Towards clinically useful neuroimaging in depression treatment: Is subgenual cingulate activity robustly prognostic for depression outcome in Cognitive Therapy across studies, scanners, and patient characteristics?

Abstract

Context

40–60% of unmedicated depressed individuals respond to Cognitive Therapy (CT) in controlled trials. Multiple previous studies suggest that activity in the subgenual anterior cingulate predicts outcome in CT for depression, but there have been no prospective replications.

Objective

This study prospectively examined whether subgenual cingulate activity is a reliable and robust prognostic outcome marker for CT for depression and whether its activity changes in treatment.

Design

Two inception cohorts were assessed with fMRI on different scanners on a task sensitive to sustained emotional information processing before and after 16–20 sessions of CT, along with a sample of control participants tested at comparable intervals.

Setting

Therapy took place in a hospital outpatient clinic.

Patients

Participants included 49 unmedicated depressed adults and 35 healthy control participants.

Main Outcome Measures

Pre-treatment subgenual anterior cingulate activity in an a priori region in response to negative words was correlated with residual severity and used to classify response and remission.

Results

As expected, in both samples, participants with the lowest pre-treatment sustained subgenual cingulate (sgACC; BA25) reactivity in response to negative words displayed the most improvement in CT (R²=.29, >75% correct classification of response, >70% correct classification of remission). Other a priori regions explained additional variance. Response/Remission in Cohort 2 was predicted based on thresholds from Cohort 1. sgACC activity remained low for remitters following treatment.

Conclusions

Neuroimaging provides a quick, valid, and clinically applicable way of assessing neural systems associated with treatment response/remission. sgACC activity, in particular, may reflect processes which interfere with treatment, e.g,. emotion generation in addition to its putative regulatory role; alternately, its absence may facilitate treatment response.

Introduction

Cognitive Therapy (CT) ¹ is a common empirically supported intervention which addresses systematic negative thinking and is effective for 40–60% of patients with unipolar depression ². Knowing which patients are likely to benefit from CT could increase response rates and decrease costs by targeted referrals. Yet, biological predictors of CT outcome e.g., ^{3, 4–6} have not been adopted clinically. This may be due to the lack of reliability and validity data for these measures or lack of ability to overcome variability across scanners/labs. Here we examined whether our previously-observed association of CT outcome with pre-treatment activity in the subgenual anterior cingulate cortex (sgACC) ⁴ replicates in multiple new samples and scanners.

The sgACC is an intuitive target to examine as a treatment predictor as it has connections to limbic regions such as the amygdala and has been suggested to serve as a proximal regulator of limbic function ⁷. It further has abnormalities in activity in depressed individuals e.g., ^{8, 9}, changes in CT and treatment with medications ^{10, 11}, as well as associations with symptom change in depression e.g., ^{12, 13}. Additionally, it is cytoarchitectonically uniform and has easily anatomically identifiable boundaries making results likely to generalize across studies. It is prognostic for clinical change in multiple CT studies using different paradigms in depression ^{4, 5} and other disorders such as PTSD ¹⁴.

To pave the way for clinical adoption of this prognostic biomarker, we examined whether sgACC prediction of outcome could be replicated in an efficacy sample (i.e., best possible conditions for observing the effect) as well as an effectiveness sample (i.e., more real-world, less ideal conditions). Thus we examined whether sgACC prediction of depression outcome in CT 1) holds in an efficacy cohort using effectively the same recruitment, design (including task), therapists, sample size, trial selection (e.g., measuring reactivity to negative words), data preparation, and analytic techniques as in ⁴; in addition to common arguments regarding the importance of replication before translation to the clinic, particularly given a single N=14 study, the need for strict replication is particularly great for voxelwise neuroimaging studies as relationships of voxelwise activity to self-report measures are notoriously unreliable ^{15, 16}, 2) holds in an effectiveness cohort with more heterogeneous clinically representative sample and community therapists with variable supervision, 3) is revealed as one of the measured neural indices which best predict treatment outcome, 4) can be formulated in a way that can be easily interpreted by clinicians and used across scanners, and 5) changes in treatment for those with the predictive marker, to inform, if not confirm, causal inferences regarding the possible role of sgACC change in clinical change.

For prediction (goals 1–2), our primary hypothesis was that sgACC activity would be prognostic for outcome in CT in multiple samples, robustly enough to garner clinical consideration. Our secondary hypotheses were that other relevant theoretically motivated regions and their functional relationships would add variance to prediction of outcome but would not completely moderate the role of the sgACC (goal 3), that the same predictive associations would be apparent in normalized data (goal 4), and that sgACC activity would normalize following treatment (goal 5).

For the “other relevant regions” (goal 3), we examined a brain network associated with emotional reactivity and implicated in depression and treatment outcome ¹⁷ including the left amygdala as in ^{4, 18} and regions associated with regulatory control that have decreased functioning in depression: the dorsolateral prefrontal cortex e.g., ^{8, 19, 20}, rostral anterior cingulate ^21–23, and their functional corticolimbic/corticocortical relationships e.g., ^{24, 25}, as well as a voxelwise analysis.

Directional hypotheses for change in sgACC activity (goal 5) are dependent on the somewhat ambiguous function of this region. On the theory that the sgACC inhibits limbic regions e.g., ^{26, 27, 28}, we have suggested that as CT teaches skills for emotion regulation, individuals who most need CT, i.e., those with the lowest pretreatment levels of sgACC activity, may respond best to CT ⁴. Thus, CT remitters would be expected to show decreased pre-treatment but increased post-treatment sgACC activity. Alternately, the sgACC may also support emotion generation / monitoring. In support of this theory, 1) neurofeedback-induced increased sgACC activity yields increased sadness ²⁹, 2) trait-related increased sgACC is associated with higher sadness ³⁰ and increased depressive severity ³¹, and 3) sgACC inhibition via deep brain stimulation facilitates recovery e.g., ³². If this role is primary, low sgACC activity may be needed for voluntary regulation to occur. In this case, remitters would be expected to show decreased pre- and post-treatment sgACC activity.

To evaluate whether sgACC activity “normalizes” in treatment, a sample of healthy never-depressed control participants were recruited to establish a “normative” baseline. The controls also afforded the ability to examine pretreatment activity with respect to normative function, i.e., as clinically interpretable Z-scores (goal 4) and to assess the test-retest reliability outside the context of depression – i.e., showing that healthy individuals do not change strongly over time and whether the measure is reliable enough to make inferences regarding pre-post measurements (e.g., as recommended in ³³). Data from controls was not used in primary prediction analyses.

Method

Participants and therapists

As shown in Figure 1 (CONSORT diagram), and Table 1 (demographics), participants from two clinical trials were tested in two cohorts on different scanners, separated by approximately 6 months (full elaboration in Author Material-I). After attrition and data cleaning (fully described in Figure 1), Cohort 1 (efficacy: from Thase, PI; ClinicalTrials.gov: NCT00183664) included 17 patients with recurrent major depressive disorder (diagnoses via SCID ³⁴) treated by the same three therapists from ⁴ who continued to receive weekly supervision with tape-review from the same master clinician (Sandar Kornblith, Ph.D.), as well as 15 healthy controls (no current or historical Axis I disorder via SCID interview). Therapeutic adherence was deemed adequate with a random selection of 10 tapes (ratings of >=2 tapes per therapist by other therapists or outside raters) receiving high marks on the Cogntive Therapy Rating Scale ³⁵ (scores of 40 represent high levels of adherence), Mean (Stdev)=52.7(8.6), Range = 40–63.

Figure 1.

CONSORT diagram

Table 1.

Demographic and clinical characteristics. T-tests confirmed that no groups or cohorts were different from any other on any demographic variable below (gender, ethnicity, age range, age, education). Cohorts did not differ in depressive severity or median # of depressive episodes. Group and cohort differences in emotion ratings and reaction times are described in the text and Author Material-VI. For BDI and HRSD, Response was defined as a 50% decrease in BDI or Hamilton. The first lines, of the clinical response section, using only CT completers, replicates the analysis done in Siegle et al (2006). Remaining lines generalize that result. Remission was defined as BDI<10 or HRSD <10

Measure	Depressed	Depressed	Control	Control	Test
	Cohort 1 CT	Cohort 2 CT	Cohort 1	Cohort 2	Depressed v. Control
N	17	32	15	20
# male	4	5	7	5
# Caucasian	14	29	12	17
Age range	24–51	19–55	25–54	21–52
Age M(SD)	37.41 (9.17)	36.13 (11.11)	38.93 (9.74)	32.95 (9.37)
Years Education M(SD)	15.35 (2.23)	15.06 (2.60)	15.67 (2.26)	16.30 (2.13)
Median # Depressive Episodes	5.59 (3.52)	4.03 (4.83)	0	0
NAART VIQ equivalent	110.31 (8.68)	108.42 (9.92)	111.61 (7.97)	109.21 (8.96)
Emotion ratings for Negative words (1=very negative-7=very positive) M(SD)	2.55 (0.66)	2.11 (0.62)	2.99 (0.65)	2.49 (0.72)	t(82)=2.89, p=0.005, d=0.64
RT to negative words M(SD) in ms	1403.82 (269.41)	1637.34 (526.16)	1411.31 (390.26)	1242.20 (374.90)	t(82)=−2.52, p=0.01, d=0.56
Personal relevance ratings for Negative Words (1=not relevant, 5=very relevant) M(SD)	3.07(.66)	3.03(.57)	2.33(.72)	2.07(.49)	t(82)=6.57, p<.0005, d=1.43
Clinical Change Completers only N=13 as in (11)
Pre-treatment BDI M(SD)	33.0 (8.8)
Post-treatment BDI M(SD) Completers only	10.7 (10.2)
BDI Change M(SD)	22.3 (13.4)
N Responders	9 (69%)
N Remitters	8 (61%)
Clinical Change – whole sample: N BDI	16	27	15	20
Pre-treatment BDI M(SD)	34.4 (8.7)	30.7 (9.2)	1.87 (3.58)	1.26 (1.63)
Post-Treatment BDI M(SD)	14.1 (12.2)	11.3 (9.6)
Change BDI M(SD)	20.3 (13.4)	19.4 (9.7)
N Responders BDI	10 (63%)	17 (62%)
N Remitters BDI	8 (50%)	14 (51%)
Clinical change – whole sample N (HRSD)	17	32
Pre-treatment HRSD M(SD)	24.6 (5.7)	20.6 (4.8)
Post-Treatment HRSD M(SD)	9.6 (6.5)	8.1 (5.8)
Change HRSD M(SD)	15.0 (7.7)	12.5 (5.6)
N Responders HRSD M(SD)	10 (63%)	20 (63%)
N Remitters HRSD M(SD)	7 (47%)	13 (41%)

Cohort 2 (effectiveness) included 32 patients who were more clinically heterogeneous than Cohort 1 as they included N=23 with recurrent major depressive disorder but also N=9 in their first-episode. Participants were drawn from the same trial or Siegle, PI, ClinicalTrials.Gov: NCT00787501 (CT election given options of CT or SSRI treatment). Cohort 2 also included 20 healthy controls (no current or historical Axis I disorder via SCID interview). Patients were treated by 6 community clinicians (Ph.D.’s, M.D.’s, M. Ed.’s, LCSW’s) who ranged in CT experience from a Ph.D. founding member of the Academy of Cognitive Therapy to a social worker who took her second CT case as part of this study. Therapists received group supervision monthly without tape review, and as-requested supervision by Dr. Kornblith (weekly for 2 therapists). Therapeutic adherence was more variable with 12 tapes of 4 therapists receiving Cognitive Therapy Rating Scale scores of Mean (Stdev)=43.3(13.1), Range=23–61 with 6/12 tapes falling below the cutoff of 40 for adequate adherence, as administered by other study therapists or outside raters.

Participants described no health problems, eye problems, or psychoactive drug abuse in the past six months and no history of psychosis, manic, or hypomanic episodes. Neither control nor depressed participants had used antidepressants within two weeks of testing (six weeks for fluoxetine) due to either medication naivity or supervised withdrawal from unsuccessful medications. Participants reported no excessive use of alcohol in the two weeks prior to testing and scored in the normal range on a cognitive screen, ³⁶; VIQ-equivalent > 85.

Procedure

After complete study description, IRB-approved written informed consent was obtained followed by a SCID interview, vision test, and unrelated physiological assessment. Patients were assessed on a different day with a battery of fMRI tasks administered in counterbalanced order; one task is reported here (others described in Author Material III-A). Participants rated their sad, anxious, and happy affect from 1 (not at all) to 5 (very) before and after the task. The Beck Depression Inventory II ³⁷ (BDI) was administered following fMRI to assess depressive severity (rationale in Author Material II-A). Depressed participants then received CT; 2 sessions/week for the first four weeks followed by 8 weekly sessions for “early-responders” (HRSD reduction <40% at session 9; 16 total sessions) or 2 sessions/week for the first 8 weeks followed by 4 weekly sessions for non-early-responders (20 total sessions). CT followed Beck’s¹ guidelines (detail in Author Material-I) with weekly Hamilton Rating Scale for Depression (HRSD) ratings. Within two weeks of completion (week 16 for controls) all participants completed the same fMRI protocol and BDI again.

Clinical data preparation

In Cohort 1, 1 participant’s pre-treatment and 1 participant’s post-treatment BDI were missing. Fortunately the BDI-I was administered clinically in the protocol, and thus these values were reconstructed via regression (described Author Material-II-B). In all other cases, final BDI scores were assessed at participants’ second scan (after completing CT or approximately 12 weeks from their first scan for those who did not complete). Final HRSD scores were imputed based on the trajectory of weekly responses as described in Author Material II-D. Response was defined as a 50% reduction in initial BDI or HRSD score (as opposed to the “early response” criterion of 40% reduction used to determine the number of sessions), and remission was defined as final-BDI <10 (rationale in Author Material-II-C) or final-HRSD <7 (as in STAR*D).

fMRI task and processing

Apparatus

Twenty-nine 3.2mm slices were acquired parallel to the AC-PC line using a reverse direction EPI pulse sequence to minimize susceptibility artifacts in the amygdala and orbitofrontal regions (3T Siemens Trio, T2*-weighted images depicting BOLD contrast; TR=1500ms, TE=27ms, FOV=24cm, flip=80), yielding 8 whole-brain images per 12 second trial. Stimuli were displayed in black on a white background via a back-projection screen (.88° visual angle). Responses were recorded using a Psychology Software Tools^™ glove.

Personal relevance rating task (PRRT)

As in our previous publications ^{4, 20}, in 60 slow-event related trials, participants viewed a fixation cue (1 s; row of X’s with prongs around the center X) followed by a positive, negative, or neutral word (200 ms; only negative words analyzed here), followed by a mask (row of X’s; 10.8 s). Participants pushed a button for whether the word was relevant, somewhat relevant, or not relevant to them or their lives (button orders balanced across participants), as quickly and accurately as they could. Participant-generated and normed words from ³⁸ were used as in our previous studies of depression ^{4, 20, 39–41} (procedures in Author Material-III; RT preparation in VI-B).

fMRI data preparation

Following standard preprocessing similar to that used in our previous study ⁴ with slight modernizations (slice time correction, motion correction, linear detrending, voxelwise outlier rescaling, conversion to percent-change, temporal smoothing (5 point middle peaked filter), 32 parameter nonlinear warping the Montreal Neurological Institute Colin-27 brain, and spatial smoothing (6mm FWHM), response time-series variability normalization across scanners; methods fully described and compared to ⁴ in Author Material-IV), the same “reactivity” index used in ⁴ yielded peak and sustained responses to negative words as the mean of the 4^th–7^th images (henceforth “scans”) of each negative-word trial minus the trial’s first (pre-stimulus) scan acquired while the fixation cue was on the screen.

Primary hypotheses were examined for mean reactivity in a 24 voxel 20mm radius sphere centered at Talairach 6, 17, −6, the centroid of the sgACC region prognostic for outcome in ⁴. Our broader network included right and left dorsolateral prefrontal cortex (DLPFC) regions, the left amygdala, and BA24 in the rostral-cingulate (criteria described Author Material-V) and to explicitly replicate methods from ⁴ we also examined regions from voxelwise analyses, subject to empirical Type I error control (Author Material IV-B).

Data Analysis: Associations between pre-treatment brain function and clinical change

Using residual severity controlled for initial severity. To replicate methods from ⁴ we examined associations of pre-treatment activity with residual severity, computed by regressing final Beck Depression Inventory scores on initial scores and retaining the residuals in participants who completed treatment in Cohort 1. To increase generalizability we then included participants who did not complete treatment in this cohort as well as Cohort 2. To predict response and remission, a within-sample grid-search found the cutoff that maximized %-correct discrimination. Standard indices of signal detection are reported (sensitivity, specificity, d′; ROC statistics in Author Material-VIII). 1000 permutation tests using the same algorithm (randomly permuting associations of response or remission with fMRI indices) assessed the significance of discrimination (%-correct and d′; tests of d′ generally agreed with %-correct, and thus are only reported when results diverged on significance). To increase generalizability outside clinical trials in which residual severity can be calculated, we examined Z-scores of sgACC activity (previously described reactivity index) normalized with respect to the same index computed for controls’ initial assessment, i.e., sgACC_Z=(X−M_controls)/sd_Controls, and severity change scores (final – initial). Multiple regions were associated with residual severity using multiple regression.

Robustness was examined via random forest regression and classification ⁴², using as features activity from all a priori structures (criteria in Author Material V) and their partial mutual information via ⁴³ to capture systemic effects involving functional-connectivity. Regression and classification forests were trained on Cohort 1 using random re-sampling (bagged); a subset of the activation and mutual information measures was selected by maximizing out-of-bag estimates of prediction and classification accuracy on Cohort 1. We tested generalizability by computing prediction and classification accuracy of the Cohort 1 trained algorithms to Cohort 2. P-values were obtained by permuting responses and computing the proportion of times the permuted regression and classification results were as good or better than the un-permuted results.

Data Analysis: Change

We examined associations of pre- and post-treatment activity via correlation, and change in the time-series of responses for participants with each combination of predicted and actual response.

Results

Demographic and Behavioral Data

Patients and controls did not differ significantly on demographic variables between cohorts or groups (Table 1). Group x cohort ANOVAs revealed that patients rated negative words as more negative and more personally relevant, and reacted more slowly than controls (Author Material-VI). Cohort 2 rated negative words as more negative t(82)=3.27, p=0.002, d=0.73 with no other Cohort main effects. There were no group x cohort interactions on valence or meaning ratings. As group x cohort effects on rts (Author Material-VI) were small and the hemodynamic window began after their likely effects, they were not used in fMRI analyses. There were no significant associations of rt, valence ratings, or personal relevance ratings with residual BDI scores or sgACC activity, |r|<.15, p>.4. Pre-treatment depressive severity (BDI) was uncorrelated with pre-treatment sgACC activity (r=−.13). Depressed participants were moderately sad and anxious and minimally happy before and slightly less so after the task whereas controls were comparably minimally sad and anxious and moderately happy before and after the task (Author Material-VI-C, Table S1).

Clinical Outcome

CT was successful at rates at least as high as those observed in the literature with enough outcome variability (N>5 in all response/remission cells) to allow further analysis (Table 1, individual trajectories, Author Material-VII).

1) Efficacy replication: sgACC activity as prognostic indicator of clinical change in CT

As shown in Figure 2 and Table 2A, as in our previous study ⁴, decreased sgACC activity was strongly correlated with stronger clinical change (more negative BDI_residual) in Cohort 1, whether or not completers were included. This prediction generalized to the HRSD, which was collected on a slightly larger subset of participants. ROC curves (Author Material-IX) yielded significant area-under-the-curve (AUC), reflecting good discrimination (Author Material-VIIIA). Voxelwise associations with BDI_residual revealed sgACC to be associated with even more of the variance (> 80% in some voxels; Figure 3c), suggesting possible optimization and, as in ⁴, relative specificity to the sgACC.

Figure 2.

Decreased anatomically defined sgACC activity was strongly correlated with stronger clinical response (decreased residual BDI) in A) Cohort 1 (green are non-completers whose final BDI scores were interpolated), moderately associated in B) Cohort 2, and more strongly associated in C) the combined cohort as well as D) the combined cohort using Z scores for sgACC and change in BDI as the outcome variable. E) Waveforms for hemodynamic responses for nonremitters (final BDI>=10; N=22) compared to controls (N=35) and remitters (final BDI < 10; N=22). Areas significant at p<.05 upon ANOVAs at each scan are highlighted in grey.

Table 2.

Associations of A) anatomical sgACC activity with clinical change using each cohort as a sample of interest and B) anatomical sgACC and a broader a priori network activity with clinical change, along with C) prediction for Cohort 2 using Cohort 1 as a training set. Note: a more elaborated version of this table including full Receiver Operating Characteristic (ROC) analyses is included in Author Material VIII-A, Table S2A.

Cohort	Measure	Residual		Response Prediction (% correct, permutation p)	Remission prediction (% correct, permutation p)
A. Prediction using sgACC with each Cohort as a sample of interest
Cohort 1 Completers (as in 4) (N=14)	BDI	R2 = 0.414, F(1,12)=7.781, p=0.018		84% correct, p=.1 (.04 based on d′), Sensitivity = 100%, Specificity = 86%, d′=2.2	84% correct, p=.08 (.04 based on d′) Sensitivity = 100%, Specificity = 75%, d′=2.1
Cohort 1 Completers and Noncompleters (estimated) (N=16)	BDI	R2 = 0.358, F(1,15)=7.79, p=0.014		81% correct, p=.07 Sensitivity = 83%, Specificity = 80%, d′=1.8	75% correct, p=.14 Sensitivity = 100%, Specificity = 50%, d′=1.4
(N=17)	HRSD	R2 = 0.341, F(1,16)=7.772, p=0.014		76% correct, p=.41 Sensitivity = 80%, Specificity = 67%, d′=1.3	82% correct, p=.045 Sensitivity = 100%, Specificity = 57%, d′=1.7
Cohort 2 (N=27)	BDI	R2 = 0.167, F(1,26)=4.91, p=0.04		78% correct, p=.04 Sensitivity = 40%, Specificity = 100%, d′=1.5	74% correct, p=.04 Sensitivity = 54%, Specificity = 93%, d′=1.6
(N=32)	HRSD	R2 = 0.040, F(1,31)=1.244, p=0.273		78% correct, p=.02 Sensitivity = 42%, Specificity = 100%, d′=1.61	66% correct, p=.38 Sensitivity = 32%, Specificity = 100%, d′=1.32
Cohorts 1 & 2 combined (N=43 with pre/post BDI scores (of N=49)	BDI	R2 = 0.288, F(1,42)=16.61, p<0.0005		79% correct, p=.003, Sensitivity = 50%, Specificity = 96%, d′=1.79	72% correct, p=.01, Sensitivity = 38%, Specificity = 95%, d′=1.39
(N=49)	HRSD	R2 = 0.12, F(1,48)=6.29, p=0.02		78% correct, p=.001 Sensitivity = 47%, Specificity = 93.8%, d′=1.46	69% correct, p=.06 Sensitivity = 34%, Specificity = 100%, d′=1.58
B. Prediction of residual symptoms using a priori network with combined cohort as a sample of interest (N=43)*
Zero order relationships All p’s <.05 – see Author Material VIII Table S2a for statistics.	BDI	sg ACC	R2= 0.29
		R Amygdala	R2= 0.16
		L DLPFC	R2= 0.20
		BA24 in the VMPFC	R2= 0.11
Multivariate relationships Full model R2=.43, F(1,38)=7.39, p=.0001. Statistics in Author Material VIII Table S2a	BDI	Constant	stβ=0, p=.21
		sgACC	stβ=.47, p=.0005
		R Amygdala	stβ=.24, p=.10
		L DLPFC	stβ=.30, p=.05
		BA24 in the VMPFC	stβ=−.10, p=.52
C. Generalizable classification from just sgACC or all 4 regions and their connectivity using random forest methodology with Cohort 1 as the training set and Cohort 2 as the test set.**
1, CT – training set − all brain regions	BDI	R2 = 0.28, p<0.02		75% correct, Sensitivity = 75%, Specificity = 90%, d′=1.28, p=.023	69% correct, Sensitivity = 62%, Specificity = 75%, d′=.99, p=.057
	HRSD	R2 = 0.21, p<0.01		76% correct, Sensitivity = 92%, Specificity = 40%, d′=1.13, p=.028	65% correct, Sensitivity = 57%, Specificity = 70%, d′=.7, p=.12
2, CT – Generalization set − initial severity alone	BDI	R2 = 0, p<0.74		48% correct, Sensitivity = 53%, Specificity = 40%, d′=−.18, p=.63	38% correct, Sensitivity = 23%, Specificity = 54%, d′=−.64, p=.92
	HRSD	R2 = 0, p=0.63		56% correct, Sensitivity = 60%, Specificity = 50%, d′=.25, p=.29	47% correct, Sensitivity = 38%, Specificity = 53%, d′=−.23, p=.7
2, CT - Generalization set − initial severity + sgACC	BDI	R2 = 0, p=0.07		74% correct, Sensitivity = 100%, Specificity = 30%, d′=1.04, p=.03	78% correct, Sensitivity = 86%, Specificity = 69%, d′=1.57, p=.003
	HRSD	R2 = 0.02, p=.07		81% correct, Sensitivity = 95%, Specificity = 58%, d′=1.86, p<.001	62% correct, Sensitivity = 69%, Specificity = 58%, d′=.7, p=.08
2, CT – Generalization set − initial severity + all brain regions (retained: Baseline BDI, LDLPFC, sgACC, phi_LDLPFC_sgACC)	BDI	R2 = 0.15, p<0.003		81% correct, Sensitivity = 100%, Specificity = 44%, d′=1.43, p=.0005	63% correct, Sensitivity = 79%, Specificity = 46%, d′=.7, p=.086
	HRSD	R2 = 0.07, p=.02, M(points off)=3.8		75% correct, Sensitivity = 95%, Specificity = 42%, d′=1.43, p=.002	66% correct, Sensitivity = 85%, Specificity = 53%, d′=1.09, p=.015

^*Classification not evaluated in the multivariate model without robust estimation (2B) given the potential for type I error – rather, evaluations in 2C reflect robust estimations from the multivariate model.

^**p-values for random forests were estimated via permutation tests on %-correct using the same random-forest methodology. Regression and classification accuracy estimates unbiased computed on out-of-bag samples for robust generalization.

Figure 3.

Decreased empirically defined sgACC activity associated strongly with response in A) our previous study ⁴ and B) the new samples (R²>.4) – green is the anatomical ROI used as a mask, orange regions were only in the new dataset, and red voxels are overlapping. These regions were reflected in predictive regions in both (C) Cohort 1 and (D) Cohort 2.

2) Effectiveness generalization to a more clinically heterogeneous sample with minimally supervised community therapists

The same general pattern of results held significantly in Cohort 2 though effect sizes were somewhat weaker (Figures 2–3 and Table 2A), potentially due to clinical and therapist heterogeneity. Importantly, prediction of responder and remitter status was still strong with >75% correct predictions for response across all samples and measures, and >65% correct predictions for remission, preserving strong specificity. When data from Cohorts 1 and 2 were combined, yielding a mix of highly supervised, and not-as-strongly supervised therapists, and a wide range of the clinical phenotype, prediction remained strong. ROC analyses (Author Material VIII, IX) were significant reflecting fair discrimination. Remitters were characterized by downward-going hemodynamic responses whereas non-remitters were characterized by upward-going hemodynamic responses (Figure 2E), with significant differences in a 3-group (control, remitter, non-remitter) ANOVA from 6 to 12 seconds, F(2,76)=5.563, p=0.006.

3) Do other brain regions or measures explain additional variance?

Other regions

A voxelwise mega-analysis combining data across cohorts, (hierarchical regression in which scanner (cohort) and initial BDI were entered on step 1 to simultaneously covary them, with final BDI on step 2), confirmed a 2-voxel region across both samples (Talairach 5, 16, −9) which was both statistically significant at p<.001 (p<.05 corrected) and clinically significant, explaining >40% of the variation in BDI_residual, and located in the a priori mask centered around the region detected in our previous study (Figure 3). No other significant >24 voxel regions explaining >= 40% of the variance were detected. Similarly, voxelwise meta-analytic conjunction analyses revealed voxels in the sgACC at p<.01 in each cohort, but no other >24 voxel regions predictive of outcome in both cohorts. Thus, sgACC associations were confirmed and no new regions were added to the a priori list.

Associations were similarly strong across all a priori regions (Table 2B, Author Material XI) suggesting that the relationship of the sgACC to outcome is not qualitatively unique. Whereas the sgAcc explained 29% of the variation, the other 3 regions explained only 12% additional variation, yielding a non-significant increment, F(3,38)=2.65, p=.06, though the final model with all a priori regions as predictors was significant, prediction equation: BDI_residual =−1.80+43.13*sgACC+17.2*Amygdala+35.7*DLPFC−12.0*BA24. Thus, decreased possibly regulatory sgACC and DLPFC activity explains better CT outcome.

Other measures

Adding common non-fMRI predictors of response including scanner, rumination, pupillary motility during the task, and demographic variables did not explain additional significant variance, and sgACC activity remained a significant predictor (Author Material XI-D).

Other valences

Positive and neutral words did not predict residual symptomatology in the combined cohort. With positive, negative, and neutral words in the same model, only negative words were a significant predictor of residual symptomatology (Author Material XIII). Thus, prediction was specific to negative words.

Robust Classification

As shown in Table 2C, classification for response and remission status in Cohort 2 was strong based on a sgACC cutoff solely derived from Cohort 1. That is, by relying on the region from a previous study with a cutoff determined in Cohort 1, we achieved 74% correct prediction of response and 78% correct prediction of remission in Cohort 2. These data suggest that assessment on one scanner yielded a predictive algorithm applicable on another scanner. Adding indices of activity among multiple regions and their connectivity improved classification and continuous prediction of BDI/HRSD_residual suggesting that accounting for a broader a priori network robustly adds to predictive power; had additional terms not statistically improved classification they would not have been retained in the final model. ROC curve analyses (Author Material VIII ROC AUC, Author Material-XI-C) further show that initial severity yielded poor classification (AUC<.7). Multivariate classifications on the test dataset yield significant estimates reflecting fair (AUC>.7) to good (AUC>.8) discriminability with just decreased severity and decreased sgACC and good discriminability with the full model, including decreased sgACC-DLPFC connectivity, predicting outcome.

4) Clinically applicable formulations: Psychometrics, Z- and symptom-change scores

Reliability and discriminant validity

sgACC_Z and sgACC reactivity had moderate test-retest reliability in control participants tested approximately 16 weeks apart (N=27, r=.39, p=.04) which was similar at each scanner (Cohort 1, r=.44, Cohort 2, r=.35) and did not change when scanner was covaried (r=.39). Controls’ sgACC mean did not change between the scans, t(26)=1.36, p=0.19, d=0.26. All but 1 had Z<0.5 pre-test, and all but 2 at post-test, suggesting stability within a restricted range, decreasing the estimated reliability. Reliability for the composite predictor associated with the sgACC, amygdala, DLPFC, and BA24 was no higher, r=.36. sgACC_Z reactivity was minimally correlated with initial depressive severity among depressed patients (N=49) r(HRSD)=−.10, p=.50; r(BDI)=−.21, p=.13.

Symptom Change scores

Associations of sgACC_Z with individually interpretable BDI_Post-Pre were nearly identical to associations of %-change with BDI_residual (Figure 2D, Author Material-VIII, X). For example, in Cohort 1 completers’ sgACC_Z predicted 54% of the variance in BDI_Post-Pre, F(1,15)=16.32, p=.001, and R²=.29 in the combined sample F(1,42)=16.87, p<.005. sgACC _Z cutoffs in the range of .46-.74 strongly predicted response across measures for both samples (76–81% correct response classification), suggesting that high levels of sgACC activity predicted non-response. Similarly, cutoffs closer to 0 predicted remission, so participants with average or higher levels of sgACC responses compared to controls were unlikely to remit in CT.

5) Towards causal inference: Did sgACC activity change?

Forty depressed participants had pre and post-treatment fMRI and BDI scores. Those participants with the lowest pre-treatment sgACC activity (primarily CT remitters) also had the lowest post-treatment activity r_pre,post=.39, F(1,39)=6.82, p=0.013, (Figure 4A). Depressed participants with pre-treatment activity below the predicted response threshold who remitted had pre- and post-treatment sgACC activity below controls throughout the trial (Figure 4B top left) and did not increase significantly (Table S4, t-test pre. v. post). In contrast, controls with low pre-treatment activity increased significantly (Table S4, Figure S7b) and depressed non-remitters with low pretreatment activity nearly so (Figure 4B top right; statistics Table S4). Moreover, depressed remitters with low pre-treatment activity had a non-significantly smaller proportion of participants who increased and lower mean level of increase than did either controls or non-remitters with low pre-treatment activity (Author Material-XII-B). Thus, we cannot conclude that sgACC activity increased as a function of treatment. Qualitatively, in contrast, 5/6 participants with low-pre-treatment activity who did not remit (Figure 4A, green) showed increased sgACC activity following treatment. Similarly, 3/5 of the remitters who had high pre-treatment sgACC activity decreased post-treatment (Figure 4A, blue squares, average Figure 4B bottom left) whereas 8 of the 12 non-remitters with high pre-treatment sgACC activity increased (Figure 4A, purple squares). Thus, the emerging picture is that participants with high post-treatment sgACC activity did not remit.

Figure 4.

Relationships of pre-treatment sgACC reactivity to post-treatment sgACC activity, A) continuously for the temporal region of interest, and B) with regard to waveforms in participants who were predicted to remit (pre-treatment %-change <.02) or not remit. Areas significant at p<.05 upon ANOVAs at each scan are highlighted in grey.

Discussion

This study replicates and extends previous results ^{4, 44} suggesting that decreased sgACC reactivity to negative words is prognostic for response, remission, and clinical change in Cognitive Therapy for depression. It was prognostic in both an efficacy sample and a more clinically diverse sample who received more clinically representative care, though effect sizes were somewhat reduced, possibly due to increased patient and therapist heterogeneity and nonstandard administration of Cognitive Therapy. Using measures of brain function interpretable at a single-subject level (Z-scores) and simple change-scores preserved prognostic utility. Response, remission, and change in severity were robustly predicted based on a few a priori regions when using thresholds derived in a different sample on a different scanner above and beyond pre-treatment severity. These data suggest that the proposed assessment could have scientific and practical utility. But participants with low pre-treatment sgACC activity who remitted did not generally demonstrate increased sgACC activity suggesting that successful treatment did not operate by normalizing this mechanism but rather remaining more like healthy individuals from pre- to post- treatment.

Though failure to observe an increase in sgACC activity following treatment could be due to the index’s low reliability or regression to the mean, these explanations would not explain the association of initial negative-going responses with the smallest sgACC change. Thus, we suggest that increasing sgACC function is not a mechanism of clinical change. Rather, sgACC activity could interfere with voluntary emotion regulation essential for CT, possibly due to its roles in sadness-upregulation ²⁹ or limbic monitoring, possibly via corticocortical connectivity, or automatic down-regulation of emotion ⁴⁵ preventing the use of considered reappraisal emphasized in CT. These interpretations support the continued development of interventions that decrease sgACC activity e.g., ³² and do not suggest that CT or other interventions (e.g., neurofeedback) should work to increase ventromedial function. Alternately, low sgACC activity could facilitate CT response if it is easiest to learn to challenge thoughts when emotion-monitoring/generation processes naturally disengage. That multiple regions associated with limbic reactivity and regulation explain overlapping and independent variance suggests that outcome reflects functioning of a wider network.

More practically, because we used a priori regions with a single 12-minute fMRI and seven minute structural acquisition, automated preprocessing, and scores that can be calculated on single subjects, our algorithm is feasible and at significantly under $1000 at most scanning centers, cost effective for use in clinics in a 30 minute MRI appointment without radiologist interpretation. Cross-scanner generalizability is likely because sgACC activity patterns were qualitatively different for remitters (activity decreased from a pre-stimulus baseline) and non-remitters (activity increased) (Table 2, Figure 2E). Thus, inter-scanner scaling differences did not affect prediction. As sensitivity using just the sgACC was low for practical use (Table 2A) but adequate when other regions were accounted for (100% sensitivity in Table 2C generalization set, BDI prediction), it may be useful to assess a network of regions in prediction, or to consider the sgACC primarily as a method to help patients decide what interventions not to try first. As predictions were stronger in the highly controlled cohort, therapists who adhere strongly to a single treatment model may make the most valid use of predictive neuroimaging.

As increased activity in anatomically proximal regions including the rostral cingulate has been shown to positively predict outcome to antidepressants using PET resting state ^{12, 46, 47}, fMRI reactivity more analogous to the current design ^{13, 48–52}, and EEG ⁵³, ventral cingulate activity could, upon prospective replication, yield an algorithm for selection into CT versus antidepressant treatment ¹⁷. Assessing sgACC function during multi-modal assessments could also help to validate mechanisms for less interpretable prognostic indicators such as psychophysiological assessments ⁵⁴. That said, as fMRI-derived sgACC activity explained variance independent of other common predictive measures (Author Material XI-D), it may represent a more unique construct in the literature.

This study had multiple limitations. Though one of the largest prognostic fMRI studies in psychiatry, our samples were small by the standards of clinical trials. Randomized trials are needed to make testable inferences regarding the relationship of sgACC function specifically to CT outcome. Future examination of medicated depressed participants will help to generalize to clinically representative conditions. If healthy individuals’ moderate reliability (r<.5) is representative of patients, refinement in multiple-baseline patient studies is needed before inferring clinical applicability to single patients. Amygdala results were not consistent with ⁴, possibly reflecting our use of an anatomically defined amygdala region versus the empirically derived smaller region in ⁴. Our index captured sustained processing of emotional information but could have missed preparatory or initial aspects of reactivity.

These limitations notwithstanding, our data suggest that a simple, automated, and theoretically motivated neuroimaging assessment yields a marker of whether an individual is likely to respond to a specific validated treatment for unipolar depression, Cognitive Therapy. Though the data do not suggest that neuroimaging should be used clinically at this time, they do indicate that neuroimaging is ready for “next step” investigations, including larger randomized clinical trials, creating age, gender, and scanner, norms, and considerations regarding dissemination.