Coordinate-Based Network Mapping of Brain Structure in Major Depressive Disorder in Younger and Older Adults: A Systematic Review and Meta-Analysis

Major depressive disorder (MDD) is one of the leading causes of disability worldwide (1), with a total yearly economic burden in the United States alone estimated at over $200 billion (2). MDD is associated with chronic physical symptoms (3), increased risk of suicide (4), and dementia (5). While many adult studies of MDD typically focus on younger and midlife adults, this disorder affects people throughout the lifespan (6). Late-life depression (LLD), defined as the presence of MDD in individuals older than 55, is associated with cerebrovascular disease (7) and Alzheimer’s disease (8). However, LLD can refer to people older than 55 with a history of depression (early onset) and those older than 55 who develop depression for the first time (late onset). Some have argued that LLD with late onset is a “different” disorder that should be studied separately from MDD (8, 9). An assessment of the brain circuitry in younger patients with MDD, older patients with early-onset MDD, and older patients with late-onset MDD could help to clarify how the same depressive syndrome in these three groups may be associated with shared as well as unique neuropathology.

Considerable efforts have been dedicated to uncovering the neural mechanisms underlying MDD in structural neuroimaging studies. Meta-analyses of brain structure in MDD suggest that structural abnormalities are present in the prefrontal cortex and medial temporal lobe regions such as the hippocampus, parahippocampal gyrus, and amygdala (10–12). A large-scale analysis (N=2,148 MDD patients) of a data set from the Enhancing Neuroimaging Genetics Through Meta-Analysis reported cortical thinning in the orbitofrontal cortex (OFC), anterior and posterior cingulate, insula, and temporal lobes (13). Further, smaller volumes of the OFC (and gyrus rectus), frontal lobe, basal ganglia, thalamus, and hippocampus have been reported as loci of structural differences in MDD (14). Meta-analytic evidence suggests that both MDD and LLD affect the hippocampus, amygdala, and ventral anterior cingulate, while LLD uniquely affects the visual cortex (11). Mechanistic theories propose that LLD (and late-onset LLD in particular) is characterized by fronto-executive impairments due to cerebrovascular disease (15).

An important challenge regarding structural neuroimaging findings in MDD and LLD is the large degree of heterogeneity (8, 10) attributable to differences in demographic or clinical characteristics (16), depression “biotypes” (17), or study methodology. For instance, greater volume or thickness in several regions affected in MDD (notably cingulate, orbitofrontal, medial, and dorsolateral prefrontal cortices) is associated with remission after antidepressant treatment (18). In LLD specifically, age at onset may be associated with different etiologies and may produce different results (19). Going beyond systematic review, activation likelihood estimation (ALE) (20) is the current gold-standard approach to synthesizing whole-brain neuroimaging findings, yet it did not detect convergence in voxel-based morphometry findings in MDD (10). Given this heterogeneity in findings, a novel, more sensitive coordinate-based network mapping approach may be needed to characterize the neural circuitry underlying MDD and LLD (21–23).

We carried out a systematic review and meta-analysis of the scientific literature to identify specific networks affected in major depression in younger to middle-aged adults with MDD and older adults with LLD. Using ALE, we expected clusters of shared structural differences to emerge in the ventromedial prefrontal cortex and the hippocampus in both MDD and LLD (10, 11). Using coordinate-based network mapping, we expected large-scale impairments in the frontoparietal control network (24), the dorsal attention network (25), and possibly the default mode network (26) in both MDD and LLD. We also hypothesized impairments in the visual network in LLD but not in MDD (11). Finally, we expected that different demographic and clinical characteristics, notably age at onset of LLD and use of antidepressants in MDD, would be associated with different findings, in particular greater involvement of the aforementioned networks in late-onset than in early-onset LLD.

METHODS

Search Strategy

We searched the PubMed, MEDLINE, and Web of Science bibliographic databases through October 27, 2020 (see Table S1 in the online supplement). The search strategy included terms related to or describing MDD, LLD, and cortical thickness or voxel-based morphometry in whole-brain studies, both as text words and as MeSH terms. The protocol of this systematic review was registered in PROSPERO (CRD42020187718).

Inclusion and Exclusion Criteria

We included any case-control studies that focused on whole-brain analyses of structural differences between patients fulfilling the DSM or ICD criteria for MDD and control subjects. From this group of studies, we then classified studies whose samples had a mean age of 55 years or older in the LLD group and the remaining studies in the MDD group. Notable exclusion criteria were the presence of comorbid psychotic disorders, severe physical conditions such as cancer, and neurological disorders such as stroke, Alzheimer’s disease, and Parkinson’s disease (see section 1.2 in the online supplement).

Data Extraction and Demographic Characteristics

Two authors (P.Z. and J.A.) screened the titles and abstracts to identify eligible studies that met all the inclusion criteria and none of the exclusion criteria (see section 1.3 in the online supplement).

Activation Likelihood Estimation (ALE) Meta-Analyses

Two ALE meta-analyses were run in the GingerALE software program (20; http://www.brainmap.org/ale/), using coordinate information and sample sizes to weight study contributions. Studies reporting anatomical regions rather than coordinates were eligible for lesion network mapping but not the ALE analyses. Significance was determined using 1,000 nonparametric permutations and cluster-level inference at p<0.05 and family-wise error (FWE) correction.

Coordinate-Based Network Mapping

Going beyond ALE, we used the novel coordinate-based network mapping approach (22, 23) to test the hypothesis that MDD and LLD result in brain structure differences in several networks. Coordinate-based network mapping leverages the difference maps derived from each study’s reported coordinates into a study-specific “network map” using the coordinates of significant differences between depressed patients and control subjects in a connectivity analysis. Combined investigations of voxel-based morphometry (VBM) and functional connectivity support the existence of a strong link between VBM and resting-state connectivity in case-control studies (27–29). Thus, we used resting-state functional connectivity and morphometric similarity (30) to identify connectivity networks corresponding to each study’s reported significant differences. For each study, we extracted the coordinates or anatomical regions of interest where significant differences were found and added these seeds together to generate a study-specific seed map. The resulting seed map was then used in resting-state and morphometric similarity analyses in the unrelated Human Connectome Project 1200 sample (N=428) (see section 1.5 of the online supplement).

Z-scored study networks were thresholded (t>3) and added together to assess which brain networks are affected in >60% of MDD and LLD studies. In addition, mean Z-values from the Yeo seven-network parcellation (hereafter referred to as Yeo networks) (31) were analyzed in a mixed-effect analysis of variance (ANOVA) using a “between” factor of group (MDD, LLD) and a “within” factor of Yeo network (visual, motor, dorsal attention, ventral attention, limbic, frontoparietal, default mode) as well as a “within” factor of modality (resting state, morphometric similarity; see Figure 2; see also Table S3 in the online supplement).

Specific Effects of Early Versus Late Onset in LLD

Going beyond identifying group mean networks underlying structural impairments in LLD, we explored the effects of early and late onset (see section 1.3 of the online supplement) of depression. We identified seven late-onset LLD studies, two early-onset LLD studies, and eight LLD studies that included patients with early or late onset (“mixed-onset” studies). In our statistical comparisons, we grouped the two early-onset studies and the eight mixed-onset studies and contrasted them with the seven late-onset studies. We assessed demographic and clinical differences between the late-onset and the early- and mixed-onset groups using one-way ANOVAs. We analyzed study-specific network maps derived from morphometric similarity and resting-state connectivity in Yeo space (seven networks) using a three-way ANOVA with group (late versus early or mixed onset), Yeo networks, and modality factors. Since pervasive differences were found, we report the results in Yeo space.

Effects of Antidepressant and Time Since Onset of MDD

We used generalized linear models (GLMs) in the Glasser surface space to test for the effects of age, sex, proportion treated with antidepressants, and depression severity on the morphometric similarity and functional connectivity networks affected in the MDD and LLD groups. Four sets of GLMs were fitted, one for each depression group (MDD, LLD) and modality (morphometric similarity, functional connectivity). In each set, there were 360 GLMs, one GLM for each region in the Glasser parcellation. A false discovery rate (FDR) corrected p threshold of 0.05 was used to control for multiple comparison correction in each set of GLMs. We included Newcastle-Ottawa Scale score as a covariate to control for study quality. The following linear models were used:

RESULTS

A total of 143 articles, reporting on 145 comparisons (128 MDD comparisons [depressed participants, N=6,362; control participants, N=6,953] and 17 LLD comparisons [depressed participants, N=535; control participants, N=468]), were included in the analyses (Figure 1). Of these, 102 studies were based on VBM results, and the remaining 43 studies used surface-based cortical thickness measures (conducted in FreeSurfer). The mean age was 37.2 years (SD=6.7, range=21–50) in the 128 MDD studies and 68.4 years (SD=6.1, range=55–79) in the 17 LLD studies. Participants in all 145 studies fulfilled the criteria for a DSM or ICD diagnosis of MDD, and all studies had high scores on the modified Newcastle-Ottawa Scale, which assesses the quality of nonrandomized studies. In all LLD studies, the mean score on the Mini-Mental State Examination was >24.

For both MDD and LLD, there were qualitative differences in brain structure in a heterogeneous set of brain areas. In MDD studies, ALE meta-analysis revealed two significant clusters (Figure 2E; see also Table S2 in the online supplement). These clusters robustly implicated medial temporal lobe regions bilaterally, including the parahippocampus (Z=5.2, p_FWE<0.001) and ventral anterior cingulate (Z=4.5, p_FWE<0.001, extending to the subgenual anterior cingulate cortex [ACC]), along with right striatal nuclei (putamen and caudate) and the right insular cortex and right inferior frontal gyrus (IFG) (p_FWE<0.01). Using ALE meta-analysis in LLD, we identified a significant cluster in the ACC and medial prefrontal cortex (Z=3.7, p_FWE<0.001) (Figure 2F; see also Table S2 in the online supplement).

Coordinate-based network mapping captured large-scale networks implicated in both MDD and LLD. Group mean network maps in MDD and LLD were similar (Figure 2A,C), reliably implicating the inferior frontal gyrus and parietal-opercular gyri (supramarginal, angular gyrus) along with the middle frontal gyrus, frontal pole, anterior cingulate, and visual cortices in over 60% of the studies (Table 1).

The three-way ANOVA identified a significant main effect of the Yeo networks (F=42.8, df=3, 430, p<0.001). Post hoc comparisons indicate that across MDD and LLD studies, differences between patients and control subjects were centered on the frontoparietal control, dorsal attention, and visual networks, since those networks showed significantly greater involvement than the limbic network (Figure 2B,D). Less differentiation was found with the ventral attention and default mode networks. No significant main effects of group were found (F=0.35, df=1, 143, p=0.85), and a small but significant effect of modality was found (F=4.7, df=1, 143, p=0.032). No significant two-way interactions between Yeo networks and group (MDD versus LLD, F=0.4, df=3, 430, p=0.75) or modality (morphometry and resting state) and group (F=1.0, df=1, 143, p=0.32) were found, suggesting that the involvement of the Yeo networks was the same for both MDD and LLD and that these two groups did not show different patterns in the resting-state versus morphometric similarity analyses, respectively. Further, no significant three-way interaction between Yeo networks, modality, and group was found (F=1.2, df=3, 426, p=0.29), suggesting that the pattern of Yeo network differences in the MDD and LLD groups was the same for the morphometric similarity and resting-state connectivity analyses.

A three-way ANOVA confined to the 17 LLD studies revealed a significant main effect of Yeo networks (F=14.9, df=2.5, 38, p<0.001), replicating results shown in Figure 2. This three-way ANOVA showed no significant three-way interaction (F=1.2, df=2.7, 40, p=0.33) or Yeo network–by–onset group interaction (F=1.8, df=2.5, 38, p=0.16). The ANOVA revealed a significant interaction of modality with early/mixed-onset versus late-onset group (F=4.9, df=1, 15, p=0.04) (Figure 3A). In a follow-up two-way ANOVA focusing on the morphometric similarity analysis, frontoparietal, dorsal, and ventral attention and visual networks were significantly more affected by late-onset LLD than early/mixed-onset LLD. Motor, limbic, and default mode networks were implicated at a similar level in both onset groups. The late- and mixed-onset studies did not differ in age, sex, depression severity, or use of antidepressants (data not shown; all p values for one-way ANOVAs >0.05). No main effect of group was found in the three-way ANOVA (F=1.2, df=1, 15, p=0.3) since the differences between the late- and mixed-onset groups were only found using morphometric similarity and not resting-state connectivity.

The final analysis investigating the effect of demographic and clinical factors on brain heterogeneity in younger adult MDD included 114 MDD studies, as 14 studies were excluded because of insufficient information on use of antidepressants. In MDD studies, the proportion of patients currently taking an antidepressant had a significant effect on the networks affected by structural abnormalities identified using resting-state network mapping (Figure 4C). More specifically, MDD studies that included more patients on antidepressants showed greater involvement of networks involving the dorsal and ventral ACC and dorsolateral prefrontal cortex (dlPFC) (t>2.36, df=1, 108, p_FDR<0.05). By contrast, studies that included fewer patients on antidepressants found greater differences in occipitotemporal regions (t<−2.39, df=1, 108, p_FDR<0.05). A sensitivity analysis included 34 first-episode studies and replicated the ACC findings (t>2.1, df=1, 27, uncorrected p<0.05; see section 2.3 in the online supplement), although it did not survive correction for multiple comparisons (p_FDR>0.05). When repeating the above analyses with morphometric similarity network mapping in MDD studies, no significant effects of sex, use of antidepressant, or depression severity were found in Glasser space (p_FDR>0.05).

Since an effect of the use of antidepressants was found, we assessed the effects of use of antidepressants on age, sex, depression severity, and years since the onset of depressive symptoms in MDD. Age at onset was not available for 34 MDD studies; therefore, we included this variable only for post hoc examination of clinical characteristics of different MDD subgroups. For these follow-up analyses, three groups of MDD studies were formed, according to the proportion of participants taking antidepressants (Figure 4).

Studies including more patients taking an antidepressant reported a higher mean age (Figure 4A) (F=21.4, df=2, 111, p<0.001) and a higher number of years since onset of MDD (Figure 4B) (F=10.6, df=2, 84, p<0.001). Twenty-nine of the 57 studies in which patients were not receiving antidepressants were focused on first-episode depression. In comparison, only five of 57 studies in which patients were receiving antidepressants included patients with first-episode depression. There were no significant associations between the proportion of participants taking antidepressants and depression severity (χ²=9.5, p=0.05) or sex (F=0.95, df=2, 114, p=0.39).

No significant effects of demographic, clinical, or quality control variables on networks impaired in LLD were found (morphometric similarity or resting-state networks, p_FDR>0.05 in Glasser space).

DISCUSSION

Using data from 143 articles, including a total of 14,318 participants, we extracted gray matter volume and cortical thickness in younger and older adults with MDD to identify networks common to both age groups and specific regional abnormalities linked to clinical characteristics. Coordinate-based network mapping showed structural differences in frontoparietal and dorsal attention networks when patients with MDD were compared with nonpsychiatric control subjects. In older patients with LLD, late onset was associated with more extensive structural differences compared with early or mixed onset. In younger patients with MDD, use of antidepressants and time since onset of MDD were associated with differences in the anterior cingulate cortex, dlPFC, and occipital-temporal regions. By contrast, traditional ALE analyses identified a limited set of ventral anterior cingulate regions common in both MDD and LLD studies. The inferior frontal gyrus was consistently implicated in MDD using both ALE and network mapping approaches. Using ALE, but not network mapping, striatal and hippocampal regions were implicated in MDD.

ALE Results

Using ALE, we showed that studies of MDD had significant clusters localized to the medial temporal lobe, anterior cingulate, and IFG. The MDD findings agreed with previous meta-analytic evidence (10) and our hypotheses. Unthresholded ALE results in LLD showed a similar pattern (e.g., involvement of left and right parahippocampi, insula, inferior frontal gyri, striatum, and anterior and posterior cingulate), although only the anterior cingulate and medial prefrontal cortices reached statistical significance. A previous meta-analysis (11) of gray matter in LLD using a different approach than ALE (signed-differential mapping) showed similar, significant results, including reductions in the ventral anterior cingulate cortex, parahippocampus, hippocampus, amygdala, and striatum and gray matter increases in the visual cortex. While our ALE findings are largely consistent with previous meta-analytic evidence for structural abnormalities in MDD and LLD, more sensitive methods are needed to understand the underlying heterogeneity.

Coordinate-Based Network Mapping Results

In contrast to our ALE analyses, our coordinate-based network mapping approach identified impairments in the IFG, dlPFC, frontal pole, medial and lateral parietal regions, and temporal gyri in both MDD and LLD. When viewed in the context of the Yeo networks, these regions encompassed frontoparietal control and dorsal attention networks in MDD and LLD. The striking continuity between MDD and LLD in both the functional connectivity and the morphometric similarity analyses suggests that MDD affects the same networks across the adult lifespan. The frontoparietal network acts as a flexible control hub over other networks (32). Impaired frontoparietal interactions with the amygdala and the striatum may interfere with prefrontal cognitive control over emotional and motivational processes in depression. The dorsal attention network, which was also implicated in MDD in our analysis, may underlie negative attentional biases in depression (33–35). Using coordinate-based network mapping, we also found evidence of visual and parietal network involvement in both MDD and LLD studies. Visual and parietal regions are strongly connected to the IFG, dlPFC, and posterior cingulate, which may explain the presence of visual networks in our results. Thus, our coordinate-based network mapping approach provides strong evidence for a common set of executive control and dorsal/visual attention networks implicated in both MDD and LLD, suggesting that LLD follows MDD along the continuum of the lifespan.

Greater Impairment in Late-Onset LLD and Potential Mechanisms

In LLD studies, late-onset LLD showed greater impairment (i.e., higher Z-scores for the network of interest) in frontoparietal, dorsal attention, and visual networks compared with mixed-onset LLD studies despite the relatively low number of LLD studies. No consensus exists on whether early- and late-onset LLD are associated with different symptomatic or cognitive profiles (16, 36–39). However, they may have a different etiology (36, 40–43). Patients with late-onset LLD consistently show higher levels of white matter hyperintensities (16, 44), a standard marker of cerebrovascular disease (45). In light of the vascular depression hypothesis, we speculate that the increased impairment of frontoparietal and dorsal attention circuits in late-onset LLD is driven by vascular burden (15, 19). Cardiovascular damage to prefrontal circuits in late-onset LLD may also help explain the heightened risk for developing dementia in late-onset compared with early-onset LLD (5). On the other hand, stress-related atrophy of medial temporal regions may underlie early-onset LLD (15). We did not detect a medial temporal versus frontostriatal dissociation between early and late-onset LLD, however. This may be due to the connectivity profiles inherent to the coordinate-based network mapping approach. Greater involvement of frontoparietal and dorsal attention networks in our late-onset LLD group is also consistent with greater impairments in executive function reported in studies of late-onset LLD (38, 39). Late-onset LLD studies and LLD studies in general have been limited by small sample sizes. Thus, our findings provide robust evidence supporting greater impairment of executive control circuits and greater disease burden in late-onset LLD.

Potential Effects of Antidepressant Treatment and Remission

In addition, we also found network differences in young MDD patients associated with greater antidepressant treatment and a longer interval since onset of MDD. Antidepressant treatment was associated with more involvement of the rostral ACC, dorsal ACC, and dlPFC. Studies of treatment response have shown that smaller ACC volumes are associated with worse clinical outcomes (46, 47). Similarly, several functional connectivity (48, 49) and activation (50–52) studies suggest a role for the rostral ACC in response to treatment. Higher functional connectivity between the rostral ACC and the dlPFC has been deemed to be critical to depression remission, and it is targeted in clinical trials of repeated transcranial magnetic stimulation (53, 54). Our meta-analysis suggests that structural abnormalities in the dlPFC and ACC-centered networks are more prevalent in patients treated with antidepressants, supporting the clinical importance of the dlPFC and rostral ACC as treatment targets for repeated transcranial magnetic stimulation. By contrast, younger patients who were not taking antidepressants, many of whom were experiencing their first episode of MDD, showed greater structural differences in occipitotemporal networks. Unmedicated patients with MDD have been shown to have gray matter reductions in the amygdala and parahippocampal regions compared with medicated MDD patients (55) and control subjects (56). Consistent with this pattern, we found greater occipitotemporal lobe involvement in studies of unmedicated patients with MDD compared with studies with medicated patients. Since most studies with participants taking antidepressants reported gray matter atrophy in the prefrontal cortex (see section 3.5 in the online supplement), one potential interpretation is that the more notable ACC involvement in medicated studies relates to greater disease burden, given that many participants were still actively depressed despite taking antidepressants. While we controlled for depression severity, individuals with more severe MDD symptoms are more likely to receive prescriptions for medication. Recent work by our group (57) has shown that in patients with LLD, remission in the context of antidepressant treatment is associated with limited or no differences in brain structure compared with control subjects; however, a prospective study is needed to compare those who achieve remission and those whose illness remains treatment resistant, while accounting for antidepressant treatment, to more fully disambiguate whether differences relate to antidepressant use or illness severity.

Limitations

While coordinate-based network mapping increases power and sensitivity for detecting network differences driven by study subgroups, its use in this study did not implicate striatal and hippocampal regions in MDD. Coordinate-based network mapping may introduce biases inherent to connectivity profiles, such as high connectivity of occipital and parietal areas and lower connectivity of the medial temporal regions. As in previous meta-analyses of whole-brain structural differences, we only provide a synthesis of significant findings. To maximize inclusion of study subjects, we did not consider the direction of the case-control differences in brain structure in this meta-analysis. A focus on increases and decreases may provide a more nuanced network mapping for MDD and LLD. Finally, the age cutoff for LLD is somewhat arbitrary and varies across published studies; a reanalysis of the LLD studies with a more stringent cutoff of age 60 did not substantially change our results (see section 2.2 in the online supplement).

In summary, our analysis shows how coordinate-based network mapping helps overcome the heterogeneity inherent to MDD studies and highlights the involvement of frontoparietal and dorsal attention executive control networks in both MDD and LLD. Collectively, our results provide a more comprehensive understanding of brain network abnormalities in depression across the adult lifespan.