Comparator abbreviations: group×time = treatment-group by time interaction; on-vs-off = stimulation-on versus stimulation-off blocks during scanning; within-group pre/post = single-arm pre-treatment versus post-treatment comparison; baseline MDD vs HC = cross-sectional patient–control comparison.

3. Direction is given relative to the stated comparator; ↑/↓ denote increase/decrease of the measured metric. “null” denotes analyses that were performed but yielded no significant cluster or association; the complete 126-row effect-level coding matrix (including 27 explicitly null rows) is provided in Supplementary Table S3.

4. Findings from Tiers 2–5 are not pooled with Tier 1 longitudinal findings in any vote count; vote-counting was across independent cohorts and cannot estimate pooled effect sizes (meta-analysis not feasible; see Methods).

4 Discussion

4.1 Principal findings

This review synthesized neuroimaging evidence from 17 reports arising from 9 independent participant cohorts (321 enrolled patients with MDD, of whom 293 entered imaging analyses, together with 144 healthy controls; participant denominators were counted once at the cohort level). Because stimulation timepoints, imaging tasks, and analysis levels differed markedly across reports, findings were organized by cohort and by five non-interchangeable imaging paradigms (Sections 3.1 and 3.3–3.8) rather than pooled; no coordinate-based meta-analysis was attempted, and synthesis relied on cohort-level vote-counting.

At the cohort level, the most reproducible pattern involved striatal–cortical regions. We use this term anatomically, to denote the nucleus accumbens (NAc), ventral and dorsal striatum (caudate, putamen, globus pallidus, ventral striatum [VSi]), and their projections to medial and orbitofrontal prefrontal cortices (mOFC; vmPFC/dmPFC; vlPFC), a distributed system anatomically consistent with the corticostriatal reward circuit. Across four independent cohorts (C2–C5), findings in these regions consistently showed bidirectional normalization: values that were abnormally high at baseline decreased after treatment, whereas values that were abnormally low increased, with post-treatment values moving toward those of healthy controls. In He et al. (2023) (), VSi ALFF, ReHo, and functional connectivity density were elevated at baseline and decreased after 8 weeks of taVNS (group × time interaction, GRF-corrected), whereas healthy controls scanned twice over the same interval showed no significant change. In Zhang et al. (2022) (), striatal–prefrontal connectivity that was low at baseline (ventral caudate–ventrolateral PFC; NAc–dmPFC/vlPFC) increased after treatment, whereas abnormally high striatal–visual connectivity (superior occipital gyrus; cuneus) decreased; within-striatum connectivity (dorsal caudate–dorsolateral putamen) increased. In Chen et al. (2022) (), globus pallidus– and caudate-seeded connectivity to sensorimotor and visual cortices decreased after 8 weeks, whereas the bilateral NAc seeds yielded no significant clusters (P = .07) and the bilateral putamen seeds were likewise null (P >.05). Beyond these longitudinal effects, two additional analyses from the C2 cohort are noteworthy: an acute experiment examining normalization of elevated mOFC regional homogeneity within a single 30-min session in treatment-resistant depression (Ma et al., 2022) (), and a longitudinal machine-learning analysis reporting that baseline connectivity features enriched in the corticostriatal–pallidal–thalamic circuit, hippocampus, and cerebellum predicted subsequent symptom change (Sun et al., 2024 (); evidence-strength qualification in Section 4.4).

Regions of the default mode network (DMN), including the precuneus/PCC, medial prefrontal cortex, angular gyrus, middle temporal gyrus, and hippocampus, were involved in 7 of 9 cohorts, but with markedly heterogeneous directions. Local activity metrics (ALFF, ReHo, degree centrality) mostly increased or normalized (precuneus clusters in He et al., 2023 (); left angular gyrus nodal degree and efficiency in Guo et al., 2024 ()), whereas connectivity metrics (seed-based FC, ICA, NBS) mostly decreased (bidirectional DMN dual-regression changes in Fang et al., 2016 (); VSi–precuneus FC in He et al., 2023 (); an NBS component spanning DMN–frontoparietal and DMN–cingulo-opercular connections in Guo et al., 2024 (); baseline DMN within-network connectivity in Liao et al., 2023 ()). The direction of precuneus findings was inconsistent across reports: precuneus ALFF was reported as decreased in Sun et al. (2023) () whereas a time-effect increase in precuneus ReHo was reported in Ma et al. (2022) (). A unidirectional account of DMN involvement is therefore not supported.

The anterior insula was the only regional target that survived between-group correction in a task/stimulation-state paradigm: in Fang et al. (2017) (), left anterior insula activation during stimulation blocks was greater with real than with sham stimulation (FWE small-volume correction) and correlated with 4-week symptom improvement in the real-stimulation group only. Longitudinal insula-seed findings (Ma et al., 2024) () showed within-group increases in insula–ACC/orbitofrontal connectivity after 8 weeks of TECAS, but post-treatment values did not differ significantly from controls (P = .075 and P = .187), so formal normalization was not demonstrated. At the whole-brain network level, one graph-theory cohort (Guo et al., 2024 ()) reported increased global efficiency, decreased characteristic path length, and increased left angular gyrus nodal degree and efficiency (FDR-corrected), with other global metrics null. Finally, in C1, amygdala–DLPFC connectivity increased after 4 weeks relative to sham (Liu et al., 2016 (), cluster-level FWE), while the left-amygdala seed analysis was null.

4.2 Putative neurobiological mechanisms

All circuit-level interpretations in this section are hypotheses generated by the synthesis (Figure 2); none was tested causally by the included studies, and network labels were assigned post hoc rather than specified in advance.

Figure 2

The afferent pathway most often invoked to explain central effects of auricular stimulation runs from auricular vagal afferents to the nucleus tractus solitarius (NTS), and thence to the locus coeruleus, raphe nuclei, thalamus, hypothalamus, and forebrain structures; this account derives from anatomical tracing, invasive vagus nerve stimulation, and transcutaneous stimulation imaging studies (–). No included study measured brainstem nuclei directly, and Fang et al. (2017) () explicitly reported that no NTS activation was detected during stimulation blocks. Garcia et al. (2021) () employed a priori seed regions in the nucleus tractus solitarius (NTS) and dorsal raphe nucleus (DRN), derived from the authors’ prior RAVANS studies and reported state-dependent connectivity changes anatomically consistent with the expected vagal afferent pathway, but direct NTS activation was likewise not measurable at the available resolution. The ascending pathway should therefore be read as an interpretive scaffold drawn from prior literature rather than as a pathway demonstrated in the reviewed studies. Furthermore, because TECAS combines auricular stimulation with transcranial electrical stimulation at cranial acupoints in trigeminal as well as vagal territories (Section 3.1), findings from the TECAS cohort cannot be attributed specifically to vagal afferent activation.

A first set of hypotheses concerns the default mode and salience systems. DMN-related changes could be hypothesized to reflect altered self-referential processing or rumination (, , 51), but the opposite directions observed for local-activity versus connectivity metrics (Section 4.1) preclude any coherent directional claim. Acute engagement of the left anterior insula in the block-design experiment () is anatomically consistent with known cortical projection targets of vagal afferents and with the role of salience-network regions in interoception and affective-state switching (); it supports a hypothetical early cortical relay, but pertains exclusively to stimulation-state and task paradigms and cannot be generalized to longitudinal treatment effects.

A second hypothesis concerns reward and motivation circuitry. The striatal–cortical normalization pattern is anatomically consistent with putative involvement of corticostriatal systems implicated in anhedonia and motivated behavior (52); several author teams interpreted their findings in these terms, and a link between striatal changes and anhedonia remains a plausible hypothesis. This interpretation is qualified by the null bilateral NAc-seed (P = .07) and putamen-seed (P >.05) results in Chen et al. (2022) (), which indicate that conclusions depend on striatal subregion and seed choice; moreover, imaging changes were observed alongside symptomatic improvement but were not shown to mediate it.

A third hypothesis concerns cognitive-control circuitry. Frontoparietal findings generate the hypothesis that treatment-associated changes extend to systems supporting attention and executive function, domains that are broadly impaired in MDD (53, 54). This is the least substantiated of the regional hypotheses: frontoparietal involvement appeared in only three cohorts, and the between-network functional connectivity results of Liao et al. (2023) () were significant only at uncorrected thresholds and were entirely null after FDR correction.

Finally, the single-cohort graph-theory results () generate the hypothesis of enhanced global network integration with reduction of abnormally elevated between-network connectivity, but this requires independent replication. Across all four hypotheses, two cautions apply. First, describing a cluster as belonging to the “DMN” or the “reward circuit” is an anatomical attribution; reverse inference from regional signal change to psychological function is not licensed by these data (). Second, the observations establish treatment-associated brain changes, not causal pathways; confirmation requires randomized, adequately powered, sham-controlled trials with preregistered network-level endpoints.

4.3 Clinical implications

From a clinical perspective, the reviewed evidence shows that vagus nerve stimulation (including TECAS) is accompanied by distributed, treatment-associated brain changes rather than effects confined to a single region, and that these changes occur in systems repeatedly implicated in the symptom dimensions of major depressive disorder. This pattern is consistent with the view of MDD as a disorder of large-scale network dysfunction rather than focal regional abnormality (, ), and provides a neurobiological rationale for continued clinical evaluation of auricular stimulation. The present data do not, however, establish that clinical benefit is produced by any particular network change; imaging changes were observed alongside symptomatic improvement, and mediation was not tested.

The distributed anatomical pattern may also help frame the heterogeneity of clinical response. Depressive symptoms span emotional regulation, reward processing, and cognitive control rather than dysfunction within a single domain (, 53), and the regional findings reviewed above correspond to these domains only at the level of anatomical attribution. Whether changes in particular systems track improvement in particular symptom dimensions remains an open question: most imaging–clinical correlations were uncorrected and are regarded as exploratory, and no study performed a formal mediation analysis. Longitudinal studies with preregistered symptom-domain endpoints are needed before neural changes can be used to explain interindividual differences in response.

With respect to clinical translation, no imaging feature in the reviewed evidence is ready for clinical use. One study reported a baseline connectivity model that discriminated subsequent treatment responders with an AUC of 0.856 under internal nested cross-validation (); because the model was developed within a single-arm design and evaluated only through internal cross-validation, it constitutes an exploratory predictive signature that requires external validation in independent, preferably sham-controlled, cohorts (, 55, 56). Candidate imaging features may in future complement symptom-based assessment for patient selection and treatment monitoring, but such applications remain hypothetical and should not inform clinical decision-making until prospectively validated in multicenter studies.

4.4 Limitations

Several limitations qualify the interpretation of this synthesis and should be read alongside the paradigm-stratified Results (Sections 3.1 and 3.3–3.8) and the cohort map in Table 1. First, the evidence base is small and interdependent. The 17 reports arose from 9 independent cohorts; analyzed cohort sizes ranged from 15 to 86 MDD participants, and four cohorts (C1, C3, C7, C8) included no healthy controls, so baseline abnormality and treatment-associated normalization could not be tested in those cohorts. Participant denominators were therefore counted once at the cohort level: linked reports within the three registered parent trials (C1, C2, C6) applied different imaging metrics or stimulation timepoints to overlapping participants and do not constitute independent replications (, , ). Most cohorts were enrolled at one or a small number of collaborating centers, and all came from a small number of collaborating research groups, which further limits the independence and generalizability of the observed patterns.

Second, the five imaging paradigms are not interchangeable. Longitudinal resting-state scans before and after 4–8 weeks of treatment, acute resting-state scans immediately around a single 30-minute session, block-design task fMRI during the first stimulation session (), and connectivity measured during continuous on/off stimulation in the scanner (, ) address different questions and cannot be pooled as one resting-state finding. Stimulation-state effects were also state- and frequency-specific: effects in the NAc-seed experiment were confined to the slow-5 band (0.01–0.027 Hz), with slow-4 and typical-band within-group analyses and all sham-group analyses null (). Coordinates, seeds, stimulation protocols, acquisition parameters, and analysis pipelines likewise varied across reports. These differences precluded the prespecified quantitative meta-analysis described in the Methods; cohort-level vote-counting can rank reproducibility across cohorts but cannot estimate pooled effect sizes or quantify between-study heterogeneity. With only nine independent cohorts, formal tests for publication bias or small-study effects were not feasible, and selective reporting of positive regional findings cannot be excluded in a literature dominated by small, exploratory neuroimaging studies with substantial analytical flexibility (57, 58). Finally, although no language filter was applied during the five-database search, eligibility was restricted to English-language reports; non-English records entered the screening pool and were excluded at screening, so relevant evidence published in other languages may have been missed.

Third, explicitly null analyses were common and are reported in parallel in Table 3 (Parts A and B), rather than treated as absences of evidence. The effect-level coding matrix compiled for this review contains 126 coded effect rows, including 27 explicitly null effect rows (Supplementary Table S3), concentrated in sham or control contrasts, particular seed choices, corrected network-level analyses, and absent imaging–clinical correlations. Examples include null findings for all frequency bands in the sham group of the NAc stimulation-state study (); a null resting-state imaging–clinical correlation and four null non-hypothalamic control seeds in the hypothalamic stimulation-state study (); no significant retest changes among healthy controls and an unchanged ventral-striatum–medial-OFC control connection (P = .934) (); null bilateral NAc (P = .07) and putamen seeds (P >.05) (); a null left-amygdala seed (); null FDR-corrected post-hoc and between-network ICA comparisons (); no significant changes in non-nodal global graph metrics (); and absent or null imaging–clinical correlations in several cohorts, including the TECAS ICA report and the acute treatment-resistant-depression report (, ). In addition, insula–prefrontal/ACC connectivity values after 8 weeks of TECAS did not differ significantly from control values at conventional thresholds (P = .075 and P = .187), so formal normalization was not demonstrated (). Most brain–symptom correlations were not protected by multiplicity correction and should be considered exploratory (57, 58); the uncorrected between-network FNC findings in the TECAS ICA report are likewise exploratory and were excluded from the main synthesis ().

Fourth, the only machine-learning prediction model remains exploratory. Sun et al. (2024) () reported R² = 0.44 and responder-classification AUC = 0.856, but these estimates came solely from internal nested leave-one-out cross-validation in a single cohort; there was no independent test cohort, external validation, sham arm, or prospective clinical evaluation. The resulting connectivity pattern is therefore an exploratory predictive signature, not a clinical marker, and model performance may be optimistic in a field where small samples and analytical flexibility are known to inflate brain-wide associations (, 57, 58). Independent replication in adequately powered, sham-controlled cohorts is required before any imaging feature can inform treatment selection (, 56).

Fifth, causal attribution is constrained by design and by the intervention itself. No included study used randomized allocation: the C1 parent investigation was a single-blind, non-randomized sequential cohort study with a sham control (), and the TECAS cohort used non-randomized assignment to TECAS versus escitalopram (). Selection, allocation, and expectancy effects therefore cannot be excluded. The credibility of sham stimulation is also uncertain because auricular stimulation produces cutaneous tingling and somatosensory sensations that may reveal group assignment and influence both neural measurements and self-reported symptoms (). No study tested mediation, so treatment-associated imaging changes were observed alongside clinical improvement but were not shown to mediate it. Moreover, TECAS combines auricular stimulation with cranial acupoints (GV20 Baihui; EX-HN3 Yintang) in trigeminal as well as vagal territories; findings from that single cohort therefore cannot be attributed to auricular vagal afferents alone (, ). Clinically, response and remission rates did not differ significantly between TECAS and escitalopram in this small sample (HAMD response 62.9% vs 57.1%, P = .55; remission 29.4% vs 35.3%, P = .67); these null differences do not support equivalence or non-inferiority (, ).

Sixth, measurement and participant-level confounds were controlled inconsistently. Although several protocols required medication washout or recruited drug-naïve participants, concurrent psychotropic medication was not excluded in every study; medication exposure, head motion, cardiac and respiratory noise, scanner and preprocessing differences, and individual variation in auricular vagal innervation and ear anatomy may each affect resting-state BOLD measures. Motion exclusions were explicitly reported in some cohorts but not uniformly, and physiological-noise regression varied across pipelines (57, 58). The evidence base was also restricted to magnetic resonance neuroimaging, so it does not cover electroencephalographic or other modalities. These factors do not invalidate the cohort-level patterns, but they limit precision, cross-study comparability, and the specificity of any anatomical or psychological interpretation.

These limitations define the appropriate agenda for future work. Confirmatory studies should use randomized, sham-controlled, multi-center designs with adequate sample sizes; preregistered imaging endpoints and analysis plans; harmonized stimulation reporting and acquisition parameters (, ); explicit reporting of null analyses; and independent, external validation of candidate predictive signatures (, 56). Designs should also separate acute stimulation-state effects from longitudinal treatment effects, include healthy controls scanned at both timepoints where normalization claims are intended, and test mediation rather than assuming that concurrent clinical improvement explains regional signal change. Until such evidence is available, the reviewed findings should be read as hypothesis-generating evidence of treatment-associated brain changes, not as causal pathways or clinically actionable imaging markers.

5 Conclusion

This systematic review synthesized neuroimaging evidence from 17 reports arising from 9 independent cohorts of patients with MDD (321 enrolled patients, 293 included in imaging analyses, and 144 healthy controls, with participant denominators counted once at the cohort level). The most reproducible pattern was bidirectional normalization in striatal–cortical regions anatomically consistent with the corticostriatal reward circuit, observed across four independent cohorts (, –); this pattern is qualified by null nucleus-accumbens- and putamen-seed findings in one of these cohorts (). Default-mode-network regions were involved in seven of nine cohorts but with markedly heterogeneous directions of change (, , , ), whereas anterior insula engagement that survived between-group correction was confined to acute stimulation-state and task paradigms (), and stimulation-state connectivity effects were state- and frequency-band-specific (). Explicitly null analyses were common throughout the literature, and the only machine-learning prediction model was evaluated by internal cross-validation within a single cohort ().

Together, these observations support a strictly hypothesis-generating account: auricular vagus neuromodulation is accompanied by distributed changes in circuits implicated in mood, reward, and cognitive control, but the evidence establishes neither causal pathways nor clinically actionable imaging markers. No included study used randomized allocation (, ), the credibility of sham stimulation was limited by cutaneous tingling and somatosensory sensations (), no study tested mediation, and findings from TECAS cannot be attributed to auricular vagal afferents alone (). Reported brain–symptom correlations were largely uncorrected and exploratory, and candidate predictive signatures require independent external validation before any clinical use (, 56).

Confirmatory progress will require randomized, sham-controlled, multi-center trials with adequate sample sizes and preregistered imaging endpoints; harmonized stimulation and acquisition parameters with standardized reporting (); explicit reporting of null analyses; designs that separate acute stimulation-state from longitudinal treatment effects and include healthy controls scanned at both timepoints; formal mediation analyses; and independent external validation of candidate predictive signatures in adequately powered cohorts (, 56). Until such evidence is available, neuroimaging findings in this field should guide mechanism-oriented research and trial design rather than clinical decision-making.

Statements

Data availability statement

The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding author/s.

Author contributions

YS: Conceptualization, Data curation, Methodology, Writing – review & editing. YL: Formal analysis, Writing – review & editing. JC: Supervision, Validation, Writing – review & editing. XL: Software, Visualization, Writing – review & editing. TL: Conceptualization, Funding acquisition, Methodology, Writing – original draft. WF: Conceptualization, Methodology, Project administration, Resources, Writing – review & editing.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This research was funded by the Basic Research Project of Shenzhen Science and Technology Innovation Bureau in 2025 (Natural Science Foundation, No. JCYJ20240813114631042), and by the 2024 Medical Scientific Research Project of the Zhuhai Municipal Health Commission (No. 2420009000106).

Acknowledgments

We thank Medjaden Inc. for scientific editing of this manuscript.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fpsyt.2026.1972200/full#supplementary-material

References

• 1

  GBD 2019 Mental Disorders Collaborators. Global, regional, and national burden of 12 mental disorders in 204 countries and territories, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Psychiatry. (2022) 9:137–50. doi: 10.1016/s2215-0366(21)00395-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  CiprianiAFurukawaTASalantiGChaimaniAAtkinsonLZOgawaYet al. Comparative efficacy and acceptability of 21 antidepressant drugs for the acute treatment of adults with major depressive disorder: a systematic review and network meta-analysis. Lancet (London England). (2018) 391:1357–66. doi: 10.1016/s0140-6736(17)32802-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  UER Group. Efficacy and safety of electroconvulsive therapy in depressive disorders: a systematic review and meta-analysis. Lancet (London England). (2003) 361:799–808. doi: 10.1016/S0140-6736(03)12705-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  PereraTGeorgeMSGrammerGJanicakPGPascual-LeoneAWireckiTS. The clinical TMS society consensus review and treatment recommendations for TMS therapy for major depressive disorder. Brain Stimul. (2016) 9:336–46. doi: 10.1016/j.brs.2016.03.010

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  RushAJTrivediMHWisniewskiSRNierenbergAAStewartJWWardenDet al. Acute and longer-term outcomes in depressed outpatients requiring one or several treatment steps: a STAR*D report. Am J Psychiatry. (2006) 163:1905–17. doi: 10.1176/ajp.2006.163.11.1905

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  RongPLiuJWangLLiuRFangJZhaoJet al. Effect of transcutaneous auricular vagus nerve stimulation on major depressive disorder: A nonrandomized controlled pilot study. J Affect Disord. (2016) 195:172–9. doi: 10.1016/j.jad.2016.02.031

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  WuCLiuPFuHChenWCuiSLuLet al. Transcutaneous auricular vagus nerve stimulation in treating major depressive disorder: A systematic review and meta-analysis. Med (Baltimore). (2018) 97:e13845. doi: 10.1097/MD.0000000000013845

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  BuYLiangAHoffmanBUSchiehserDMCaseOSimmonsAet al. A review of vagus nerve stimulation for disease: Comprehensive theory and evidence for mechanisms of action. Compr Physiol. (2026) 16:e70109. doi: 10.1002/cph4.70109

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  PeukerETFillerTJ. The nerve supply of the human auricle. Clin Anat (New York NY). (2002) 15:35–7. doi: 10.1002/ca.1089

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  FarmerADStrzelczykAFinisguerraAGourineAVGharabaghiAHasanAet al. International consensus based review and recommendations for minimum reporting standards in research on transcutaneous vagus nerve stimulation (Version 2020). Front Hum Neurosci. (2020) 14:568051. doi: 10.3389/fnhum.2020.568051

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  FoxMD. Mapping symptoms to brain networks with the human connectome. N Engl J Med. (2018) 379:2237–45. doi: 10.1056/NEJMra1706158

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  KaiserRHAndrews-HannaJRWagerTDPizzagalliDA. Large-scale network dysfunction in major depressive disorder: A meta-analysis of resting-state functional connectivity. JAMA Psychiatry. (2015) 72:603–11. doi: 10.1001/jamapsychiatry.2015.0071

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  Whitfield-GabrieliSFordJM. Default mode network activity and connectivity in psychopathology. Annu Rev Clin Psychol. (2012) 8:49–76. doi: 10.1146/annurev-clinpsy-032511-143049

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  MenonV. Large-scale brain networks and psychopathology: a unifying triple network model. Trends Cognit Sci. (2011) 15:483–506. doi: 10.1016/j.tics.2011.08.003

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  NestlerEJCarlezonWAJr.The mesolimbic dopamine reward circuit in depression. Biol Psychiatry. (2006) 59:1151–9. doi: 10.1016/j.biopsych.2005.09.018

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  FoxMDSnyderAZVincentJLCorbettaMVan EssenDCRaichleME. The human brain is intrinsically organized into dynamic, anticorrelated functional networks. PNAS. (2005) 102:9673–8. doi: 10.1073/pnas.0504136102

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  HaberSNKnutsonB. The reward circuit: linking primate anatomy and human imaging. Neuropsychopharmacology. (2010) 35:4–26. doi: 10.1038/npp.2009.129

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  SupekarKCaiWKrishnadasRPalaniyappanLMenonV. Dysregulated brain dynamics in a triple-network saliency model of schizophrenia and its relation to psychosis. Biol Psychiatry. (2019) 85:60–9. doi: 10.1016/j.biopsych.2018.07.020

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  PoldrackRA. The future of fMRI in cognitive neuroscience. Neuroimage. (2012) 62:1216–20. doi: 10.1016/j.neuroimage.2011.08.007

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  ZangYFHeYZhuCZCaoQJSuiMQLiangMet al. Altered baseline brain activity in children with ADHD revealed by resting-state functional MRI. Brain Dev. (2007) 29:83–91. doi: 10.1016/j.braindev.2006.07.002

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  ZangYJiangTLuYHeYTianL. Regional homogeneity approach to fMRI data analysis. Neuroimage. (2004) 22:394–400. doi: 10.1016/j.neuroimage.2003.12.030

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  TomasiDVolkowND. Functional connectivity density mapping. PNAS. (2010) 107:9885–90. doi: 10.1073/pnas.1001414107

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  MarekSTervo-ClemmensBCalabroFJMontezDFKayBPHatoumASet al. Reproducible brain-wide association studies require thousands of individuals. Nature. (2022) 603:654–60. doi: 10.1038/s41586-022-04492-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  PageMJMcKenzieJEBossuytPMBoutronIHoffmannTCMulrowCDet al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. (2021) 372:n71. doi: 10.1136/bmj.n71

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  BarkerTHHabibiNAromatarisEStoneJCLeonardi-BeeJSearsKet al. The revised JBI critical appraisal tool for the assessment of risk of bias for quasi-experimental studies. JBI Evidence Synthesis. (2024) 22:378–88. doi: 10.11124/JBIES-23-00268

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  RollsETHuangCCLinCPFengJJoliotM. Automated anatomical labelling atlas 3. Neuroimage. (2020) 206:116189. doi: 10.1016/j.neuroimage.2019.116189

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  MakrisNGoldsteinJMKennedyDHodgeSMCavinessVSFaraoneSVet al. Decreased volume of left and total anterior insular lobule in schizophrenia. Schizophr Res. (2006) 83:155–71. doi: 10.1016/j.schres.2005.11.020

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  FangJRongPHongYFanYLiuJWangHet al. Transcutaneous vagus nerve stimulation modulates default mode network in major depressive disorder. Biol Psychiatry. (2016) 79:266–73. doi: 10.1016/j.biopsych.2015.03.025

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  LiuJFangJWangZRongPHongYFanYet al. Transcutaneous vagus nerve stimulation modulates amygdala functional connectivity in patients with depression. J Affect Disord. (2016) 205:319–26. doi: 10.1016/j.jad.2016.08.003

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  FangJEgorovaNRongPLiuJHongYFanYet al. Early cortical biomarkers of longitudinal transcutaneous vagus nerve stimulation treatment success in depression. NeuroImage Clin. (2017) 14:105–11. doi: 10.1016/j.nicl.2016.12.016

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  TuYFangJCaoJWangZParkJJorgensonKet al. A distinct biomarker of continuous transcutaneous vagus nerve stimulation treatment in major depressive disorder. Brain Stimul. (2018) 11:501–8. doi: 10.1016/j.brs.2018.01.006

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  WangZFangJLiuJRongPJorgensonKParkJet al. Frequency-dependent functional connectivity of the nucleus accumbens during continuous transcutaneous vagus nerve stimulation in major depressive disorder. J Psychiatr Res. (2018) 102:123–31. doi: 10.1016/j.jpsychires.2017.12.018

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  MaYWangZHeJSunJGuoCDuZet al. Transcutaneous auricular vagus nerve immediate stimulation treatment for treatment-resistant depression: A functional magnetic resonance imaging study. Front Neurol. (2022) 13:931838. doi: 10.3389/fneur.2022.931838

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  SunJGuoCMaYGaoSLuoYChenQet al. Immediate modulatory effects of transcutaneous auricular vagus nerve stimulation on the resting state of major depressive disorder. J Affect Disord. (2023) 325:513–21. doi: 10.1016/j.jad.2023.01.035

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  SunJSunKChenLLiXXuKGuoCet al. A predictive study of the efficacy of transcutaneous auricular vagus nerve stimulation in the treatment of major depressive disorder: An fMRI-based machine learning analysis. Asian J Psychiatr. (2024) 98:104079. doi: 10.1016/j.ajp.2024.104079

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  LiaoLZhangLLvJLiuYFangJRongPet al. Transcutaneous electrical cranial-auricular acupoint stimulation modulating the brain functional connectivity of mild-to-moderate major depressive disorder: An fMRI study based on independent component analysis. Brain Sci. (2023) 13:274. doi: 10.3390/brainsci13020274

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  MaYHeJLuXSunJGuoCLuoYet al. Transcutaneous electrical cranial-auricular acupoint stimulation versus escitalopram for modulating the brain activity in mild to moderate major depressive disorder: An fMRI study. Neurosci Lett. (2023) 814:137414. doi: 10.1016/j.neulet.2023.137414

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  MaYXuPGuoCLLuoYGaoSSSunJFet al. Modulating the default mode network: Antidepressant efficacy of transcutaneous electrical cranial-auricular acupoints stimulation targeting the insula. Psychiatry Res Neuroimaging. (2024) 339:111787. doi: 10.1016/j.pscychresns.2024.111787

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  ChenLMSunJFGuoCLLiXJWangZYangHet al. Preliminary single-arm study of brain effects during transcutaneous auricular vagus nerve stimulation treatment of recurrent depression by resting-state functional magnetic resonance imaging. J Traditional Chin Med. (2022) 42:818–24. doi: 10.19852/j.cnki.jtcm.2022.05.010

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  HeJKLiSYWangYZhaoBXiaoXHouXBet al. Mapping the modulating effect of transcutaneous auricular vagus nerve stimulation on voxel-based analyses in patients with first-episode major depressive disorder: a resting-state functional magnetic resonance imaging study. Braz J Psychiatry. (2023) 45:93–101. doi: 10.47626/1516-4446-2022-2788

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 41

  ZhangSHeJKZhongGLWangYZhaoYNWangLet al. Prolonged longitudinal transcutaneous auricular vagus nerve stimulation effect on striatal functional connectivity in patients with major depressive disorder. Brain Sci. (2022) 12:1730. doi: 10.3390/brainsci12121730

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 42

  YiSWangZYangWHuangCLiuPChenYet al. Neural activity changes in first-episode, drug-naive patients with major depressive disorder after transcutaneous auricular vagus nerve stimulation treatment: A resting-state fMRI study. Front Neurosci. (2022) 16:1018387. doi: 10.3389/fnins.2022.1018387

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 43

  GuoZPLiaoDChenLWangCQuMLvXYet al. Transcutaneous auricular vagus nerve stimulation modulating the brain topological architecture of functional network in major depressive disorder: An fMRI study. Brain Sci. (2024) 14:945. doi: 10.3390/brainsci14090945

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 44

  GarciaRGCohenJEStanfordADGabrielAStowellJAizleyHet al. Respiratory-gated auricular vagal afferent nerve stimulation (RAVANS) modulates brain response to stress in major depression. J Psychiatr Res. (2021) 142:188–97. doi: 10.1016/j.jpsychires.2021.07.048

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 45

  BerthoudHRNeuhuberWL. Functional and chemical anatomy of the afferent vagal system. Auton Neurosci. (2000) 85:1–17. doi: 10.1016/S1566-0702(00)00215-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 46

  FrangosEEllrichJKomisarukBR. Non-invasive access to the vagus nerve central projections via electrical stimulation of the external ear: fMRI evidence in humans. Brain Stimul. (2015) 8:624–36. doi: 10.1016/j.brs.2014.11.018

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 47

  YakuninaNKimSSNamEC. Optimization of transcutaneous vagus nerve stimulation using functional MRI. Neuromodulation. (2017) 20:290–300. doi: 10.1111/ner.12541

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 48

  NeuhuberWLBerthoudHR. Functional anatomy of the vagus system - Emphasis on the somato-visceral interface. Auton Neurosci. (2021) 236:102887. doi: 10.1016/j.autneu.2021.102887

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 49

  KrahlSEClarkKB. Vagus nerve stimulation for epilepsy: A review of central mechanisms. Surg Neurol Int. (2012) 3:S255–9. doi: 10.4103/2152-7806.103015

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 50

  Andrews-HannaJRSmallwoodJSprengRN. The default network and self-generated thought: component processes, dynamic control, and clinical relevance. Ann N Y Acad Sci. (2014) 1316:29–52. doi: 10.1111/nyas.12360

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 51

  HamiltonJPFarmerMFogelmanPGotlibIH. Depressive rumination, the default-mode network, and the dark matter of clinical neuroscience. Biol Psychiatry. (2015) 78:224–30. doi: 10.1016/j.biopsych.2015.02.020

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 52

  TreadwayMTZaldDH. Reconsidering anhedonia in depression: lessons from translational neuroscience. Neurosci Biobehav Rev. (2011) 35:537–55. doi: 10.1016/j.neubiorev.2010.06.006

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 53

  SnyderHR. Major depressive disorder is associated with broad impairments on neuropsychological measures of executive function: a meta-analysis and review. Psychol Bull. (2013) 139:81–132. doi: 10.1037/a0028727

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 54

  McTeagueLMHuemerJCarreonDMJiangYEickhoffSBEtkinA. Identification of common neural circuit disruptions in cognitive control across psychiatric disorders. Am J Psychiatry. (2017) 174:676–85. doi: 10.1176/appi.ajp.2017.16040400

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 55

  LenerMSIosifescuDV. In pursuit of neuroimaging biomarkers to guide treatment selection in major depressive disorder: a review of the literature. Ann N Y Acad Sci. (2015) 1344:50–65. doi: 10.1111/nyas.12759

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 56

  CohenSEZantvoordJBWezenbergBNBocktingCLHvan WingenGA. Magnetic resonance imaging for individual prediction of treatment response in major depressive disorder: a systematic review and meta-analysis. Transl Psychiatry. (2021) 11:168. doi: 10.1038/s41398-021-01286-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 57

  PoldrackRABakerCIDurnezJGorgolewskiKJMatthewsPMMunafoMRet al. Scanning the horizon: towards transparent and reproducible neuroimaging research. Nat Rev Neurosci. (2017) 18:115–26. doi: 10.1038/nrn.2016.167

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 58

  Botvinik-NezerRWagerTD. Reproducibility in neuroimaging analysis: challenges and solutions. Biol Psychiatry Cognit Neurosci Neuroimaging. (2023) 8:780–8. doi: 10.1016/j.bpsc.2022.12.006

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

Keywords

functional connectivity, functional neuroimaging, resting-state fMRI, transcutaneous auricular vagus nerve stimulation, treatment response

Citation

Su Y, Liu Y, Chen J, Li X, Liu T and Fu W (2026) Large-scale brain network modulation by auricular vagus neuromodulation in major depressive disorder: a systematic review of neuroimaging evidence. Front. Psychiatry 17:1972200. doi: 10.3389/fpsyt.2026.1972200

Received

19 August 2026

Revised

11 September 2026

Accepted

17 September 2026

Published

30 September 2026

Volume

17 - 2026

Edited by

Marta Pecina, University of Pittsburgh, United States

Updates

Copyright

© 2026 Su, Liu, Chen, Li, Liu and Fu.

This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Wenbin Fu, fuwenbin@139.com; Tianzhu Liu, hadesfantasy012@163.com

Disclaimer

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.