Functional dysconnectivity in youth depression: Systematic review, meta-analysis, and network-based integration

Youth depression has been associated with heterogenous patterns of aberrant brain connectivity. To make sense of these divergent findings, we conducted a systematic review encompassing 19 resting-state fMRI seed-to-whole-brain studies (1400 participants, comprising 795 youths with major depression and 605 matched healthy controls). We incorporated separate meta-analyses of connectivity abnormalities across the levels of the most commonly seeded brain networks (default-mode and limbic networks) and, based on recent additions to the literature, an updated meta-analysis of amygdala dysconnectivity in youth depression. Our findings indicated broad and distributed findings at an anatomical level, which could not be captured by conventional meta-analyses in terms of spatial convergence. However, we were able to parse the complexity of region-to-region dysconnectivity by considering constituent regions as components of distributed canonical brain networks. This integration revealed dysconnectivity centred on central executive, default mode, salience


Introduction
Major depression has a lifetime prevalence of 11.1-14.6%(Kessler and Bromet, 2013;Kovess-Masfety et al., 2013).Early onset depression is associated with functional impairments across academic and occupational settings (Kessler, 2012).It is therefore unsurprising that for those aged 10-24 years, depressive disorders represent the leading cause of mental health disability (Gore et al., 2011;Vos et al., 2016).Unfortunately, early-onset is also a risk factor for treatment resistance (Dold and Kasper, 2017), persistent and severe clinical course, and suicidal behaviours (Zisook et al., 2007).The high prevalence together with poor psychosocial and clinical prognosis highlight the importance of delineating the neurobiological mechanisms mediating youth depression, which may facilitate effective early intervention.
Youth depression is associated with changes in brain structure (e.g., reduced hippocampal volume and certain cortical surface areas (Schmaal et al., 2017(Schmaal et al., , 2016) ) and function.Neurobiological abnormalities in brain function may be examined using functional magnetic resonance imaging (fMRI), during task or rest.Meta-analyses of task-based fMRI studies in youth depression have identified altered activations during reward, emotion, and/or cognitive processing (Keren et al., 2018;Li and Wang, 2021;Miller et al., 2015).An increasing number of studies have also focused on changes in endogenous brain activity at rest (i.e., in the absence of any explicit task and referred to as resting-state fMRI; Raichle et al., 2001), although integrating these findings is not straightforward and warrants comprehensive review.
Resting-state fMRI (rs-fMRI) is primarily used to infer dysfunctional connectivity across distributed brain regions, and is measured based on statistical dependencies in neural activity between distinct regions.Seed-based correlation is one of the most frequently used approaches for assessing functional connectivity.This approach seeks to identify abnormalities in functional connectivity between a seed placed in an a priori region-of-interest (ROI) and the rest of the brain (Lee et al., 2013).To date, only the literature on amygdala-seeded dysfunction has been meta-analysed (Tang et al., 2018).However, a vast body of literature exists on dysfunction across other brain regions, and has yet to be synthesised.These studies examined the connectivity of other key limbic structures, including the hippocampus, anterior cingulate cortex (ACC), and the striatum.Current data appear divergent both in terms of implicated brain regions and directionality (Chattopadhyay et al., 2017;Connolly et al., 2013;Feng et al., 2022;Gabbay et al., 2013;Geng et al., 2016;Kerestes et al., 2015).Further work is therefore needed to integrate prior work and provide a comprehensive understanding of the broader extent of functional connectivity disruptions across regions other than the amygdala in youth depression.
While meta-analyses typically focus on establishing co-localisation of abnormalities at distinct anatomical loci, it is also conceivable that resting-state connectivity disruptions extend to and reflect broader abnormalities in communication between distributed brain networks (Cash et al., 2023).In adult depression, abnormal intra-and inter-network dysconnectivity anchoring primarily in the default mode (DMN), limbic/affective, and central executive networks (CEN) have been implicated (Brakowski et al., 2017;Cooney et al., 2010;Hamilton et al., 2015;Kaiser et al., 2015;Li et al., 2018Li et al., , 2022)).While the majority of research in youth depression has employed an ROI-based approach, it is possible to establish whether these findings implicate distributed brain networks using "network-based mapping" methodology.This involves mapping anatomical regions implicated across prior work to the functional brain network within which they reside.This approach also permits concurrent analysis of multiple ROIs, enabling commonalities between studies to be detected with greater statistical power, and is gaining popularity in meta-analytical literature (Kaiser et al., 2015;O'Neill et al., 2019;Taylor et al., 2023;Young et al., 2023;Zhukovsky et al., 2021).
The present work encompassed extant seed-to-whole-brain rs-fMRI studies in youth depression.Our systematic review set out to (i) identity individual brain regions that were most frequently implicated in hyperand hypo-connectivity (after systematically assigning each reported peak coordinate to an atlas-based parcellation); (ii) leverage networkbased mapping for the first time in youth depression literature to identify dysconnectivity in functional networks, and (iii) summarise findings of brain regions associated with the severity of depression to explore clinical relevance.We additionally performed network-based meta-analyses of findings grouped by seed networks to quantitatively determine convergence of findings at the level of canonical brain networks.Due to a flurry of recently published papers, we also conducted an updated meta-analysis of amygdala seed-based findings, building on seminal work by Tang and colleagues (2018).

Search strategies
Study selection was conducted in accordance with the PRISMA 2020 guidelines.This review was registered with PROSPERO (ID: CRD42022379640).PubMed/Medline, PsychINFO, and Embase databases were systematically searched to identify journal articles that explored resting-state functional connectivity disruptions in youth depression, published until December 3, 2022.
The following search strategy was used: OR "MDD" [Title/Abstract]), with human research, peer-reviewed, and English language as the filters.

Eligibility criteria
Inclusion criteria include original research articles that (i) were published in a peer-review journal; (ii) were published in English; (iii) used resting-state fMRI; (iv) employed seed-to-whole brain analytical approach as primary or secondary analysis; (v) included a clinical diagnosis of major depression based on standardised diagnostic criteria (e.g., DSM); (vi) involved participants above the age of 11 years and a mean + standard deviation age of < 25 years; and (vii) included comparison with a healthy control group.Two studies that included a partially overlapping sample but used distinct seed regions were considered as two separate studies (Table 1).

Data collation, coordinate identification and conversion
The following information was extracted systematically from each selected study where available: (i) name of first author; (ii) year of publication, (iii) country, (iv) patient and healthy control sample sizes, (v) participant demographic characteristics (i.e., age, education years, and sex), (vi) patient clinical characteristics (i.e., illness duration, illness severity, the diagnostic and symptom measures used, percentage of other comorbid diagnosis, and percentage of first-episode and/or medication-naïve patients where available); (vi) scanning protocol details (i.e., eyes open or closed, scanning duration, repetition time, and voxel size); (vii) neuroimaging analysis methodology details (i.e., size of seeds, the inclusion of strategies to control for physiological and/or head motion variables, spatial smoothing kernel size, pre-processing and processing pipeline, and method for multiple comparison correction); (viii) seed coordinates or the brain atlas that the ROI was derived from; and (ix) peak coordinates of each significant between-group effect, and the associated t/z and p values.For studies where peak effect coordinate information was not available, the corresponding author was contacted.
The seed and peak coordinates were converted to Montreal Neurological Institute (MNI) space if Talairach coordinates were reported, using the meta-analytical tool, GingerALE version 3.0.2.For ROIs extracted from a standard brain atlas, the centre of mass for the relevant parcellation was computed to obtain the seed coordinate.

Systematic review
We performed a comprehensive review of sample and methodological characteristics of the surveyed literature.We then tested whether functional dysconnectivity co-located to distinct regions or canonical brain networks (Fig. 1a and c).Lastly, brain regions implicated in depressive symptom severity were reviewed.

Cohort and methodological characteristics of the surveyed literature
Cohort characteristics and neuroimaging methodologies were first reviewed to examine potential sources of variation between studies.

2.2.2.
Seed-to-peak regional functional connectivity disruptions 2.2.2.1.Peak brain region assignment.Each of the peak coordinates was assigned to one region of the Glasser (surface-based) or Melbourne subcortex atlases (Fig. 1a; Glasser et al., 2016;Tian et al., 2020).The Glasser atlas was chosen as it is a multimodal parcellation based on cortical architecture, functional activation and connectivity, and/or topography, with high anatomical precision involving a detailed delineation of 180 areas per hemisphere.

Vote counting.
We next looked to identify the most robust functional connectivity changes in youth depression.This involved tallying the above peak-coordinate parcellated regions, irrespective of their corresponding seed regions.Separate counts were maintained for  increased and decreased connectivity (Fig. 1a).

Correlation with depression severity
A qualitative synthesis of functional dysconnectivities previously reported as being associated with depressive symptom severity was conducted to elucidate brain regions of potentially greater clinical relevance.

Seed network-to-peak network functional connectivity disruptions
We next examined whether individual functional connectivity disruption findings may be underpinned by specific functional network alterations (Fig. 1c).
2.2.4.1.Seed and peak network assignment.Consistent with previous meta-analytical literature focused on network-based disruptions in psychiatric conditions (Kaiser et al., 2015;O'Neill et al., 2019), each seed and peak coordinate was assigned to its canonical resting-state functional network, as defined by the Yeo network parcellation (the fine 17-network solution; Fig. 1c; Buckner et al., 2011;Choi et al., 2012;Yeo et al., 2011) and are referred to as seed network and peak network respectively thereafter.Use of this parcellation thus enables direct comparison with prior meta-analyses adopting a similar coordinate-based network mapping approach.
For the small subset of subcortical seed regions that fell outside the Yeo network parcellation, a network was assigned in accordance with previous network-based adult depression meta-analysis (Kaiser et al., 2015).

Vote counting.
The frequency of abnormal connectivity between each seed and peak network was summed for decreased and increased connectivity separately (Fig. 1c).The resultant counts were then summed to obtain the absolute total values to elucidate the most common patterns of network-to-network alterations.

Network-based meta-analyses
To quantitatively elucidate convergence of findings for each network, we next performed network-based meta-analyses of findings grouped by seed networks (Fig. 1b).Peak coordinates from included studies were stratified into groups based on the network within which  the seed ROI resided.This yielded a distinct set of peak coordinates for each of the 7 resting-state networks.Separate meta-analyses were then conducted for each set of peak coordinates.Seed networks with inadequate sample sizes (i.e., employed in <10 studies) were excluded from further analysis.For studies that included more than one seed region, seed regions that fell within a different resting-state network were considered a distinct study (see Supplementary Table 1 for assignment of seed networks).

Activation Likelihood Estimation (ALE). ALE meta-analyses
were first conducted.The null hypothesis was that the spatial distribution of coordinates would be uniform across the grey matter volume.The peak effect coordinates (i.e., foci) in MNI space were imported into GingerALE version 3.0.2(this tool is accessible at https://www.brainmap.org/ale/).The foci were modelled as three-dimensional Gaussian probability distributions using full-width half-maximum (FWHM) kernel estimated based on study sample size.A tighter, taller Gaussian was used to blur the foci for larger samples due to greater statistical certainty, and vice versa.A union map (i.e., ALE image) of all study-specific probability distributions was generated and compared to a null distribution created from simulated datasets with randomly placed foci based on characteristics of the imported data (1000 permutations; Eickhoff et al., 2012).
Results were then family-wise error (FWE) corrected for multiple comparisons using cluster-level inference of p < .05 with a cluster-forming threshold of uncorrected p < .001.

Seed-based d mapping with permutation of subject images (SDM-PSI).
We further interrogated the data using SDM-PSI (version 6.21; can be accessed via https://www.sdmproject.com/),a meta-analytic approach that has the advantages of accounting for the directionality of results (i.e., increased vs decreased functional connectivity changes) and taking in effect size estimates, rather than a binary measure of significance.Accounting for these factors may ensure reproducibility of findings.

2.3.1.3.
Meta-analyses controlling for potential confounds.We repeated the above meta-analyses after excluding five studies whose cohorts might have limited the generalisability of outcomes.Of these, two studies focused on specific comorbid symptom manifestations, (i.e., disruptive behaviour (Kim et al., 2015) and suicidality (Cao et al., 2021), one study included patients in remission (Jacobs et al., 2016), and two studies (Hu et al., 2019;Yan et al., 2022) included participants above the age of 25 years (these two studies were initially included as their samples' mean + standard deviation ages fell below 25 years).
To address the potential influence of acquisition variability, we performed additional meta-analyses excluding six studies.These studies either utilised an eye-open (Connolly et al., 2017(Connolly et al., , 2013;;Gabbay et al., 2013;Jacobs et al., 2016) or music-listening (Cullen et al., 2009) protocol during resting-state scanning or lacked information on the specific protocol used (Kim et al., 2015).

Sensitivity analysis.
To objectively evaluate the robustness and reproducibility of meta-analysis findings, jack-knife sensitivity analysis was conducted.This involved repeating all network-based ALE and SDM-PSI meta-analyses discarding one study at a time to assess disproportionate influences of any single studies.
Using the SDM-PSI tool, we further conducted meta-regression analyses with age and the respective percentage of female participants in patient and healthy control group as regressors to account for the combined effects of age and sex on neurodevelopmental variability and in turn connectivity aberrancies.Next, to detect connectivity disruptions of the greatest relevance to clinical outcomes, a meta-regression by depressive severity as measured by the Hamilton Depression Rating Scale (HAM-D; Hamilton, 1960) was conducted.Conversion to HAM-D scores was performed where possible including the Beck Depression Inventory (BDI; Furukawa et al., 2019) and Montgomery-Asberg Depression Rating Scale (MADRS; Leucht et al., 2018), resulting in a total of 8 studies for both limbic (Chi et al., 2021;Connolly et al., 2013;Cullen et al., 2014Cullen et al., , 2009;;Davey et al., 2012;Gabbay et al., 2013;Jacobs et al., 2016;Kerestes et al., 2015) and default mode (Chi et al., 2021;Connolly et al., 2013;Davey et al., 2012;Gabbay et al., 2013;Jacobs et al., 2016;Kerestes et al., 2015;Pan et al., 2020;Yan et al., 2022) seed-networks.The same statistical threshold as per main analyses was used for all sensitivity and meta-regression analyses.

Supplementary amygdala-seed-based meta-analysis.
An updated meta-analysis of amygdala seed-based findings was conducted, to include publications since the seminal work by Tang and colleagues in 2018.To ensure comparability, we utilised SDM-PSI and parameters identical to that in Tang et al. (2018), including an uncorrected threshold of p < .005and peak z > 1.

Risk of bias assessments
Risk of bias assessment was conducted using the NIH Quality Assessment Tool for Observational Cohort and Cross-Sectional Studies (https://www.nhlbi.nih.gov/health-topics/study-quality-assessment-tools) to assess study methodological quality (Ma et al., 2020;Wang et al., 2006).It consists of 14 items under six quality domains, namely, exposure, outcome assessment, loss-to-follow-up, confounding, and others.Within the context of rs-fMRI studies, exposure (i.e., independent variable) was defined as a formal diagnosis of depression, and outcome as the rs-fMRI results.As such, four of 14 items pertaining to sufficient timeframe between exposure and outcome, different levels of exposure, longitudinal/repeated assessment of exposure, and blinding to exposure status, were not applicable and hence excluded from the current review.Further, the item assessing loss-to-follow-up (i.e., "was loss to follow-up after baseline 20% or less?") was modified to "was exclusion of scans/enrolled participants 20% or less?" to suit the current review.Each item was assigned a score of either 1 ("Yes"), 0.5 ("partially fulfilled"), or 0 ("No" or "Not reported"), with the modified checklist yielding a total score of 10.The quality score, as measured by the proportion of affirmative ratings (including ratings of 0.5) across the 10 items, was computed for each study.These scores were then assigned a quality rating of "poor" (<60% of affirmative ratings), "fair" (60-80%), or "good" (>80%).

Study selection
We identified 2752 articles published between June 1975 and December 2022.The abstracts of 586 articles, including full texts of 188 studies were individually screened for eligibility.Searching the reference list of the full texts yielded an additional 5 articles to be reviewed.In total, 19 studies were found eligible, comprising 1400 participants (795 youths with major depression and 605 demographically matched healthy controls; refer to Fig. 2 for reasons of exclusion, Table 1 for participant characteristics, and Table 2 for neuroimaging characteristics).

Systematic review
3.2.1.Cohort and methodological characteristics of the surveyed literature 3.2.1.1.Sample characteristics.Detailed examination revealed considerable variation between studies in cohort characteristics.More than half of the studies (n = 11) encompassed early to late adolescence (within the age range of 11-19 years), whereas the remaining 8 studies focused solely on young adulthood (e.g., from 18 years onwards with an upper age limit of 30 years; Fig. 3c and Table 1).While such wide withinsample age range may have been accounted for through the inclusion of demographic covariates during analyses in some studies (typically age and sex), three studies did not include covariates during analysis (Table 2).
Heterogeneity was similarly evident in clinical characteristics.Some studies included patients with current antidepressant use (Chattopadhyay et al., 2017;Cullen et al., 2009;Davey et al., 2012;Jacobs et al., 2016) or previous suicide attempt/ideation (Cao et al., 2021;Yan et al., 2022).Others, however, selectively focused on medication-naïve, medication-free and/or first-episode cohorts (Chi et al., 2021;Connolly et al., 2017Connolly et al., , 2013;;Feng et al., 2022;Gabbay et al., 2013;Kerestes et al., 2015;Kim et al., 2015;Pan et al., 2020;Pannekoek et al., 2014;Zhang et al., 2014).This likely represented differing illness severity and variable confounding influence of medication use both within and between past studies.Of note, whether these first-episode patients were in first treatment seeking episode or very first depressive episode was not defined in any of the studies.Moreover, the proportion of patients with at least one comorbid diagnosis (predominantly an anxiety disorder) also varied widely, ranging from 0% to almost 85% across studies (Fig. 3c).Despite this, only one study attempted to account for clinical variations of sample in their analysis (i.e., the presence of comorbid anxiety diagnosis in Davey et al., 2012; Table 2).

Choice of seed regions and methodology.
We stratified the 19 included studies based on the resting-state networks assigned to the seed regions.Most seed regions resided either within the DMN (68.4% of the studies) or the limbic network (73.7%;Fig. 3a; Supplementary Table 1).Less than half of the studies employed seeds within the salience or central executive (36.8%) network.Somatomotor (SMN) and visual networks were selected in less than 11% of the studies, and none of the seed regions fell within the dorsal attention network (DAN).Seed regions were therefore largely biased towards the DMN and limbic network.
DMN seed regions were diverse in location, including various anterior (encompassing subgenual, pregenual and supragenual ACC) and posterior cingulate areas (i.e., posterior cingulate cortex; PCC; Fig. 3b).For the limbic network, seeds were predominantly centred on the amygdala and other subdivisions of the subgenual ACC (sgACC; Fig. 3b).
A detailed summary of neuroimaging variables can be found in Table 2. Of particular importance, different pre-processing and analysis pipelines involving various software packages were employed in past studies.Considerable variation across studies was also evident in the extent of spatial smoothing (ranging from 4 mm to 8 mm), scanning duration (five to 12 min), and method used for multiple comparison correction.While most studies scanned participants with eyes closed (i.e., 13 studies), four studies adopted an eyes-open protocol and one study played music during the scan.Overall, substantial differences were present across all levels of fMRI methodologies in the surveyed literature.

Seed-to-peak regional functional connectivity disruptions
Next, peak results, following systematic standard brain atlas regional assignment, were stratified broadly into increased and decreased functional connectivity to identify the most implicated brain regions within each direction of aberrancy.
3.2.2.1.Regional hyperconnectivity.Of all 19 selected studies, increased functional connectivity largely localised within the frontal regions (40.6%; predominantly comprising the DLPFC and inferior frontal  regions), followed by visual (12.5%; mainly early and primary visual regions) and auditory (10.9%; auditory association regions) regions (Fig. 4 and Supplementary Figure 1).We found that frontal region hyperconnectivity was mainly associated with subcortical seed regions, including the ventral ACC (encompassing subgenual and pregenual ACC), striatum and amygdala.Most visual and auditory regions showed hyperconnectivity with ventral ACC seeds.At the level of seed ROI, we identified a greater representation of the ventral ACC seed regions in elevated connectivity, contributing to 50% of peak findings (Fig. 4).This was followed by the amygdala and striatum.

Regional hypoconnectivity. The distribution of peak regions
showing hypoconnectivity was largely similar to that showing hyperconnectivity.The frontal areas (37.5%) were again implicated to be the core peak regions, including the anterior cingulate and medial and dorsolateral prefrontal regions.Following these, visual (20.5%; primarily early visual areas) and parietal (14.8%; inferior and superior parietal areas) regions tended to emerge as notable peak regions (Fig. 4 and Supplementary Figure 1).
Examination of the seed regions underlying the peak findings noted similar contributions from the amygdala, ventral ACC, and striatum (Fig. 4).
Overall, there was no notable distinction between the precise patterns of hyper-and hypoconnectivity.Mixed involvement of predominantly the ventral ACC, amygdala, and striatum seed regions were featured across both hyper-and hypo-connectivity profiles.

Seed network-level functional connectivity disruptions
Studies considering seed regions in networks other than limbic and default mode networks were too few to enable meaningful summary.It should be noted that the different divisions of the sgACC were assigned to either the default mode or limbic network according to Yeo network boundaries, and therefore represented distinct findings with no overlap  3 and Supplementary Figure 2, a wide range of regions showed abnormal connectivity with the amygdala, including frontal, temporal, parietal and occipital cortices as well as the brain stem and cerebellum (Chattopadhyay et al., 2017;Connolly et al., 2017;Cullen et al., 2014;Pannekoek et al., 2014;Straub et al., 2017).However, discrepancy in the pattern of connectivity disruptions was evident, with selective increases and decreases in connectivity reported by Chattopadhyay et al. (2017) and Connolly et al. (2017), respectively, but mixed directions in the remaining studies (Cullen et al., 2014;Pannekoek et al., 2014;Straub et al., 2017).Further, widespread alterations were not always replicated, with two studies reporting null findings (Chi et al., 2021;Cullen et al., 2009) and other studies implicating selective insular (Jacobs et al., 2016), OFC (Zhang et al., 2014), or subcallosal gyrus/sgACC and parahippocampal (Kim et al., 2015) involvement.Notably, these studies comprised mixed samples of patients with (Chattopadhyay et al., 2017;Cullen et al., 2014;Straub et al., 2017) and without (Connolly et al., 2017;Pannekoek et al., 2014;Zhang et al., 2014) medication treatment, as well as patients in remission (Jacobs et al., 2016) or with disruptive behaviours (Kim et al., 2015).
Other key limbic network regions, particularly the sgACC and striatum, have been studied in a small number of studies.Connolly et al. (2013) reported elevated connectivity between the sgACC and distributed frontal, insular, parietal, temporal, occipital, and striatal (i.e., lentiform nucleus) structures in medication-naïve, first-episode patients (Supplementary Table 3).This study did not control for any demographic or clinical variables.In contrast, opposing observations of solely reduced connectivity between the sgACC and frontal, insular, temporal, and striatal (i.e., putamen) regions were reported in another study with a large proportion of the patient group on antidepressants (Cullen et al., 2009).Accounting for the effects of age, medication status, and presence of anxiety disorder, Davey et al. (2012) found only increased FC between bilateral sgACC and dorsomedial frontal cortex.Within the striatum, Gabbay et al. (2013) reported increased connectivity of the nucleus accumbens with primarily dorsomedial PFC in a predominantly first-episode and medication-naïve cohort.In contrast, Kerestes et al. (2015) did not identify any connectivity disruptions with the nucleus accumbens in medication-free patients.
3.2.4.2.DMN seed network findings.DMN seeds employed in past studies were generally located in the anterior and posterior cingulate regions (Fig. 3b).Within the ACC, abnormal connection with diverse cortical and subcortical structures has been reported.Broad convergence was evident and implicated largely temporal, occipital, and cerebellar regions in studies using sgACC seeds (Supplementary Table 4; Connolly et al., 2013;Pan et al., 2020;Straub et al., 2017).However, the direction of FC changes reported were discrepant.One study reported an absence of differences, although more than one-third of the patients in this study were receiving antidepressants (Chattopadhyay et al., 2017) in contrast to the aforementioned studies reporting significant findings involving predominantly medication-naïve and first-episode clinical samples (Connolly et al., 2013;Pan et al., 2020;Straub et al., 2017).A similar profile of inconsistent findings was identified for the limited number of studies that used PCC as the ROI.In contrast to the absence of significant findings in a study of medication-naïve patients (Pannekoek et al., 2014), Kim et al. (2015) reported decreased FC with precentral gyrus, inferior parietal lobe and insula in a medication-naïve and first-episode cohort with comorbid disruptive behaviours (Supplementary Table 4).An opposing pattern of increased connectivity limited to the middle frontal gyrus was implicated in Jacobs et al. (2016) involving a divergent sample of active and remitted major depression young patients.Despite the attempt to integrate findings by seed networks, variable brain regions (along with substantial heterogeneity in cohort characteristics) remained evident in studies probing the default mode or limbic seed network.

Network-to-network functional connectivity disruptions
We next sought to examine whether such heterogeneous patterns of altered functional connectivity may be reconciled in terms of broader disruptions at a network-to-network level.Given that only three and two seeds belonged to the SMN and visual network, respectively, and none of the studies employed seeds mapping onto DAN, these networks were not included.As such, only DMN, CEN, limbic and salience seed networks were reviewed.
Overall, we found distributed connectivity alterations centring on limbic, default mode, central executive, and salience seed networks.Specifically, limbic seed network demonstrated a relatively greater number of increased connections with the CEN and salience network (Figs.5a and 5d), along with reduced connectivity with the DMN (Figs. 5b and 5e).In terms of the DMN, hyperconnectivity with the CEN and within the DMN emerged (Figs.5a and 5d).
Upon tallying the number of times each peak network was reported as a significant finding, DMN and CEN were the most commonly implicated, followed by salience and visual networks (Fig. 5c).

Network-based meta-analyses
To quantitatively elucidate convergence in locations of functional dysconnectivity based on seed networks, we next performed networkbased meta-analyses, comprising all the findings (i.e., peak regions) implicated by all seeds contained within each Yeo-7 network.
Meta-analyses were conducted for only the DMN and limbic seed network, as the number of studies considering seeds in other networks was deemed inadequate (i.e., <10 studies) to enable robust metaanalysis of sufficient statistical power (Supplementary Table 1).The limbic network meta-analysis included 14 studies involving 829 participants (448 patients and 381 matched healthy controls; Supplementary Table 3) and 78 foci, while the DMN meta-analysis comprised 13 studies including 916 participants (533 adolescent major depression patients and 383 demographically-matched healthy controls; Supplementary Table 4) and 41 foci.

Activation likelihood estimation (ALE)
Neither limbic nor default mode seed network meta-analysis revealed any significant clusters, using the ALE.

Seed-based d mapping with permutation of subject images (SDM-PSI)
Similar to ALE, no significant clusters were detected for limbic or default mode seed network by SDM-PSI meta-analyses.

Meta-analyses controlling for potential confounds
Repeated ALE and SDM-PSI analyses with stricter criteria were conducted to account for potential confounds.Excluding the 5 studies that targeted specific comorbidities (i.e., major depression with disruptive behaviour or suicidality) or included remitted major depression patients or participants aged > 25 years, as well as the 6 studies that adopted an eye-open or music-listening resting-state acquisition protocol in separate analyses similarly did not yield any significant findings.

Sensitivity analysis
No significant clusters were identified for all limbic-and default mode-seed-network-based ALE and SDM-PSI leave-one-study-out jackknife sensitivity and meta-regression analyses, suggesting that the findings were unlikely to be biased by disproportionate influence of any single studies and between-study variation in sample age and sex distribution.Meta-regression by depressive symptom severity similarly did not detect any significant clusters.

Amygdala-seed-based meta-analysis
The amygdala-seed-based meta-analysis involved 10 studies including 648 participants (365 young patients with major depression and 283 demographically-matched healthy controls; Supplementary Table 5) and 47 foci (Supplementary Figure 2).Using a more lenient, uncorrected threshold identical to that used in previous meta-analysis (Tang et al., 2018), no significant clusters emerged between the amygdala seed and other brain regions.

Risk of bias assessment
On quality assessment, approximately half of the selected studies were evaluated to have a quality rating of "Good" (n = 10), while the remaining nine studies were deemed "Fair" (Supplementary Table 6).More specifically, almost all studies clearly defined the research objective, study population, and outcome measures; recruited participants from the same/similar populations with uniform inclusion and exclusion criteria; and measured exposure of interest (i.e., a formal diagnosis of depression) prior to measurement of outcome (i.e., rs-fMRI scanning).A majority of the studies (89.5%) controlled for the confounding influences of head motion and physiological variables in their neuroimaging analyses.However, a lower proportion of the studies reported 20% or less exclusion of scans/ enrolled participants (52.6%) or confirmed depression diagnosis by at least one medical specialist (63.2%).None of the studies provided a justification for the sample size, and 11 of the 19 studies included a clinical sample size of ≤ 32 patients with major depression (Fig. 3c), suggestive of limited statistical power.

Discussion
This work extends substantially on the only prior meta-analysis of resting-state functional connectivity disruptions in youth depression, which focused exclusively on amygdala connectivity.Here we have comprehensively synthesised the breadth of previously reported abnormalities across the levels of individual brain regions and canonical resting-state networks.The most consistently reported peak clusters were within frontal and visual systems, although the underlying seed regions and direction of change varied considerably.Meta-analyses of connectivity abnormalities for the most commonly seeded brain networks (default-mode and limbic networks) and the most common seed region (amygdala), did not indicate any significant peak clusters.However, assigning all seed and peak regions to their respective networks in a more integrated network-to-network analysis, suggested evidence of functional dysconnectivities representing inter-and intranetwork disruptions involving predominantly the limbic, default mode, central executive, and salience networks.

Review of regional connectivity disruptions in youth depression
Our systematic review highlighted convergence centring on aberrant frontal (particularly the DLPFC) and visual areas (particularly early visual regions; refer to Fig. 4), which were also implicated in depression severity (detailed in Supplementary Table 2).While the underlying ROIs were diverse, predominant involvement of the ventral ACC, amygdala, and striatal seed regions emerged.Together, the key regions implicated in our work appeared congruent with other lines of neuroimaging metaanalytical research in youth depression.Altered limbic and striatal activation are well documented in reward and/or affective processing in task-fMRI meta-analyses (Keren et al., 2018;Li and Wang, 2021;Miller et al., 2015).Abnormal striatal activation could also precede onset of depression (Keren et al., 2018) and emerged as the most robust neuroimaging biomarker for onset and increased depression symptoms in a systematic review of longitudinal studies (Toenders et al., 2019).Aberrant DLPFC connectivity has also been identified during negative valence tasks but not executive function tasks, pointing towards the unique role of frontal alterations in emotion processing (Miller et al., 2015).Abnormalities in sensory systems, such as the visual regions, also emerged as one of the most robust observations.Interestingly, these regions have also been implicated in structural and functional aberrations (resting-state and task-based fMRI) in adolescent and adult depression as well as antidepressant response.In youth depression, there is meta-analytical finding of more pronounced occipital activation abnormalities during emotional processing in depressed youth compared to depressed adults (Li and Wang, 2021).A large-scale mega-analysis of structural abnormalities in youth depression indicated reductions in cortical surface area that were confined to frontal, visual, and somatomotor regions (Schmaal et al., 2017), converging with the cortical findings of frontal and visual dysfunctional connectivity observed here.Visual regions have been implicated as a network hub (Oldham and Fornito, 2019;Tomasi and Volkow, 2011), habouring connections with key emotional processing areas including the hippocampus, striatum, amygdala, and ACC (Han et al., 2018;Lu et al., 2022).In adult depression, the cuneus and middle occipital gyrus emerged as the only significant clusters of altered resting-state spontaneous brain activity in a recent whole-brain-based meta-analysis (Yuan et al., 2022).Interestingly, the visual cortex has been associated with depression treatment efficacy across non-invasive brain stimulation (Zhijun Zhang et al., 2021) and pharmacological treatment (Wu et al., 2023) as well as as a marker of illness severity (Ray et al., 2021).It has been posited that the role of visual area in interoceptive and exteroceptive functions (Ray et al., 2021) as well as emotional facial processing essential for social functions (Yuan et al., 2022) may underpin its involvement in major depression.In task-based fMRI studies, abnormalities in visual cortex connectivity can be more directly attributed to differences in active social, emotional or attentional processing involving multiple brain systems, however the persistence of aberrant visual cortex functional connectivity at rest is more difficult to explain.One possibility, articulated in frameworks such as predictive coding and active inference, is that sensory perception is actively influenced by top-down cognitive and emotional influences (Ray et al., 2021) and that these influences are abnormal in depression.The implication would be that aberrant connectivity between visual cortex and regions such as the ventral ACC, amygdala, and striatum (Fig. 4) impacts on how adolescent individuals with depression perceive the world around them.However, this interpretation is speculative and requires additional dedicated research.

Null results of conventional meta-analyses
Established meta-analysis tools were unable to detect significant convergence in findings among studies that used seeds located within either the limbic or default mode network.Further, our supplementary meta-analysis of amygdala-seed-based findings did not detect any significant findings, in contrast to an earlier meta-analysis (Tang et al., 2018).The latter discrepancy may relate to a narrower age range of 13-18 years and lesser number of empirical studies (n = 8) available for review by Tang et al. (2018), further complicated by the inclusion of a fully remitted sample (not included in current analysis; Peters et al., 2016).Moreover, an earlier version of SDM (i.e., anisotropic effect-size SDM; AES-SDM; Radua et al., 2014) was implemented by Tang et al. (2018).Here, we utilised a more recent version of this software (i.e., SDM-PSI), which allows for better control of the FWE rate (Albajes- Eizagirre et al., 2019b;Albajes-Eizagirre and Radua, 2018) and minimises the well-established risk of spurious meta-analytical findings particularly when a small number of studies are included (Eickhoff et al., 2016;Müller et al., 2017).The null-findings were consistent across both ALE and SDM-PSI meta-analytical methodologies and robust to all leave-one-study-out jack-knife sensitivity and meta-regression analyses, strongly increasing confidence in the observed absence of significant findings for these conventional approaches.

Network-to-network synthesis
Conventional meta-analyses aim to detect spatial convergence across peak regions, but this may be compromised particularly in cases where considerable sources of heterogeneity within the population or across studies are present.It is also increasingly recognised that peak regions that appear random may actually constitute components of spatially distributed brain networks (Cash et al., 2023;Kaiser et al., 2015;Young et al., 2023;Zhukovsky et al., 2021).To this end, we conducted the first network-to-network synthesis in this population assigning each seed and peak region to canonical brain networks.This method revealed altered inter-and intra-network connectivity involving limbic, salience, central executive, and default mode networks.These findings converge with meta-analytical work in adult depression, which has similarly implicated abnormal connectivity within the DMN and between the DMN and limbic network (Cooney et al., 2010;Hamilton et al., 2015;Kaiser et al., 2015;Li et al., 2018Li et al., , 2022)).Such commonality suggest that widespread functional network aberrancies may be shared by youth and adult patients and ultimately may represent a trajectory of disruption to a set of common networks that persists into late illness course.
More importantly, our network-based findings broadly align with regions implicated in meta-analyses and systematic reviews of task-based fMRI studies in youth depression.Task-based research has highlighted altered activations of frontal (e.g., DLPFC), striatal (e.g., caudate and putamen), limbic (e.g., amygdala) regions, as well as key nodes of the salience (e.g., the insula) and default mode (e.g., rostral ACC/sgACC/ medial PFC and precuneus) networks.These differences were identified during tasks focusing on reward, emotion, cognitive and/or selfreferential processing (Butterfield et al., 2023;Keren et al., 2018;Kerestes et al., 2014;Li and Wang, 2021;Miller et al., 2015;Rakesh et al., 2020).There is overall convergent evidence supporting the robust role of limbic, central executive, salience, and default mode network involvement in youth depression across task-based and resting-state fMRI literature.
Involvement of the DMN is understood to relate to maladaptive and excessive self-focused ruminative thought patterns, while aberrant connectivity between DMN and limbic networks is interpreted as representing abnormal emotion processing (Cooney et al., 2010;Hamilton et al., 2015;Kaiser et al., 2015;Li et al., 2018Li et al., , 2022)).Abnormal connectivity of CEN with the limbic network and DMN has also been reported, and is postulated to represent aberrant top-down frontal-executive control of emotional regulation (Brakowski et al., 2017;Kaiser et al., 2015;Li et al., 2018Li et al., , 2022)).In parallel with the trajectories of neurodevelopment, the protracted development of transmodal association networks (i.e., regions that are involved in the integration of multimodal sensory representations and higher-order cognitive and socioemotional abilities) may represent an extended window of cortical plasticity and has been linked to heightened vulnerability to environmental influences (Kolk and Rakic, 2022;Sisk and Gee, 2022;Sydnor et al., 2023).Altered progression of this critical trajectory of transmodal association network development may therefore give rise to discoordination of external and internal cognitive, emotional, and attentional processes, and confer risk for emotional disturbance at a period of significant psychosocial transitions (Casey et al., 2019).

Comparison to whole-brain-based findings
The primary motivation for focusing on the synthesis of rs-fMRI dysconnectivity derived from seed-based methods in the present work is its extensive adoption in individual youth depression studies and in the meta-analytic literature of psychiatric disorders (Brandl et al., 2019;Dong et al., 2018;Gürsel et al., 2018).A disadvantage of utilising seed-based studies is that dysfunctional connectivity in non-seeded regions may be overlooked.We thus provide a brief overview of findings from studies using whole-brain-based methodologies.Utilising independent component analysis (ICA), aberrant internetwork connectivity anchoring in key DMN, CEN and/or salience network structures has been implicated across young depressed patients with and without previous suicide attempt compared to healthy controls (Cao et al., 2020;Zhang et al., 2016;Zhu et al., 2012), as well as those with non-suicidal self-injurious behaviour (Ho et al., 2021).Select studies that used amplitude of low-frequency fluctuations (Alff) and regional homogeneity (ReHo) measuring alterations in local neural activity/interaction have on the other hand implicated diverse sets of brain regions.The most consistent observation emerged for superior frontal gyrus, insula, and lingual, middle occipital, and postcentral gyrus (Hu et al., 2019;Mao et al., 2020;Zhang et al., 2014;Zijian Zhang et al., 2021).Taken together, and in line with our current findings, heterogeneity is similarly evident in youth depression rs-fMRI whole-brain-based literature, although broad convergence of sensory and higher-order association system involvement emerged.An important extension of the current work is therefore for future meta-analysis to extend beyond networks/regions traditionally implicated in depression and synthesise findings from whole-brain-based methodologies.This will aid in the reconciliation of functional dysconnectivity in youth depression in an unbiased manner.Future meta-analysis may also benefit from the inclusion of studies targeted at remission to maximise the sample size and better differentiate trait vs. state effects of youth depression.

Heterogeneity across the literature
A major source of heterogeneity across the literature derives from substantial variation across cohort characteristics (i.e., age range, comorbidity, medication status, previous history of suicide attempt/ideation, and number of depressive episodes).Additionally, we observed considerable differences in methodologies across the levels of fMRI protocols (e.g., eyes closed vs open during scan, scan duration) and preprocessing and analysis workflows (e.g., the extent of spatial smoothness, software package used).Importantly, some of which have previously been shown to produce differential patterns of results even on an identical fMRI dataset (Botvinik-Nezer et al., 2020).Depending on the effect sizes of brain-behaviour relations, the small sample sizes of past empirical studies can also have a negative influence in terms of propensity for false positives (see reviews by Ioannidis, 2018;Poldrack et al., 2017) and reproducibility (Botvinik-Nezer et al., 2020;Marek et al., 2022).Future standardised large-scale investigations combined with an impartial, whole-brain approach will be helpful in further delineating robust changes in brain connectivity in youth depression.

Conclusion
This work represents the most comprehensive systematic review and meta-analysis of rs-fMRI studies in youth depression to date.It includes the first network-based meta-analysis and network-to-network synthesis in this population.Our systematic review highlights frontal, visual, subgenual and pregenual anterior cingulate, striatum, and amygdala involvement.Neither ALE nor SDM-PSI identified significant clusters for seeds located within the limbic or default mode networks, nor for seeds in the amygdala.Our network-to-network synthesis revealed dysconnectivity in distributed intra-and inter-network alterations converging on central executive, limbic, salience and default mode networks.The variation we observed in methodological, demographic, and clinical characteristics across studies may have been an additional critical factor in the null findings across conventional meta-analyses.Future largescale studies would be helpful for reliably characterising functional connectivity dysfunction in youth depression.

Fig. 1 .
Fig.1.Schematic of the meta-analysis and systematic review methods.(a) To systematically identify the most common patterns of regional connectivity disruptions, each reported peak coordinate in MNI space was assigned to a region of the standard Glasser or Melbourne Subcortex Atlas.The parcellated regions were then tallied for increased and decreased functional connectivity separately.(b) Next, separate network-based meta-analyses of findings of seeds located within either the limbic or default mode network, determined by Yeo network parcellation (the fine 17-network resolution), were conducted using both Activation Likelihood Estimation (ALE) and Seed-based d Mapping with Permutation of Subject Images (SDM-PSI) approaches (here, we illustrated peak coordinate extraction for the default mode seed network as an example).(c) Lastly, to examine whether the wide range of region-to-region functional connectivity abnormalities may be underpinned by broader alterations in communications between brain networks, we sought to fully integrate findings in a network-to-network manner.This involved assigning each seed and peak coordinate to one of the Yeo network parcellation within which it resided to derive the seed and peak network respectively.The frequency of abnormal connections between each seed network and peak network was then summed for increased and decreased functional connectivity separately, and the associated heatmaps were generated.CEN = Central executive network; DAN = Dorsal attention network; DMN = Default mode network; FC = Functional connectivity; SMN = Somatomotor network.

Fig. 2 .
Fig. 2. PRISMA flowchart outlining the study selection procedure and the number of articles included and excluded at each stage.

Fig. 3 .
Fig. 3. (a) Percentage of seed regions residing within each of the canonical resting-state networks.(b) Locations of seed regions residing within the default mode (left) or limbic (right) network utilised in the selected studies.(c) Key study clinical and demographic characteristics including age range, proportion of female participants, percentage of patients with a history of antidepressant use, percentage of patients with a comorbid diagnosis in addition to major depression as a primary diagnosis, and clinical (red bars) and control (green bars) sample sizes.*Patients refrained from antidepressant use for at least one month before the MRI scan.* * None of the included patients had another DSM-IV axis I major psychiatric disorders.* ** With a mean + standard deviation age of less than 25 years.HC = healthy controls; MDD = major depression; N/A = Information not available; sgACC = subgenual anterior cingulate cortex.

Fig. 4 .
Fig. 4. Bar charts depicting the frequency of each brain region reported to show abnormally increased (left) and decreased (right) connectivity across the 19 selected studies, stratified by respective seed regions.The ventral anterior cingulate cortical regions (encompassing subgenual and pregenual anterior cingulate cortex), amygdala and striatum were among the most common seed regions associated with youth depression-related hyperconnectivity (increased connectivity) and hypoconnectivity (reduced connectivity).Bolded text indicates broad cortical and subcortical brain region categorisation, where length indicates summation of frequencies of its respective subregions.ACC = Anterior cingulate cortex.

Fig. 5 .
Fig. 5. Network-to-network aberrancies in young patients with major depression compared to demographically matched healthy controls.(a) Heatmap depicting the frequency of abnormally increased functional connectivity between networks across 19 included studies; (b) Heatmap showing the frequency of abnormally reduced between-network connectivity; (c) Stacked bar chart demonstrating the most commonly implicated peak networks and the associated contributing seed networks; (d) Hyperconnectivity between limbic and salience and central executive network regions, within the default mode network, and between default mode and central executive networks; and (e) Hypoconnectivity between limbic and default mode network regions.Seed and peak effect coordinates are represented in the left and right brain, respectively."Unassigned" encompasses peak coordinates that fall outside the grey matter.
N.Y.Tse et al.

Table 1
Participant demographic and clinical characteristics of the 19 selected studies.

Table 1
(continued ) Albajes-Eizagirre et al., 2019b, and associated effect sizes (e. g., reported or converted t values from z or p values) were submitted to SDM-PSI.Using the effect sizes, SDM-PSI first generated maps of lower and upper bounds of possible effect sizes for each included study by means of an anisotropic un-normalized Gaussian Kernel, such that voxels more correlated with the peak coordinate had effect-sizes similar to those of the peak.The mean map was then generated by voxel-wise calculation of the random effects mean of the dataset maps, weighted by the sample size and variance of each study, accounting for inter-study heterogeneity (for further details, seeAlbajes-Eizagirre et al., 2019b, Radua, 2018)es-Eizagirre andRadua, 2018).Default kernel size and thresholds (a 20 mm full width at half maximum anisotropic Gaussian kernel, TFCE FWE-corrected threshold of p < .05,and 1000 permutations) were used.

Table 2
Scanning protocol and neuroimaging analytical parameters of the 19 included studies.
DPARSF = Data Processing Assistant for Resting-State fMRI; FD = framewise displacement; FDR = False discovery rate; FSL = FMRIB Software Library; FWE = Familywise error; GRETNA = Graph Theoretical Network Analysis Toolbox; N/A = Information not available; Nil = No covariate was included in the analysis; REST = Resting-State fMRI Data Analysis Toolkit; RESTplus = Resting-State fMRI Data Analysis Toolkit plus; SPM = Statistical Parametric Mapping; and TFCE = Threshold-Free Cluster Enhancement.*Thereported duration of the entire functional sequence.N.Y.Tse et al.