Functional connectivity of salience and affective networks among remitted depressed patients predicts episode recurrence

Abstract

Recurrent episodes in major depressive disorder (MDD) are common but the neuroimaging features predictive of recurrence are not established. Participants in the Predictors of Remission in Depression to Individual and Combined Treatments (PReDICT) study who achieved remission after 12 weeks of treatment withcognitive behavior therapy, duloxetine, or escitalopram were prospectively monitored for up to 21 months for recurrence. Neuroimaging markers predictive of recurrence were identified from week 12 functional magnetic resonance imaging scans by analyzing whole-brain resting state functional connectivity (RSFC) using seeds for four brain networks that are altered in MDD. Neuroimaging correlates of established clinical predictors of recurrence, including the magnitude of depressive (Hamilton Depression Rating Scale), anxiety (Hamilton Anxiety Rating Scale) symptom severity at time of remission, and a comorbid anxiety disorder were examined for their similarity to the neuroimaging predictors of recurrence. Of the 344 patients randomized in PReDICT, 61 achieved remission and had usable scans for analysis, 9 of whom experienced recurrence during follow-up. Recurrence was predicted by: 1) increased RSFC between subcallosal cingulate cortex (SCC) and right anterior insula, 2) decreased RSFC between SCC and bilateral primary visual cortex, and 3) decreased RSFC between insula and bilateral caudate. Week 12 depression and anxiety scores were negatively correlated with RSFC strength between executive control and default mode networks, but they were not correlated with the three RSFC patterns predicting recurrence. We conclude that altered RSFC in SCC and anterior insula networks are prospective risk factors associated with MDD recurrence, reflecting additional sources of risk beyond clinical measures.

Introduction

Major depressive disorder (MDD) is, for most patients, a recurrent illness, though the time to subsequent depressive episodes varies from months to decades. Recurrence is a major contributor to the morbidity of MDD, with each subsequent episode increasing risk for chronicity [1] and likely increasing the risk of medical comorbidities, suicide, substance abuse, and functional impairment [2]. Improved understanding of the biological contributors to depressive episode recurrence could help identify patients at risk for recurrence and enable the development of specific recurrence-prevention interventions, which may differ from those required to achieve remission from a major depressive episode (MDE) [3, 4].

In community settings, recurrence rates of MDEs range from 35–42% across 20 or more years of follow-up [5–7]. In recurrence-prevention clinical trials, which typically employ a design with double-blind placebo-controlled discontinuation of an effective antidepressant, recurrence over 6–12 months of follow-up averages 18% in those maintained on ADM versus 37% in those switched to placebo [8]. In these trials, higher recurrence rates are observed over longer follow-up periods [9, 10]. The highest recurrence rates have been reported from studies of patients receiving naturalistic treatment in mental health specialty settings, ranging from 33–50% during one year of follow-up [11, 12] to 67–85% at 15 years of follow-up [13, 14]. Patients who are treatment naïve at the time of enrollment may be less likely to experience recurrence, with one prior study reporting recurrence at 21% at one year and 30% at two years [15].

Although several clincial risk factors for recurrence have been identified [16], the underlying biological features that demarcate risk for future episodes are poorly understood. A variety of methods have examined biological markers [17], most commonly involving neuroimaging, immunology, or hormonal approaches, with additional individual studies examining neurotrophins, neurotransmitters, oxidative stress, or metabolomic markers [18]. Functional MRI (fMRI) using resting state functional connectivity (RSFC) or task-based analyses have reported mixed results for predicting recurrence [19–22]. Variability in these fMRI predictors may be due to differences across samples’ clinical features, medication status at time of scanning, and the analytic methods applied. Altered subcallosal cingulate cortex (SCC) (also referred to as subgenual anterior cingulate cortex, sgACC) functional connectivity has been the most commonly identified marker distinguishing recurring and non-recurring patients. No studies have evaluated risk for recurrence using whole-brain RSFC analyses with seeds for each of the major brain networks implicated in the pathophysiology of MDD [23], and it is unknown whether imaging markers are associated with established clinical predictors of recurrence.

Using fMRI and clinical data from the 21-month follow-up phase for remitters in the Predictors of Remission in Depression to Individual and Combined Treatments (PReDICT) trial [24], we evaluated whether brain functional connectivity at the time of remission from MDD could identify patients who would experience recurrence versus those who remained well. Based on our previous work identifying the importance of RSFC between the SCC and anterior insula for the pathophysiology of MDD and its response to treatment [25, 26], we hypothesized that increased connectivity of the SCC with the salience network would predict subsequent recurrence. We also examined whether RSFC markers of recurrence risk correlated with clinical predictors of recurrence previously reported for this study [27].

Methods

Study overview

Previous publications have reported the PReDICT study design [24], acute treatment outcomes [25], and recurrence outcomes [27]. To summarize, PReDICT was designed to identify biological and clinical predictors of acute treatment response and long-term relapse/recurrence in adult (ages 18–65 years) outpatients with MDD. The study was conducted at two clinics associated with Emory University: 1) the Emory Mood and Anxiety Disorders Program in Atlanta, including a satellite location in Stockbridge, Georgia, and 2) a purely Spanish-speaking clinic at Grady Hospital in Atlanta [28]. Study approval was granted by the Emory Institutional Review Board and the Grady Hospital Research Oversight Committee, and all patients provided written informed consent prior to beginning study procedures.

Participants

The study enrolled participants aged 18–65 years, with a primary psychiatric diagnosis of MDD without psychotic features who were not considered at imminent risk for suicide or homicide as assessed by a study psychiatrist. Eligibility required that patients had never previously received treatment for MDD, dysthymia, or other mood disorder, with previous treatment defined as receiving a marketed antidepressant at a minimum effective dose for ≥ 4 consecutive weeks or receiving ≥ 4 sessions of an evidence-based psychotherapy for depression (i.e., CBT, behavior therapy, interpersonal therapy, or behavioral marital therapy). All psychiatric diagnoses were determined by a study psychiatrist and confirmed with the Structured Clinical Interview for DSM-IV [29]. To proceed to randomization, patients had to score ≥ 18 at the screening visit and ≥ 15 at the baseline visit on the 17-item Hamilton Depression Rating Scale (HAM-D) [30]. Key exclusion criteria included any medically significant or unstable medication condition (determined by medical history, physical exam, and laboratory testing), or a current eating disorder, obsessive compulsive disorder, or substance abuse or dependence. Complete inclusion/exclusion criteria are listed in the Supplement.

Treatments

At baseline, patients were randomized 1:1:1 to one of three acute phase (12 weeks) treatments: 1) escitalopram (ESC; 10–20 mg/d), a selective serotonin reuptake inhibitor (SSRI); 2) duloxetine (DUL; 30–60 mg/d), a serotonin-norepinephrine reuptake inhibitor (SNRI); or 3) CBT (16 one-hour individual sessions). Remission to acute phase treatment was defined as a HAM-D score of ≤ 7 at both weeks 10 and 12. Patients who remitted at the end of the 12-week acute treatment phase were eligible to enter a 21-month follow-up period during which they were assessed every three months to monitor for recurrence. Remitters to ESC or DUL were encouraged to remain on medication through the first 9 months of follow-up (i.e., 12 months after baseline), after which they could choose to maintain or discontinue their medication for the second year of follow-up (months 12–24). In the overall trial, 70% of remitters remained on medication throughout the follow-up phase and recurrence rates did not differ between those who maintained or discontinued medication [27]. Remitters to CBT were offered up to 3 booster sessions (at least one month apart) during the first 9 months of follow-up and up to 3 additional booster sessions during the second year of follow-up. In addition, one “crisis session” was also allowed during each year of follow-up. All patients continued in follow-up regardless of whether they continued their medication or attended the CBT booster sessions. The only concomitant psychiatric medications permitted during the maintenance phase were over-the-counter and prescription treatments for insomnia (eszopiclone, zolpidem, zaleplon, ramelteon).

Follow-up procedures

Assessments administered at 3-month intervals during the follow-up phase included a Longitudinal Interval Follow-up Evaluation interview [31], blinded ratings of depression and anxiety symptoms using the HAM-D and Hamilton Anxiety Rating Scale (HAM-A) [32], and a clinical visit with a study psychiatrist. Follow-up visits continued until 24 months from study baseline (at least 21 months after achieving remission) or until the patient experienced a depressive recurrence or was lost to follow-up. Participating patients received $50 per follow-up visit.

Definition of recurrence

All patients in this analysis who met the recurrence criteria did so at the 9-month follow-up visit or later, and thus fall under the ACNP Task Force definition of recurrence [12]. Patients were determined to have experienced recurrence if any of the four a priori defined thresholds were met: 1) meeting criteria for a major depressive episode based on a LIFE score of 3 or greater; 2) a HAM-D score ≥ 14 for two consecutive weeks (patients with an HAM-D ≥ 14 at a follow-up visit were asked to return the following week for an additional rating); 3) a HAM-D ≥ 14 at any follow-up visit and at which time the patient requested an immediate change in treatment; and 4) high risk of suicide, as determined by the study psychiatrist [24].

Image acquisition

MRI scans were performed during the week prior to the randomization visit and 1–5 days prior to the week 12 visit. All imaging was carried out using a 3 T Siemens TIM Trio (Siemens Medical Systems, Erlangen, Germany). Resting-state fMRI (rsfMRI) scanning was performed with patients’ eyes open for 7.4 min using the following protocols: a Z-SAGA sequence [33] to recover areas affected by susceptibility artifact; 150 measurements; 30 axial slices; voxel resolution = 3.4 × 3.4 × 4 mm; matrix = 64 × 64, repetition time/ echo time = 2950 ms/30 ms. A high-resolution anatomical T1-weighted image was also acquired using a magnetization-prepared rapid acquisition gradient-echo sequence (MPRAGE: repetition time = 2600 ms, inversion time = 900 ms, echo time = 3.02 ms; flip angle = 8°, voxel Resolution = 1 × 1 × 1 mm; number of slices = 176; matrix = 224 × 256).

Preprocessing of fMRI analysis

Preprocessing of all imaging data was conducted using Analysis of Functional NeuroImages (AFNI; http://afni.niml.nih.gov/afni/) software [34]. The standard preprocessing pipeline implemented in AFNI package, including despiking and local white matter regression approach to minimize sensitivity to motion, was used for processing RSFC [35]. The time series of rsfMRI data were despiked, and corrected for slice time acquisition differences and head motion [36]. Unusable scans were determined by head motion > 3 mm in any direction, signal drop or severe distortion artifact, as determined by visual inspection of the raw data, or head-coil artifacts, determined by visual inspection in each individual subject correlation map, all criteria that are consistent with our prior work with this dataset [26]. Mean head motion of included subjects, assessed by framewise displacement [37], did not significantly differ between the recurrence and non-recurrence groups (p = 0.51). The remaining effects of the noise signal, including head motion inferences, signals from the CSF and local white matter, were removed [38]. Subsequently, data were band-pass filtered (0.01 < f < 0.1 Hz) and spatially smoothed up to 8 mm full-width at half-maximum (FWHM) using 3dBlurToFWHM. The functional images were aligned to the corresponding T1-weighted anatomical image (with visual confirmation) and normalized to standard Montreal Neurological Institute (MNI) 1-mm template.

Functional connectivity analysis

A seed-based correlation approach was used to assess the RSFC of four brain networks implicated in MDD [23] and previously evaluated for change during acute treatment in this sample [26]. Bilateral spheres of 5 mm radius were used as seeds for each network: 1) default mode network (DMN), using a posterior cingulate cortex (PCC) seed [39]; 2) executive control network (ECN, also known as the cognitive control network or fronto-parietal network), using a DLPFC seed [40]; 3) the salience network (SN) using an anterior insula seed [41]; and 4) the affective network using an SCC seed [25] (see Supplementary Fig. S1). The time series within the bilateral seeds were averaged, and Pearson’s correlation coefficient maps were created for each individual subject. The correlation maps were converted to z-scores using Fisher’s r-to-z transformation to enable statistical analysis. The z-values determined the levels of functional connectivity of each seed.

Group analysis

To determine the differences in the functional connectivity between recurrence and non-recurrence groups, a non-parametric approach using whole brain voxel-wise two sample t-tests with 10,000 permutations were performed using the scans at the end of 12 weeks of acute treatment. Our primary analysis used a significance threshold of family-wise error (FWE) corrected α < 0.05 (equivalent to uncorrected p < 0.005 with a cluster size minimum 1146 mm³). The cluster size for multiple comparison correction was estimated by simulated noise volumes assuming the spatial auto-correlation function with the 10,000 Monte-Carlo simulations using 3dClustSim [42]. The significant clusters were mapped to networks defined by Yeo and colleagues [43]. A secondary analysis enabling greater sensitivity for identifying potentially important smaller regions used a threshold of p < 0.001 with 100 mm³ minimum cluster size. For our primary analyses, we included the subjects (n = 10) who were lost to follow-up (LTFU) before experiencing a recurrence with the subjects (n = 44) who completed the follow-up phase without recurrence; this approach was conservative in that if LTFU subjects did experience an un-observed recurrence after their last study visit, they would have been erroneously classed in the non-recurrent group, which would have served to reduce imaging differences between the recurrent and non-recurrent groups. To evaluate this approach, we subsequently performed post hoc comparisons between the three groups’ (recurrent, non-recurrent completers, LTFU) RSFC strength within the three recurrence-associated patterns identified in the primary analysis of the week 12 scans (see Results).

To identify the biological signature of the clinical predictors of recurrence in this dataset, we conducted the same four seed-based whole-brain analyses to identify network regions correlated with the week 12 HAM-D and HAM-A scores. The threshold was also set at FWE-corrected α < 0.05.

Results

Of the 344 patients randomized, 94 remitted to acute treatment and entered the long-term follow-up phase. Of those, 61 (17 CBT, 24 duloxetine, 20 escitalopram) completed MRI scanning during week 12 and had imaging data of sufficient quality for the current analysis. Nine patients experienced recurrence (3 in each group) and 52 did not, including 10 patients who terminated early (mean month of last visit for drop-outs = 15.3). Demographic and clinical characteristics are presented in Table 1. Presence of a comorbid anxiety disorder and Week 12 HAM-D and HAM-A scores were all significantly higher in the patients who went on to experience recurrence. Among the medication-treated patients, four were on medication at the time of recurrence and two were not. Neither site nor treatment was associated with recurrence, so the data from both sites and all three treatments were combined in the primary analyses examining neuroimaging predictors of recurrence.

Table 1.

Clinical and demographic characteristics.

Characteristic	All Patients (N = 61)		No Recurrence (n = 52)		Recurrence (n = 9)
	M	SD	M	SD	M	SD	F	p
Age (yrs)	38.8	11.0	39.1	11.1	37.4	11.2	0.41	0.69
Age at first episode (yrs)	28.7	13.1	28.9	13.3	27.3	12.3	0.34	0.74
Current episode duration (wks)*	103.5	198.7	102.9	210.2	107.0	121.8	−0.06	0.95
CTQ total	44.6	15.9	43.3	13.9	51.9	24.0	−1.0	0.32
Baseline HAM-D	18.2	3.2	18.4	3.3	17.6	2.4	0.70	0.49
Baseline HAM-A	14.2	4.4	14.3	4.4	13.9	4.8	0.25	0.80
Week 12 HAM-D	3.0	2.3	2.7	2.2	4.9	1.9	−2.9	0.01
Week 12 HAM-A	2.8	2.4	2.5	2.4	4.4	1.9	−2.1	0.04
	n	%	N	%	n	%	X2*	p
Sex							0.40	0.53
Male	33	54.1	29	55.8	4	44.4
Female	28	45.9	23	44.2	5	55.6
Race							2.30	0.51
Black	10	16.4	10	19.2	0	0
Multiple	5	8.2	4	7.7	1	11.1
Native American	9	14.8	7	13.5	2	22.2
White	37	60.7	31	59.6	6	66.7
Ethnicity							0.13	0.72
Hispanic	11	18.0	9	17.3	2	22.2
Non-Hispanic	50	82.0	43	82.7	7	77.8
Married/Cohabitating							0.01	0.90
Yes	35	57.4	30	57.7	5	55.6
No	26	42.6	22	42.3	4	44.4
Education							0.87	0.35
< High school	3	4.9	2	3.8	1	11.1	.
≥ High school	58	95.1	50	96.2	8	88.9
Employed full-time							0.55	0.46
Yes	34	55.7	30	57.7	4	44.4
No	27	44.3	22	42.3	5	55.6
Lifetime major depressive episodes							0.66	0.72
1	32	52.5	28	53.8	4	44.4
2	9	14.8	8	15.4	1	11.1
≥ 3	20	32.8	16	30.8	4	44.4
Chronic episode (≥ 2 yrs)*							1.1	0.31
Yes	18	29.5	14	26.9	4	44.4
No	42	68.9	37	71.2	5	55.6
Lifetime substance use disorder							0.77	0.38
Yes	8	13.1	9	11.5	2	22.2
No	53	86.9	43	88.4	7	77.8
Anxiety disorder at baseline							4.3	0.04
Yes	22	36.1	16	30.8	6	66.7
No	39	63.9	36	69.2	3	33.3

^*Length of current episode not recorded for one patient.

Neuroimaging markers of recurrence

Table 2 shows the three significant RSFC patterns (depicted in Fig. 1) from the primary analysis threshold at the Week 12 (remitted) scan that predicted subsequent recurrence. Recurrence was predicted by 1) increased connectivity of the SCC seed with the right anterior insula of the Salience network and 2) reduced connectivity of the SCC seed with bilateral V1 in the Visual B network. Of the other seeded networks, the only RSFC pattern that survived correction for multiple comparisons was the finding that 3) reduced connectivty of the anterior insula seed with the bilateral dorsal caudate predicted recurrence. In the three-group comparison of recurrent, non-recurrent completers, and LTFU subjects, the LTFU group’s mean RSFC strengths in the three identified patterns did not significantly differ, and were essentially identical to, those of the non-recurrent group, indicating the identified patterns were specific to recurrence (Supplementary Fig. S2).

Table 2.

Significant connectivity patterns predicting depressive episode recurrence.

Seed (Network)	Region	Brodmann Area	Yeo Network	R/L	Peak MNI coordinates			Cluster Size (mm3)	t-Value	Effect size
					x	y	z
SCC (AN)	Anterior Insula	47,16	Salience B	R	+31	+21	−10	1775	4.12	2.10
	V1	17	Visual B	R	+6	−72	+11	1337	−3.81	−1.45
Ant.INS (SN)	Caudate			R	+13	+4	+12	1653	−3.91	−1.66

Family-wise error corrected at p < 0.05.

Fig. 1. Significant resting state connectivity patterns predicting recurrence and their change from baseline to week 12.

A Increased RSFC between SCC seeds and right anterior insula. B Decreased RSFC between SCC seeds and bilateral V1. C Decreased RSFC between anterior insula seeds and bilateral caudate. For each significant RSFC pattern (a–c): (a) Week 12 fMRI RSFC pattern; (b) Week 12 RSFC difference between recurrent and non-recurrent patients overall and by treatment arm; (c) Change from baseline to week 12 in the RSFC pattern identified as predictive of recurrence at week 12. ***p < 0.001, **0.001 < p < 0.01, *0.01 < p < 0.05. ADM Antidepressant medication, ant-INS Anterior insula, CBT Cognitive behavior therapy, SCC Subcallosal cingulate cortex.

To evaluate whether recurrence could have been predicted from the pre-treatment scan, we performed post hoc analyses using the baseline RSFC values for the three RSFC patterns that predicted recurrence in our primary analysis. As shown in the violin plots of Fig. 1, pre-treatment RSFC strength did not differ at baseline between those who did or did not experience recurrence; hence, prediction of eventual recurrence could not be predicted from the pre-treatment scan. The Week 12 scan differences predicting recurrence occurred as a result of the changes from baseline RSFCs in the patients who went on to recurrence. Specifically, the SCC-anterior insula RSFC increased significantly from baseline to Week 12 among those who experienced recurrence. In contrast, for the SCC-V1 and anterior insula–caudate the RSFC strengths significantly decreased from baseline among those who experienced recurrence. Both CBT- and medication-treated patients who experienced recurrences showed similar Week 12 scan RSFC patterns and similar changes from baseline to Week 12 compared to non-recurring patients (Fig. 1A–C, b, c).

The secondary analysis using the small cluster threshold identified significant differences in RSFC patterns at Week 12 for all four seeds, most notably for the SCC, for which recurrence was associated with connectivity differences with bilateral anterior insula, frontal regions, cerebellum, and putamen. The DLPFC seed demonstrated reduced connectivity with the hippocampus and increased connectivity with V1 regions as markers for eventual recurrence. Increased connectivity of the DMN PCC seed with regions of the supplementary motor area were also predictive of eventual recurrence (Supplementary Table 1).

Neuroimaging signatures of clinical predictors of recurrence

None of the identified RSFC patterns predictive of recurrence from our primary analysis significantly correlated with the predictors of clinical outcomes, namely HAM-D or HAM-A score at Week 12 or a diagnosis of a comorbid anxiety disorder at baseline (all p > 0.05). In our secondary analyses of the RSFC patterns associated with recurrence, the only significant association to emerge was that reduced RSFC between the DLPFC and left hippocampus was significantly correlated with week 12 HAMD score (R² = 0.162, p = 0.001) and HAMA score (R² = 0.110, p = 0.009).

Given our previous report in this sample that anxiety and residual depressive symptoms among remitters at Week 12 independently predicted recurrence [27], we regressed the Week 12 HAM-D and HAM-A scores against whole brain RSFC patterns of the four seeded networks. For the HAM-A score, there was an inverse correlation with the ECN DLPFC seed and two regions of the DMN, specifically the medial prefrontal cortex and the PCC (Fig. 2a, Table 3). Week 12 HAM-D score similarly showed a significant inverse correlation between the DLPFC and medial prefrontal cortex (Fig. 2b, Table 3). These correlations were evident among all patients at Week 12, regardless of whether they went on to experience recurrence, thus representing neural correlates of residual symptoms rather than neural recurrence risks per se.

Fig. 2. Significant correlations between residual anxiety and depression severity scores and functional connectivity of the DLPFC.

A Significant negative correlation between week 12 HAM-A score and RSFC of the DLPFC seeds with anterior (mF) and posterior (PCC) components of the default mode network. B Significant negative correlation between week 12 HAM-D score and RSFC of the DLPFC seeds with anterior (mF) component of the default mode network. DLPFC Dorsolateral prefrontal cortex, HAM-A Hamilton Anxiety Rating Scale, HAM-D Hamilton Depression Rating Scale, mF medial frontal cortex, PCC Posterior cingulate cortex, RSFC Resting state functional connectivity.

Table 3.

Connectivity patterns significantly correlating with anxiety and depression severity scores at Week 12.

Seed (Network)	Region	Brodmann Area	Yeo Network	R/L	Peak MNI coordinates			Cluster Size (mm3)	t Value
					x	y	z
A. HAMA
DLPFC (ECN)	Medial Frontal	32,10, 25	Default A	R	+2	+46	−15	2635	−4.27
	Posterior Cingulate	23	Default A	L	−3	−34	+31	1228	−4.35
	Cerebellum			L	−17	−70	−25	1859	3.96
SCC (AN)	Cerebellum			R	+13	−89	−31	1526	−3.89
B. HAMD
DLPFC (ECN)	Medial Frontal	32	Default A	R	+1	+47	−15	1158	−4.01
	Middle Frontal	10	Control B	R	+28	+47	−7	1219	−4.69
	Inferior Frontal Junction	8	Default C	L	−46	+12	+47	1441	−4.62
SCC (AN)	Cerebellum			L	−4	−65	−34	2695	−4.39
Ant.INS (SN)	Secondary Visual	18	Visual B	R	+11	−89	+20	1977	4.47

Family-wise error corrected at p < 0.05.

Discussion

This study identified patterns of brain functional connectivity that predicted subsequent depressive episode recurrence over 21 months of follow-up among recently remitted adult treatment-naïve MDD patients. Recurrence was predicted by: 1) greater RSFC between the SCC of the affective network and the right anterior insula of the salience network, 2) reduced RSFC between the SCC and bilateral V1 regions of the Visual B network, and 3) reduced RSFC between the anterior insula and the bilateral caudate. Notably, the risk for recurrence identified by these connectivity patterns was present among patients acutely treated with either antidepressant medication or with CBT, and the recurrences occurred despite CBT booster sessions or maintenance medication treatment. This result suggests that at least some of the biological risk for recurrent episodes is not mitigated by the first-line treatments for MDD. The identified connectivity patterns among remitted patients who went on to recurrence did not correlate with the level of anxiety or depressive symptoms at the time of the Week 12 scan. Moreover, none of the RSFC patterns that predicted eventual recurrence were present at the baseline scan, nor were shared with the RSFC change patterns associated with remission during acute treatment [26]. These findings suggest that brain network connectivity in the remitted state may have unique explanatory value for understanding recurrence risk in depressed adults and has the potential to identify those patients who will need additional recurrence-prevention treatments [4, 44].

The study results overlap to some extent with the limited work that has employed functional neuroimaging to identify risk of future major depressive episodes. Among resting state fMRI studies, Workman and colleagues [21], using a region of interest analysis with left anterior sgACC seed among 47 un-medicated remitters, found that greater RSFC between the left and right anterior subgenual anterior cingulate regions predicted recurrence over 14 months of follow-up. This study scanned subjects in an eyes-closed condition, so associations with visual regions may have been unlikely to emerge. Studying 55 un-medicated remitted young adults with MDD, Langenencker and colleagues [22] reported that those who went on to experience recurrence had increased RSFC between the left sgACC and bilateral inferior frontal, middle frontal, inferior parietal, supplementary motor, precuneus and posterior inferior temporal gyrus regions. These authors concluded that among patients at risk for MDD recurrence, connectivity between the cognitive control network and the salience-emotional network was increased; the results of the current study did not replicate that finding despite using seeds for both networks.

Among task-based fMRI studies, future recurrence has been predicted by altered reactivity to viewing sad movie clips among 32 medicated and remitted MDD patients, specifically greater reactivity in medial prefrontal regions (Brodmann area 32) and lower reactivity in V1 (Brodmann area 17) [19]. Although our analysis used RSFC rather than reactivity to stimuli, the reduction in SCC-V1 connectivity associated with recurrence in our sample is consistent with the visual region findings of Farb and colleagues. Another study, during which participants processed self-blame statements, found greater RSFC of a right superior anterior temporal lobe seed with both the right putamen and the septal region adjacent to the SCC [20]. Notably, this latter study only seeded the right temporal lobe region, which may have limited its ability to identify contralateral regions involved in recurrence risk. Finally, a whole brain analysis using a DMN seed found no statistically significant clusters associated with recurrence among 39 remitted MDD patients clinically re-assessed two years after scanning [45]. Similarly, our primary analysis did not find any connectivity of the DMN seed to be associated with recurrence, though the secondary analysis identified greater DMN-supplementary motor area RSFC among recurrent patients.

Several other RSFC network connectivity patterns in our analysis that initially appeared to predict recurrence did not survive correction for multiple comparisons but have been implicated in the pathophysiology of MDD, including increased SCC RSFC with the left putamen, reduced insula-caudate, and reduced DLPFC-hippocampus RSFC [46, 47]. The latter finding is particularly noteworthy, given substantial neuroimaging research implicating both the DLPFC [23] and the hippocampus in the pathophysiology of MDD and response to treatments, possibly related to the role of the DLPFC in inhibiting limbic reactivity [48]. Segal and colleagues, in a non-imaging study, previously demonstrated that increased dysfunctional cognitions in reaction to sad mood provocation by music predicted recurrence over 18 months of follow-up among patients who had achieved remission after treatment with either antidepressant medication or CBT [49]. Unfortunately, PReDICT did not have specific assessments of ECN- or hippocampus-dependent functions, such as regulation of mood reactivity or cognitive functioning, which could be included in future neuroimaging studies of recurrence risk. Among the RSFC patterns of recurrence that emerged from our secondary analyses, only the DLPFC-hippocampus RSFC was significantly correlated with the level of residual depressive and anxiety symptoms among remitters at Week 12.

None of the primary analysis RSFC patterns that predicted recurrence were significantly associated with the clinical predictors of recurrence in our sample: Week 12 HAM-D score, HAM-A score, or comorbid anxiety disorder at baseline. Low power due to the limited number of recurrences or limitations in the imaging analyses (described below) may have contributed to the absence of associations between the imaging and clinical predictors. It is noteworthy, however, that Workman and colleagues also found no overlap between the level of residual depressive symptoms and their sgACC neuroimaging signature predictive of recurrence [21].

Given that most of the patients who experienced recurrence in this study were maintained on maintenance medication or received CBT booster sessions, these results support the value of additional specific recurrence-prevention interventions. Multiple studies have demonstrated the utility of preventive cognitive therapy [50, 51] or similar psychotherapy approaches [4, 52] aimed at preventing depressive recurrence, which is effective for patients who remitted to acute treatment with either psychotherapy or antidepressant medication. Less well studied are the recurrence prevention benefits of pharmacotherapy addition after acute-phase improvement with CBT, for which some trials suggest benefit [27, 44].

There are several strengths to the current analysis. Patients were followed prospectively for 21 months after achieving remission with their first adequate lifetime treatment for MDD, and recurrence assessments conducted every three months using rigorous criteria. We analyzed the four brain networks most consistently implicated in the pathophysiology of MDD with a whole-brain resting-state analysis, thus avoiding the multiple comparison concerns of task-based region of interest analyses [53]. The clinical predictors of recurrence identified in the sample, including the level of residual depressive and anxiety symptoms and the presence of an anxiety disorder, are consistent with the clinical predictors identified by others. Finally, focusing on the individual-level outcome of recurrence provides the most immediate relevance to the clinical decision making involved in the care of patients; however, it is possible that an analytic approach that employed depressive symptoms as a continuous measures, rather than as a categorical outcome, may have increased statistical power for RSFC pattern identification.

The primary limitation to this analysis was the relatively small number of patients experiencing recurrence. Several characteristics of our clinical trial sample contributed to the relatively low recurrence rate, including their treatment-naïve status, being in remission at the start of follow-up, and the continuation of treatment during maintenance. Although this analysis was limited to the patients with usable pre- and post-treatment MRI scans, the recurrence rate of 14.7% is very similar to the 15.5% overall rate of recurrence in PReDICT [27] and other studies of treatment-naïve patients [15]. Comparison to other imaging-based recurrence prediction studies was limited due to our patients continuing active treatment at the time of the prediction scan; conducting scans while on treatment, however, is better suited to any eventual clinical deployment for recurrence prediction, as the decision to taper treatments would be made based on recurrence risk. The limited duration of the 7.4-minute resting state scan could have reduced the power and reliability of the functional connectivity analyses [54]; future studies should consider acquiring longer periods of resting state data. Another potential limitation was our decision to apply a whole-brain analysis using seeds for four networks of relevance to MDD. Although this approach was consistent with our prior work with this dataset [26], a more constrained approach focusing on specific regions of interest might have offered greater power to detect potential RSFC patterns that reflected recurrence risk. Finally, alternative seeds could have been selected for the SN (e.g., an anterior cingulate seed) or the DMN (e.g., a medial prefrontal cortex seed), which may have produced differing results.

Given the high prevalence and adverse life impacts of recurrent episodes in MDD, it is remarkable that the number and size of biological studies exploring the prediction and moderation of recurrence are so sparse. There is a clear and pressing need for studies that biologically characterize and longitudinally follow large samples of MDD patients to understand better the longitudinal course of illness; this would allow effective preventative treatments to be applied to those who need them and the harms of unnecessary treatments could be avoided in those who do not. Further development of brain-based biomarkers that indicate probable MDD recurrence, supplemented by clinical or peripheral biological surrogate measures [55], could support clinical decision-making regarding maintenance or tapering of treatment and contribute to the development of interventions that specifically target recurrence-risk biology. Given our observed lack of association between neuroimaging and clinical predictors of recurrence risk, optimized individual-level prediction of recurrence is likely to require the combination of clinical and biological variables.

Author contributions

BWD drafted the manuscript. BWD, CBN, WEC, and HSM designed the study. BWD, WEC, and HSM collected the study data. JC and KC analyzed the imaging data and prepared the figures. All authors edited the manuscript and approved the final version.

Funding

Supported by NIH grants P50 MH077083, RO1 MH080880, UL1 RR025008, M01 RR0039, K23 MH086690 and funding from the Fuqua family foundations. Forest Laboratories and Elli Lilly donated the study medications (escitalopram and duloxetine, respectively).

Competing interests

Dr. Dunlop has received research support from Boehringer Ingelheim, Compass Pathways, NIMH, Otsuka, Sage, Usona Institute, and Takeda and has served as a consultant for Biohaven, Cerebral Therapeutics, Greenwich Biosciences, Myriad Neuroscience, NRx Pharmaceuticals, Otsuka, Sage, and Sophren Therapeutics. Dr. Cha reports no financial relationships with commercial interests. Dr. Choi has served as a consultant for Abbott Laboratories. Dr. Nemeroff has served as a consultant for AbbVie, ANeuroTech (division of Anima BV), Signant Health, Magstim, Inc., Intra-Cellular Therapies, Inc., EMA Wellness, Sage, Silo Pharma, Engrail Therapeutics, Pasithea Therapeutic Corp., GoodCap Pharmaceuticals, Inc., Senseye, Clexio, Ninnion Therapeutics, EmbarkNeuro, SynapseBio, BioXcel Therapeutics, and Relmada Therapeutics; he has served on scientific advisory boards ANeuroTech (division of Anima BV), Brain and Behavior Research Foundation (BBRF), Anxiety and Depression Association of America (ADAA), Skyland Trail, Signant Health, Laureate Institute for Brain Research (LIBR), Inc., Heading Health, Pasithea Therapeutic Corp., and Sage; he holds stock in Seattle Genetics, Antares, Inc., Corcept Therapeutics Pharmaceuticals Company, EMA Wellness, Naki Health, Relmada Therapeutics; he serves on the Board of Directors for Gratitude America, ADAA, and Lucy Scientific Discovery, Inc.; and he is named on U.S. patents 6,375,990B1 and 7,148,027B2. Dr. Craighead serves on the National Advisory Board for the George West Mental Health Foundation, as a board member of Hugarheill ehf (an Icelandic company dedicated to the prevention of depression), and as a scientific advisory board member for AIM for Mental Health and the Anxiety and Depression Association of America; he is supported by the Mary and John Brock Foundation, the Pitts Foundation, and the Fuqua family foundations; and he receives book royalties from John Wiley. Dr. Mayberg has received consulting and intellectual property licensing fees from Abbott Neuromodulation.

Supplementary information

The online version contains supplementary material available at 10.1038/s41386-023-01653-w.