Introduction
Due to cognitive changes, medical comorbidity, disability, and polypharmacy, treatment of late-life depression (LLD) is inherently complex (Taylor, Reference Taylor2014; Taylor, McQuoid, & Krishnan, Reference Taylor, McQuoid and Krishnan2004). Older depressed adults often do not respond as robustly to antidepressant treatment as do younger adults (Beekman et al., Reference Beekman, Geerlings, Deeg, Smit, Schoevers, De Beurs and van Tilburg2002; Tedeschini et al., Reference Tedeschini, Levkovitz, Iovieno, Ameral, Nelson and Papakostas2011) and persistent depression is associated with poorer outcomes of medical illness, impaired cognition and dementia, and high rates of suicide (Katon, Unützer, & Russo, Reference Katon, Unützer and Russo2010; Mulsant, Blumberger, Ismail, Rabheru, & Rapoport, Reference Mulsant, Blumberger, Ismail, Rabheru and Rapoport2014; Nelson, Delucchi, & Schneider, Reference Nelson, Delucchi and Schneider2013). While clinical, behavioral, and neuropsychological data provide insight on the likelihood of how patients will respond to treatment (Alexopoulos et al., Reference Alexopoulos, Kiosses, Heo, Murphy, Shanmugham and Gunning-Dixon2005; Nelson et al., Reference Nelson, Delucchi and Schneider2013; Sheline et al., Reference Sheline, Pieper, Barch, Welsh-Boehmer, McKinstry, MacFall and Doraiswamy2010), such markers are indirect measures of brain function. In contrast, brain-based measures have the potential to disentangle brain network differences that contribute to clinical heterogeneity and variability in the response to treatment (Aizenstein, Khalaf, Walker, & Andreescu, Reference Aizenstein, Khalaf, Walker and Andreescu2014).
Earlier work in LLD examined age-related structural brain changes, often focusing on the hippocampus or white matter hyperintensities (WMH), cerebrovascular-related structural abnormalities common in LLD. Both hippocampal atrophy and greater WMH severity has been associated with poorer antidepressant responses, although these findings are not always consistent across studies (Gunning-Dixon et al., Reference Gunning-Dixon, Walton, Cheng, Acuna, Klimstra, Zimmerman and Alexopoulos2010; Hsieh et al., Reference Hsieh, McQuoid, Levy, Payne, MacFall and Steffens2002; Sheline et al., Reference Sheline, Pieper, Barch, Welsh-Boehmer, McKinstry, MacFall and Doraiswamy2010; Sneed et al., Reference Sneed, Roose, Keilp, Krishnan, Alexopoulos and Sackeim2007; Taylor, Aizenstein, & Alexopoulos, Reference Taylor, Aizenstein and Alexopoulos2013a; Taylor, Kudra, Zhao, Steffens, & MacFall, Reference Taylor, Kudra, Zhao, Steffens and MacFall2014a; Taylor et al., Reference Taylor, McQuoid, Payne, Zannas, MacFall and Steffens2014b). A current hypothesis (Taylor et al., Reference Taylor, Aizenstein and Alexopoulos2013a) is that in order for cerebrovascular damage to influence treatment outcomes, WMH would need to disrupt key fiber tracts and impair connectivity between regions of canonical functional networks implicated in depression. These include the default mode network (DMN), a network associated with negativity bias and rumination (Andrews-Hanna, Reidler, Sepulcre, Poulin, & Buckner, Reference Andrews-Hanna, Reidler, Sepulcre, Poulin and Buckner2010; Buckner, Andrews-Hanna, & Schacter, Reference Buckner, Andrews-Hanna and Schacter2008) that often fails to appropriately deactivate in depressed individuals (Sheline et al., Reference Sheline, Barch, Price, Rundle, Vaishnavi, Snyder and Raichle2009) and the cognitive control network (CCN) that is involved in executive function, emotional regulation, and guiding externally directed tasks (Seeley et al., Reference Seeley, Menon, Schatzberg, Keller, Glover, Kenna and Greicius2007; Zilverstand, Parvaz, & Goldstein, Reference Zilverstand, Parvaz and Goldstein2017). Past work reports that, compared to normal elderly subjects, depressed elders exhibit altered resting-state functional connectivity across DMN regions (Alexopoulos et al., Reference Alexopoulos, Hoptman, Kanellopoulos, Murphy, Lim and Gunning2012; Gandelman et al., Reference Gandelman, Albert, Boyd, Park, Riddle, Woodward and Taylor2019) and lower functional connectivity within the CCN (Alexopoulos et al., Reference Alexopoulos, Hoptman, Kanellopoulos, Murphy, Lim and Gunning2012). The limbic network, involved in emotion processing, the emotional response, and memory (Helm et al., Reference Helm, Viol, Weiger, Tass, Grefkes, Del Monte and Schiepek2018), is a third network for consideration. Altered limbic network function, particularly hyperactivity, is associated with greater depression severity (Peluso et al., Reference Peluso, Glahn, Matsuo, Monkul, Najt, Zamarripa and Soares2009). Network functional connectivity patterns are dynamic and connectivity patterns in the DMN and CCN change with antidepressant treatment (Karim et al., Reference Karim, Andreescu, Tudorascu, Smagula, Butters, Karp and Aizenstein2017). It remains unclear whether measures of resting-state functional connectivity pre-treatment can predict antidepressant response in LLD.
Studies examining pre-treatment connectivity as a predictor of response in LLD are sparse and often limited by open-label trial designs or smaller sample sizes. However, they do support that variability in the antidepressant response is associated with functional connectivity differences in the CCN, salience network, reward network, and even sensorimotor regions (Alexopoulos et al., Reference Alexopoulos, Hoptman, Kanellopoulos, Murphy, Lim and Gunning2012; Andreescu et al., Reference Andreescu, Tudorascu, Butters, Tamburo, Patel, Price and Aizenstein2013; Karim et al., Reference Karim, Andreescu, Tudorascu, Smagula, Butters, Karp and Aizenstein2017; Steffens, Wang, & Pearlson, Reference Steffens, Wang and Pearlson2019). Randomized clinical trials in midlife major depressive disorder (MDD) support that resting-state connectivity patterns may predict antidepressant treatment response. Higher connectivity between DMN hub regions, specifically the posterior cingulate cortex (PCC) and anterior cingulate cortex (ACC)/medial prefrontal cortex (mPFC) regions, predicted remission to first-line antidepressant regimens (Goldstein-Piekarski et al., Reference Goldstein-Piekarski, Staveland, Ball, Yesavage, Korgaonkar and Williams2018). Treatment response is further associated with connectivity differences between the subgenual anterior cingulate cortex (sgACC) and prefrontal regions (Dunlop et al., Reference Dunlop, Rajendra, Craighead, Kelley, McGrath, Choi and Mayberg2017). The EMBARC study (Trivedi et al., Reference Trivedi, McGrath, Fava, Parsey, Kurian, Phillips and Weissman2016), a multisite, randomized, controlled trial, identified a number of within-network and across-network moderators related to antidepressant response (Chin Fatt et al., Reference Chin Fatt, Jha, Cooper, Fonzo, South, Grannemann and Trivedi2020). These findings included within-network DMN connectivity and cross-network CCN connectivity, while supporting an important role of limbic network connectivity with the hippocampus emerging as a key region (Chin Fatt et al., Reference Chin Fatt, Jha, Cooper, Fonzo, South, Grannemann and Trivedi2020; Trivedi et al., Reference Trivedi, McGrath, Fava, Parsey, Kurian, Phillips and Weissman2016).
This study aimed to determine whether regional resting-state functional connectivity measures obtained prior to randomization and treatment were associated with change in depression severity over a randomized, controlled trial. We hypothesized that functional connectivity in the DMN, CCN, and limbic networks would be related to clinical improvement. Based on past work (Alexopoulos et al., Reference Alexopoulos, Hoptman, Kanellopoulos, Murphy, Lim and Gunning2012; Trivedi et al., Reference Trivedi, McGrath, Fava, Parsey, Kurian, Phillips and Weissman2016), our primary hypotheses were that higher resting within-network connectivity for both the DMN and CCN would be associated with better antidepressant responses. Additionally, given recent work (Chin Fatt et al., Reference Chin Fatt, Jha, Cooper, Fonzo, South, Grannemann and Trivedi2020; Dunlop et al., Reference Dunlop, Rajendra, Craighead, Kelley, McGrath, Choi and Mayberg2017), we also tested for select cross-network relationships involving the sgACC and a possible role of the hippocampus. In primary analyses, we focused on change in depression severity during a blinded, controlled trial of escitalopram. In secondary analyses, we tested for moderating effects of regional connectivity on treatment-specific response and change in depression severity over time. In an exploratory aim, we tested for similar relationships during subsequent open-label treatment with bupropion.
Methods
Participants
Participants were recruited at Vanderbilt University Medical Center (VUMC; Nashville, TN) through outpatient referrals and response to community advertisements. Enrollment ranged from June 2015 through March 2020.
Criteria for inclusion required subjects be age 60 years or older and meet DSM-IV-TR criteria for MDD with a Montgomery-Asberg Depression Rating Scale (MADRS) (Montgomery & Asberg, Reference Montgomery and Asberg1979) score of 15 or more. Participation required a Mini-Mental State Exam (MMSE) (Folstein, Folstein, & McHugh, Reference Folstein, Folstein and McHugh1975) score of 24 or greater with no diagnosis of dementia or other neurological disorder. Exclusion criteria included: (1) other Axis 1 diagnoses, other than anxiety symptoms occurring during depressive episodes; (2) history of substance use disorder in the last 3 years; (3) history of psychosis; (4) acute suicidality; (5) acute grief; (6) MRI contraindications; (7) a failed trial of escitalopram in the current episode; (8) ECT in the last 6 months; and (9) current psychotherapy. Antidepressant medication use at study entry was not an exclusion criterion. After eligibility was confirmed, individuals taking antidepressant medication had those medications tapered and discontinued over several weeks. They were clinically assessed weekly for worsening depression, safety concerns such as emergent suicidality, or development of other adverse events. They could be withdrawn and return to clinical care if these problems developed. Participants were off antidepressant medications for at least two weeks prior to baseline assessments.
All participants provided written informed consent. The VUMC Institutional Review Board approved the study. The study was registered with ClinicalTrials.gov (NCT02332291).
Assessments
The Mini-International Neuropsychiatric Interview (Sheehan et al., Reference Sheehan, Lecrubier, Sheehan, Amorim, Janavs, Weiller and Dunbar1998) evaluated psychiatric diagnoses, with findings confirmed by a geriatric psychiatrist. Depression severity was quantified using the MADRS and medical burden was quantified with the Cumulative Illness Rating Scale (CIRS) (Miller et al., Reference Miller, Paradis, Houck, Mazumdar, Stack, Rifai and Reynolds1992). Age of initial depressive episode onset and duration of the current depressive episode was obtained by clinical interview with a geriatric psychiatrist and review of medical records. The MADRS was similarly obtained by a geriatric psychiatrist at each visit.
Study intervention and clinical visits
Participants were randomized to either escitalopram or placebo in a 2 to 1 allocation. The study statistician (HK) created a sequential predetermined assignment managed by the Vanderbilt Investigational Drug Service to assign participants to each treatment arm. As WMH severity may influence treatment outcomes (Gunning-Dixon et al., Reference Gunning-Dixon, Walton, Cheng, Acuna, Klimstra, Zimmerman and Alexopoulos2010; Sheline et al., Reference Sheline, Pieper, Barch, Welsh-Boehmer, McKinstry, MacFall and Doraiswamy2010; Taylor et al., Reference Taylor, Aizenstein and Alexopoulos2013a, Reference Taylor, Kudra, Zhao, Steffens and MacFall2014a), randomization was stratified by ‘high’ or ‘low’ WMH severity based on a median WMH volume derived from earlier datasets in LLD. The initial cutoff was a WMH volume of 3.86 mL, the median WMH volume observed on 3 T MRI in 145 depressed older adults across prior studies (Chang et al., Reference Chang, Yu, McQuoid, Messer, Taylor, Singh and Payne2011; Taylor et al., Reference Taylor, Kudra, Zhao, Steffens and MacFall2014a, Reference Taylor, Zhao, Ashley-Koch, Payne, Steffens, Krishnan and MacFall2013b). This stratification threshold was adjusted downward to 2.00mL by the end of the study based on the median WMH volume observed in the current study population. Participants, study physicians, and staff were blinded to treatment allocation.
For phase 1, study medication was started at one tablet daily (either 10 mg of over-encapsulated escitalopram or matching placebo), with the option to increase to two tablets daily as early as week 2. The decision to increase the dose was based on change in depression severity, clinical judgment, tolerability, and patient preference. Participants were assessed every two weeks, by telephone at weeks 2 and 6, and in clinic at weeks 4 and 8.
Participants who could not tolerate study medication or did not remit after 8 weeks had their phase 1 drug tapered over one week before progressing to phase 2, an 8-week open-label trial of bupropion, using the 24-h extended dose formulation. Dosage started at 150 mg daily and increased to 300 mg daily in 2–4 weeks if tolerated. Participants had the option to withdraw if they did not tolerate the 300 mg dose. They could continue on the 300 mg dose or increase to a maximum 450 mg daily as early as week 4 if they tolerated the medication and were not experiencing clinical improvement. Study assessments and depression severity scoring through MADRS followed the same protocol as phase 1.
MRI acquisition
Participants completed pre-randomization MRI at the Vanderbilt University Institute for Imaging Sciences on a research-dedicated 3.0 T Philips Achieva whole-body scanner (Philips Medical Systems, Best, the Netherlands) using body coil radiofrequency transmission and a 32-channel head coil for reception. Structural imaging included a whole-brain T1-weighted MPRAGE image with TR = 8.75 ms, TE = 4.6 ms, flip angle = 9°, and spatial resolution = 0.89 × 0.89 × 1.2 mm3 plus a FLAIR T2-weighted imaging conducted with TR = 10 000 ms, TE = 125 ms, TI = 2700 ms, flip angle = 90°, and spatial resolution = 0.7 × 0.7 × 2.0 mm3. Resting-state functional MRI was conducted with eyes open (TR = 2000 ms, echo time = 35 ms, flip angle = 77°, spatial resolution = 2.75 × 2.75 × 3.7 mm3, 35 axial slices). WMH volumes were measured on FLAIR images using the Lesion Segmentation Toolbox (Schmidt et al., Reference Schmidt, Gaser, Arsic, Buck, Forschler, Berthele and Muhlau2012) as previously described (Gandelman et al., Reference Gandelman, Albert, Boyd, Park, Riddle, Woodward and Taylor2019).
Functional MRI analyses
Resting-state functional images were preprocessed using the CONN toolbox (version 15.g) in SPM12, including realignment of the functional runs and correction for head motion, coregistration of functional and anatomical images for each participant, normalization of the anatomical and functional images to the standard MINI template, and spatial smoothing with a Gaussian filter (6 mm at full width at half maximum). Motion artifacts were further detected by applying the Artifact Detection Toolbox as implemented in CONN. We used a displacement threshold of 0.9 mm and a global signal threshold of Z = 5. To effectively mitigate the effects of head motion, denoising in CONN was conducted for white matter (five components extracted) and cerebrospinal fluid (five components extracted) signal, and realignment parameters (Muschelli et al., Reference Muschelli, Nebel, Caffo, Barber, Pekar and Mostofsky2014) with outlier volumes identified by the Artifact Detection Toolbox. We retained all participants with >5 min of scan time after excluding outlier volumes. The resulting blood oxygen level–dependent time series were band-pass filtered (0.01 to 0.1 Hz) to further reduce noise and increase sensitivity.
We selected seed regions of interest (ROIs) for primary seed-to-seed resting state functional connectivity analyses, identified from the original study hypotheses and recent literature (Chin Fatt et al., Reference Chin Fatt, Jha, Cooper, Fonzo, South, Grannemann and Trivedi2020; Trivedi et al., Reference Trivedi, McGrath, Fava, Parsey, Kurian, Phillips and Weissman2016). These cortical and subcortical ROIs focused on the DMN, CCN, and limbic networks. Using methodology adapted from the EMBARC trial (Chin Fatt et al., Reference Chin Fatt, Jha, Cooper, Fonzo, South, Grannemann and Trivedi2020; Schaefer et al., Reference Schaefer, Kong, Gordon, Laumann, Zuo, Holmes and Yeo2018), cortical ROI seeds were identified with the Yeo atlas (Yeo et al., Reference Yeo, Krienen, Sepulcre, Sabuncu, Lashkari, Hollinshead and Buckner2011). Cortical DMN ROIs included the (1) PCC, (2) mPFC, and (3) rostral/pregenual ACC (rACC). Cortical CCN ROIs included the (4) dorsal ACC (dACC) and (5) dorsolateral prefrontal cortex (DLPFC), while the limbic network included the (6) sgACC and (7) orbitofrontal cortex (OFC). Subcortical ROIs were identified using the WakeForest Anatomical Atlas and included the (8) anterior hippocampus and the (9) amygdala. Aside from the PCC and sgACC, where the seed crossed midline, other regions were measured bilaterally in separate ROIs (refer to online Supplementary Table S1 for full ROI details). Following a priori hypotheses, we generated seed-to-seed pairs for evaluation of functional connectivity (refer to online Supplementary Table S2 for all seed-to-seed pairs examined) and extracted individual-level beta values for each ROI pair of interest.
To account for individual differences in gray matter volume within each ROI, each subject was processed with FreeSurfer7 using the standard recon-all procedure. As outlined by the Yeo group (https://bit.ly/3wv0rZo), the Schaefer parcellations were projected to each subject's surface using the FreeSurfer procedure ‘mri_surf2surf’ and then transferred to labels using ‘mri_aparc2seg’. These labels were then used to calculate the volume of each ROI with ‘mri_segstats.’
Statistical analyses
Statistical analyses were conducted in R Statistical Software (version 4.0.3, https://cran.r-project.org). Summary statistics were used to characterize the participants.
First, we sought to determine what functional connectivity pairs were associated with post-treatment depression severity. These primary analyses of the initial blinded phase examined the relationship between pre-randomization resting functional connectivity and clinical improvement assessed by the final assessed MADRS score. We selected final MADRS score as the primary outcome over categorical characterizations such as remission or response to preserve power given the relatively small number of individuals assigned to placebo who achieved those thresholds. We created a general linear model predicting final MADRS score, including all pairwise seed-to-seed connectivity measures and key covariates (baseline MADRS score, treatment assignment, time in the study, age, gender, and WMH volume). Using this approach, we had no missing data for individuals with usable fMRI data. Backward stepwise elimination was used to determine which seed-to-seed connectivity measures were most strongly associated with final MADRS score. Using the step() function implemented in R Statistical Software, an initial linear mixed model with all ROI pairs and covariates was specified. Key covariates were retained in the final model and fixed ROI effects were dropped iteratively based on improvement of Akaike information criterion (AIC) value until either (1) subsequent models no longer improved AIC or (2) a single independent variable of interest remained. All connectivity pairs and total WMH volume were scaled using variable means and standard deviations to keep all predictors comparable during backward elimination.
To account for potential regional volumetric differences that would affect study results, a composite gray matter volume measure was then added as a covariate to the final backwards elimination model. This measure derived from a principal component analysis (PCA) that accounted for composite gray matter across all regions included within that final backward elimination model. A single principal component was estimated from standardized brain region gray matter volumes using varimax rotation, and component scores were extracted as a covariate.
Secondary analyses examined whether pre-randomization resting connectivity measures were associated with treatment-specific changes in the trajectory of depression severity change over time. This approach used longitudinal mixed effects models examining MADRS score as a repeated measure and independent variables of seed connectivity, treatment assignment, and time, controlling for covariates of age, gender, and WMH volume. Initial models tested for a three-way statistical interaction between connectivity, treatment, and time. When that interaction term did not achieve statistical significance at the false positive rate less than 0.05, we removed the three-way interaction term and examined interactive effects between seed connectivity and time. These models also included a treatment by time interaction, but that was not the focus of analyses. These secondary analyses were considered as exploratory. The sample was not sufficiently powered to detect differences in the relationship between resting functional connectivity and clinical course between the treatment arms, particularly given the unequal randomization between arms. However, these would be useful hypothesis-generating data. For these reasons we did not adjust for multiple comparisons.
Finally, we conducted exploratory analyses of the subsample of individuals who progressed to the subsequent open-label bupropion phase. Due to the smaller sample size, we did not pursue the backwards elimination approach but rather used similar approaches as in the secondary analyses of the blinded trial. These analyses did not include treatment assignment as a dependent variable as all were on the same treatment.
To account for regional volumetric differences in these secondary and exploratory analyses, gray matter was added as a covariate to each mixed model with statistically significant findings. Standardized gray matter volumes (i.e. (raw regional gray matter volume – mean regional gray matter volume)/standard deviation of regional gray matter volume) from each region of the ROI-to-ROI pair were included as a covariate in final models.
These secondary and exploratory analyses were affected by missing outcome data. Time points which did not have an outcome measurement (i.e. missing total MADRS score) were excluded via listwise deletion, but subjects with outcome data at any time point were included. Mixed effect modeling accounted for missing data by calculating a maximum likelihood estimate, which produced an unbiased parameter estimate since data met the assumption of being either missing at random or missing completely at random.
Results
The study enrolled 162 depressed elders (Fig. 1), with 95 individuals completing baseline procedures and subsequent randomization. The majority of withdrawn individuals were excluded due to concerns for MRI safety based on prior surgeries or medical procedures identified after obtaining initial informed consent and before progressing to the baseline visit. Although study eligibility allowed for a MMSE score of 24 or greater, all randomized participants exhibited a score of 26 or greater, so could be considered as cognitively intact. There were no significant treatment group differences in baseline demographic data or dose equivalents (Table 1), only a treatment effect where the escitalopram cohort had significantly lower final MADRS scores. Overall, the population exhibited depression chronicity, with the mean duration of the current depressive episode approaching three years (range 15–5141 days).
CIRS, Cumulative Illness Rating Scale; MADRS, Montgomery-Asberg Depression Rating Scale; MMSE, Mini-Mental State Exam; WMH, white matter hyperintensities (in milliliters).
Continuous variables presented as mean (standard deviation), with categorical variables presented as percent (N). Continuous variables were compared between treatment arms using pooled, two-tailed tests with 88 degrees of freedom. Categorical variables were compared using a χ2 test with 1 degree of freedom.
Of the 95 participants who were randomized (Fig. 1), one participant withdrew from the study after randomization but before receiving study drug. Four participants were excluded from analyses due to motion during MRI. Of the remaining 90 participants, 59 received escitalopram and 31 received placebo. Three individuals randomized to escitalopram and 9 individuals randomized to placebo withdrew early from the blinded phase due to worsening depression or poor tolerability, with the remainder completing the blinded phase. Forty-one participants (22 from the escitalopram arm and 19 from the placebo arm) subsequently entered the open-label bupropion phase, including 2 individuals in the escitalopram arm and 4 individuals in the placebo arm who withdrew early from the blinded phase. Eleven of these individuals withdrew early and the remaining 30 participants completed the open-label phase.
Primary analyses predicting final MADRS score
We integrated all a priori regional resting-state functional connectivity pairs (online Supplementary Table S1) into a single model. After completing backwards elimination and adding the gray matter covariate to the model, the final model identified six regional resting functional connectivity pairs that were significantly associated with final MADRS score (Table 2). Regions in the CCN (left DLPFC – left dorsal ACC) and in the limbic network (right OFC – left amygdala) exhibited a positive relationship, with greater functional connectivity being associated with a higher final MADRS score. Regions in the DMN (PCC – left mPFC, PCC –sgACC, and right mPFC – sgACC) and the limbic network (right OFC – left hippocampus), exhibited a negative relationship, where greater functional connectivity was associated with lower final MADRS score.
MADRS, Montgomery-Asberg Depression Rating Scale; DLPFC, dorsolateral prefrontal cortex; dACC, dorsal anterior cingulate cortex; CCN, cognitive control network; OFC, orbitofrontal cortex; PCC, posterior cingulate cortex (bilateral); mPFC, medial prefrontal cortex; DMN, default mode network; rACC, rostral anterior cingulate cortex; sgACC, subgenual anterior cingulate cortex (bilateral).
In this general linear model examining data from the blinded trial of escitalopram and placebo, the outcome variable was final MADRS score. A positive relationship indicated that higher functional connectivity between ROI seeds was associated with a higher final MADRS score. A negative relationship indicated that higher functional connectivity between ROI seeds was associated with a lower final MADRS score.
Secondary analyses examining treatment and time effects
We observed a single significant three-way interaction between resting-state connectivity, treatment, and time (full statistical details in online Supplementary Table S2). Greater connectivity between the right mPFC and left rACC was associated with lower MADRS scores over time in the escitalopram arm, but less change in MADRS score over time for those allocated to placebo (t = −2.37, 326df, p = 0.0184; Fig. 2a).
After removing the three-way interaction term, we tested for an interactive effect between resting functional connectivity and time predicting MADRS score. We observed a significant interaction between PCC – left hippocampal connectivity and time (t = −2.07, 327df, p = 0.0388; Fig. 2b), where increased connectivity was associated with a greater decrease in MADRS score over time.
Exploratory analysis of subsequent open-label bupropion trial
In the open-label bupropion trial we observed isolated interactive effects between pre-randomization resting state functional connectivity and time on MADRS score. Within the DMN, higher connectivity between the sgACC and left mPFC was associated with lower MADRS scores over time (t = −2.16, 124, df, p = 0.0324). In parallel, connectivity between the right OFC and right amygdala was associated with higher MADRS scores over time (t = 1.53, 124, df, p = 0.0149).
Discussion
In this single-site, two-phase randomized, controlled antidepressant trial in LLD, pre-treatment resting-state functional connectivity in DMN, limbic, and CCN regions significantly predicted clinical outcomes. Beyond primary analyses associating pre-treatment regional resting connectivity measures in these networks with post-treatment depression severity, secondary analyses suggested that resting connectivity within the DMN differentially moderated response to treatment assignment and was associated with change in depression severity over time (Fig. 2). Exploratory analyses of individuals who did not respond to initial treatment and progressed to the second open-label study phase associated change in depression severity with sgACC resting state functional connectivity.
These findings are largely concordant with past work in younger adult cohorts associating pre-treatment DMN, sgACC, and hippocampal resting functional connectivity patterns with response to antidepressant medications (Chin Fatt et al., Reference Chin Fatt, Jha, Cooper, Fonzo, South, Grannemann and Trivedi2020; Dunlop, Talishinsky, & Liston, Reference Dunlop, Talishinsky and Liston2019). Past work suggests that higher resting connectivity between anterior and posterior nodes of the DMN predicts better pharmacotherapy response (Andreescu et al., Reference Andreescu, Tudorascu, Butters, Tamburo, Patel, Price and Aizenstein2013; Dunlop et al., Reference Dunlop, Talishinsky and Liston2019; Goldstein-Piekarski et al., Reference Goldstein-Piekarski, Staveland, Ball, Yesavage, Korgaonkar and Williams2018),a finding replicated in our primary analysis. Data from the EMBARC trial expanded these results, associating higher within-network DMN functional connectivity more broadly with better response to sertraline over placebo (Chin Fatt et al., Reference Chin Fatt, Jha, Cooper, Fonzo, South, Grannemann and Trivedi2020).
In primary analyses, higher pre-treatment resting connectivity between the sgACC with both anterior and posterior DMN hubs was also associated with lower post-trial depression severity (Table 2) and sgACC connectivity with additional regions was associated with change in depression severity over the subsequent open-label trial. Substantial work associates the response to antidepressant medications and cognitive behavioral therapy with both sgACC activity (Konarski et al., Reference Konarski, Kennedy, Segal, Lau, Bieling, McIntyre and Mayberg2009; Mayberg et al., Reference Mayberg, Brannan, Mahurin, Jerabek, Brickman, Tekell and Fox1997) and sgACC functional connectivity (Dunlop et al., Reference Dunlop, Rajendra, Craighead, Kelley, McGrath, Choi and Mayberg2017, Reference Dunlop, Talishinsky and Liston2019; Kozel et al., Reference Kozel, Rao, Lu, Nakonezny, Grannemann, McGregor and Trivedi2011). Our findings associating higher sgACC resting connectivity with better treatment responses are concordant with this literature, extending those findings into older adults. Analyses of the subsequent open-label bupropion trial further suggest that broader differences in sgACC connectivity may be seen in individuals who did not respond to either study trial. Given the study design, such individuals cannot clearly be described as being treatment resistant, however they may potentially benefit from pharmacological augmentation, neuromodulation or rapidly-acting antidepressants such as ketamine (Baeken, Duprat, Wu, De Raedt, & van Heeringen, Reference Baeken, Duprat, Wu, De Raedt and van Heeringen2017; Nakamura et al., Reference Nakamura, Tomita, Horikawa, Ishibashi, Uematsu, Hiraki and Uchimura2021).
Just as the hippocampus emerged as a key hub predicting antidepressant response in the EMBARC study (Chin Fatt et al., Reference Chin Fatt, Jha, Cooper, Fonzo, South, Grannemann and Trivedi2020), our findings also highlight the hippocampus. Involvement of the hippocampus may be particularly salient in LLD given past work associating LLD with smaller hippocampal volumes and hippocampal atrophy (Hsieh et al., Reference Hsieh, McQuoid, Levy, Payne, MacFall and Steffens2002; Taylor et al., Reference Taylor, McQuoid, Payne, Zannas, MacFall and Steffens2014b). The bilateral hippocampi are integral components of the DMN (Greicius, Supekar, Menon, & Dougherty, Reference Greicius, Supekar, Menon and Dougherty2009) and connectivity between the hippocampus and DMN regions such as the PCC may have functional consequences, such as contributing to episodic memory deficits (Bai et al., Reference Bai, Zhang, Watson, Yu, Shi, Yuan and Qian2009; Schott et al., Reference Schott, Wustenberg, Wimber, Fenker, Zierhut, Seidenbecher and Richardson-Klavehn2013; Sestieri, Corbetta, Romani, & Shulman, Reference Sestieri, Corbetta, Romani and Shulman2011). Such cognitive deficits are in turn associated with poor antidepressant response (Sheline et al., Reference Sheline, Pieper, Barch, Welsh-Boehmer, McKinstry, MacFall and Doraiswamy2010). Intriguingly, we previously associated a poorer antidepressant response with WMH damage to the posterior limb of the cingulum bundle (Taylor et al., Reference Taylor, Kudra, Zhao, Steffens and MacFall2014a), the fiber tract serving as the structural connection between the anterior hippocampus and PCC.
We previously reported volumetric differences in the OFC in LLD (Taylor et al., Reference Taylor, MacFall, Payne, McQuoid, Steffens, Provenzale and Krishnan2007). In this study we observed a differential effect of OFC resting functional connectivity. Higher resting connectivity with the hippocampus was associated with lower final MADRS scores, but conversely higher connectivity with the amygdala was associated with higher final MADRS scores. Limbic regions are associated both with depression (Bremner, Fani, Cheema, Ashraf, & Vaccarino, Reference Bremner, Fani, Cheema, Ashraf and Vaccarino2019; Siegle, Steinhauer, Thase, Stenger, & Carter, Reference Siegle, Steinhauer, Thase, Stenger and Carter2002) and with the physiologic response to stress (Rajmohan & Mohandas, Reference Rajmohan and Mohandas2007). As the limbic network has reciprocal excitatory and inhibitory projections (Radley, Reference Radley2012), it is possible that greater connectivity with the hippocampus may facilitate hippocampal efforts to regulate that stress response (Herman et al., Reference Herman, Figueiredo, Mueller, Ulrich-Lai, Ostrander, Choi and Cullinan2003). In contrast, higher OFC connectivity with the amygdala may challenge stress or emotional regulation, contributing to both depression and potentially decreasing the likelihood of a treatment response. This finding deserves further study, as past studies have associated a better response to antidepressant medications with higher amygdala functional connectivity with frontocingulate regions (Klimes-Dougan et al., Reference Klimes-Dougan, Westlund Schreiner, Thai, Gunlicks-Stoessel, Reigstad and Cullen2018; Vai et al., Reference Vai, Bulgarelli, Godlewska, Cowen, Benedetti and Harmer2016). This reflects broader issues in the field about challenges in understanding inconsistencies in findings across functional neuroimaging studies.
Finally, higher resting-state connectivity within the CCN was associated with poorer clinical response, or conversely, lower within-network CCN connectivity was associated with better response. This is surprising given previous findings in LLD demonstrating that lower CCN connectivity is related to persistent depression, executive dysfunction, and poor antidepressant medication response (Alexopoulos et al., Reference Alexopoulos, Hoptman, Kanellopoulos, Murphy, Lim and Gunning2012). Our finding may reflect heterogeneity in the LLD population. Some past work (Alexopoulos et al., Reference Alexopoulos, Hoptman, Kanellopoulos, Murphy, Lim and Gunning2012) has focused on executive dysfunction, which may enrich samples for CCN dysfunction. In contrast, our sample exhibited intact cognitive performance at screening and poorer performance on executive function tests were not a requirement for study entry.
Exploratory analyses examining outcomes from the open-label bupropion trial should be viewed cautiously. Sample size and multiple comparisons are an issue with these analyses, as less than half of study participants progressed to that study phase. Moreover, by definition, this approach eliminated individuals with a more ‘favorable’ network connectivity pattern who responded during the blinded trial.
A strength of this study included its rigorous clinical design as a blinded, controlled trial. However, limitations include a modest overall sample size, with the allocation resulting in a small placebo arm. The number of subjects excluded because of past surgical history due to concerns for MRI safety may have reduced study generalizability to more medical ill elders. Moreover, while antidepressant trial durations of 8 weeks are common, some individuals may need 12 weeks or longer to exhibit a clinical response. Thus our design may have classified some individuals who needed more time on medication as ‘nonresponders’. Multiple comparisons are an additional limitation in our secondary and exploratory analyses. In order to reduce the number of total comparisons, we tested a set number of a priori seed-to-seed regions. This approach negates the ability to identify connectivity patterns related to treatment response that involve regions outside our a priori seeds. However, the study was not powered to detect differences in the relationship between allocation groups in connectivity measures and clinical change. Moreover, our findings would not have survived statistical correction for multiple comparisons. Thus, even though our results are generally concordant with past work, they should be viewed cautiously as hypothesis-generating findings. Additionally, while the use of backwards elimination for primary analyses allowed a focus on a single rather than multiple models, it does carry the limitation that variables removed early in the process are not reintroduced, even if they would have been statistically significant in the final model (Chowdhury & Turin, Reference Chowdhury and Turin2020). Finally, as connectivity patterns change during antidepressant treatment (Karim et al., Reference Karim, Andreescu, Tudorascu, Smagula, Butters, Karp and Aizenstein2017), obtaining only a pre-treatment MRI precluded us from examining changes in connectivity patterns over time that may be related to recovery.
In conclusion, pre-treatment resting state functional connectivity patterns across multiple intrinsic networks are associated with the response to pharmacotherapy in older depressed adults. This advances our understanding of the neurobiological profile that characterizes an individual who will likely respond to first-line antidepressant treatment and extends it into older adults. When combined with previous work in this area (Chin Fatt et al., Reference Chin Fatt, Jha, Cooper, Fonzo, South, Grannemann and Trivedi2020; Gandelman et al., Reference Gandelman, Albert, Boyd, Park, Riddle, Woodward and Taylor2019; Goldstein-Piekarski et al., Reference Goldstein-Piekarski, Staveland, Ball, Yesavage, Korgaonkar and Williams2018; Karim et al., Reference Karim, Andreescu, Tudorascu, Smagula, Butters, Karp and Aizenstein2017), our findings support that network connectivity patterns may serve as proximal identifiers of favorable response to antidepressant treatment in complex patient populations. Given the single-site nature of the study and relatively small sample size for a clinical trial, future research should work to both replicate these observations and translate these findings into accessible clinical markers. This could allow for clinical stratification of patients into those likely to have a good response to first- or second-line pharmacotherapy, or inform the identification of a treatment-resistant phenotype who may benefit from earlier intervention with pharmacological augmentation or neuromodulation.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291722003579.
Financial support
This study was supported by National of Institute of Health grants R01 MH102246, R01 MH121620, K24 MH110598, S10 OD021771, and UL1TR002243.
Conflicts of interest
All authors (Ahmed, Boyd, Elson, Albert, Begnoche, Kang, Landman, Szymkowicz, Andrews, Vega, Taylor) report no conflicts of interest.