Gamified Attention Bias Modification Training to Augment Cognitive-Behavioral Therapy for Youth Anxiety Disorders: A Randomized Controlled Trial

Abstract

Objective:

Augmenting cognitive-behavioral therapy (CBT) with attention-bias modification training (ABMT) may maximize therapeutic benefit for anxiety through targeting different underlying processes. Yet, corresponding neural mechanisms are unclear. This pre-registered double-blind, randomized controlled trial assigned youth with anxiety disorders receiving CBT to active or sham gamified ABMT. Group differences in clinical efficacy, and associations between amygdala connectivity and treatment outcomes, were examined.

Method:

Pediatric patients (N=121; 8–17y) with anxiety disorders were randomized to active or sham gamified ABMT alongside 12-weeks of manualized CBT. Primary outcomes were clinician-rated Pediatric Anxiety Rating Scale (PARS) and Clinical Global Impression-Improvement (CGI-I) Scale. Amygdala seed-based connectivity during an fMRI dot-probe attention-bias task was measured before and after treatment in patients, and across the same time-interval in healthy controls.

Results:

Over treatment, PARS decreased significantly in both groups (F(2, 214)=82.84; p<.001, η_p²=0.44). Contrary to hypotheses, active ABMT did not enhance symptom reductions (F(2, 214)=0.61; p=.54, η_p²<0.01). There was no difference in the proportion of CGI-I treatment responders vs. non-responders between active and sham (X²=0.76, p=.38). Amygdala connectivity is reported, but should be interpreted with caution due to poor observed test-retest reliability in healthy controls (ICC(3,1)<0.4). Lower post-treatment anxiety was associated with weaker left-amygdala connectivity with several cortical regions on threat-congruent trials.

Conclusion:

Hypotheses that youth receiving active gamified ABMT+CBT would show greater improvement were not supported. Findings highlight the need to explore alternate attention-bias modification strategies for augmenting the clinical response to CBT for anxiety. Future research should continue to examine reliability of attention-bias tasks adapted for fMRI.

Clinical trial registration information:

A Clinical Trial of a Gamified Attention Bias Modification Training in Anxious Youth; https://clinicaltrials.gov/study/NCT03283930.

Study registration information:

The Nature of the Effect; https://osf.io/fxc54

INTRODUCTION

Exposure-based cognitive-behavioral therapy (CBT) produces meaningful reductions in anxiety for many youth.¹ However, ~50% of patients fail to respond.² Augmentation therapies could address this problem. While CBT targets explicit, goal-directed processes through exposure exercises and involves cognitive restructuring, attention bias modification training (ABMT) was developed to target a potential implicit mechanism of anxiety, attentional bias toward threat.³ Specifically, ABMT trains participants to attend away from threats. CBT and ABMT may target different aspects of anxiety, and thus may have complementary effects when combined.^4–6 Indeed, some research has demonstrated an enhanced effect of CBT combined with ABMT on youth anxiety.^4,7–9 This preregistered,¹⁰ double-blind, randomized controlled trial examines the efficacy of a novel, gamified ABMT + CBT on anxiety symptoms among youth with anxiety disorders.

Imaging studies relate anxiety to functional connectivity between the amygdala and frontal regions during threat attention tasks.⁴ However, mixed findings exist on ABMT efficacy,¹¹ with some studies finding no differences in efficacy between active and sham ABMT.¹² This is possibly due to small sample sizes and low participant engagement, particularly in younger children. Indeed, low task engagement moderates cognitive training efficacy.¹³ Moreover, youth show poorer ABMT learning gains than adults.¹⁴ To address these limitations, the current trial recruited 121 youth patients and incorporated game-like elements into ABMT in an attempt to enrich engagement and enhance ABMT efficacy.

As preregistered,¹⁰ the study combines a modified dot-probe task¹⁵ and visual search task, as opposed to a traditional dot-probe-only-based ABMT.¹⁶ Gamification is described in the Methods. In active dot-probe-based ABMT, a target probe follows neutral rather than threat cues, thus implicitly training attention to enhance performance. ABMT is hypothesized to alter attentional bias through “bottom-up” mechanisms via salience contingencies.¹⁷ In the novel visual search component of the gamified ABMT, participants search for non-threat targets in the context of threatening distractors. A similar approach has been successfully used in other ABMT¹⁸ and related to activity in the fronto-parietal attention network.^19,20 By combining the dot-probe and visual search, which target different and overlapping neurobiology, the current gamified ABMT was expected to yield enhanced treatment effects compared to prior ABMT augmentation trials.⁴

The current study had three primary preregistered aims.¹⁰ First, this study aimed to extend the findings of White and colleagues¹⁶ by testing the efficacy of a novel, gamified ABMT augmenting CBT in a clinic setting for youth with an anxiety disorder. It was hypothesized that youth receiving active ABMT+CBT would demonstrate greater improvements, compared to youth receiving sham ABMT+CBT, on the Pediatric Anxiety Rating Scale (PARS) and the Clinical Global Impressions-Improvement (CGI-I) Scale. Second, it was hypothesized that baseline amygdala connectivity (attention bias contrast) would differ between pediatric anxiety patients and a healthy comparison sample.⁵ Third, it was hypothesized that baseline amygdala connectivity would moderate the association between ABMT condition and clinical outcomes. Lastly, this study aimed to explore additional mechanisms of treatment response, namely, pre-to- post-treatment changes in amygdala connectivity. The study was registered at Clinicaltrials.gov on September 14, 2017 (NCT00018057, NCT03283930) and at OSF on December 1, 2017 (https://osf.io/fxc54). The peer-reviewed study protocol was published online on August 7, 2019.¹⁰

METHOD

Participants

Treatment-seeking patients with an anxiety disorder and healthy comparison youths (8–17-years-old) were recruited. All participants had an IQ > 70 as assessed by the Wechsler Abbreviated Intelligence Scale-version 2,²¹ were fluent in English, and had no current use psychoactive substances or medication at enrollment. Presence/absence of psychiatric diagnoses were established by semi-structured clinical interview, the Kiddie Schedule for Affective Disorders and Schizophrenia for School-Aged Children-Present and Lifetime version (K-SADS).²² Healthy controls had no lifetime anxiety disorders. Patients were included if they had a primary diagnosis of generalized anxiety disorder, social anxiety disorder, and/or separation anxiety disorder. For detailed exclusion criteria, see Supplement 1, available online. All study procedures were approved by the National Institute of Mental Health (NIMH) Institutional Review Board. Parents and youths provided written informed consent and assent, respectively.

A total of 121 patients were randomized to receive active (n=59) or sham (n=62) ABMT; all received manualized CBT.²³ Patients, families, and participant-facing staff were masked to ABMT condition. Data were collected from February 2017 to August 2023. The flow of participants is illustrated in Figure 1. Of 121 patients, 84 participated in an fMRI visit (active: n=47; sham: n=37), as did 100 healthy youths, who provided amygdala connectivity comparison data.

Figure 1: Consolidated Standards of Reporting Trials (CONSORT) Diagram of Study Design.

Procedure

Primary Clinical Outcome Measures.

The Pediatric Anxiety Rating Scale (PARS) is a clinician-rated 50-item checklist (0–5 scale) of anxiety symptoms and functional impairment, based on child and parent report.²⁴ The PARS assesses seven subscales of global severity/impairment: 1) number of symptoms; 2) frequency of symptoms; 3) severity of distress associated with symptoms; 4) severity of physical symptoms; 5) avoidance; 6) interference at home; and 7) interference outside the home. The PARS total score (range=0–25) is the sum of subscales 2, 3, 5, 6, and 7. Higher scores indicate greater levels of anxiety, with scores greater than 11.5 reflecting the likely presence of an anxiety disorder. The PARS demonstrates moderate test-retest reliability,²⁴ and excellent interrater reliability.²⁵ In the current patient sample, baseline PARS internal consistency was acceptable (Cronbach’s α=.78).

The Clinical Global Impression of Improvement (CGI-I) scale is a single-item clinician-rated measure of global symptom improvement. Scores range from 1 to 8, with lower scores indicating greater improvement and 5 indicating no change. Ratings ≤ 3 at the end of treatment are considered “responders,” while ratings > 3 are considered “non-responders.”²⁶

Secondary clinical outcome measures are listed in Supplement 1, available online.

Cognitive Behavioral Therapy.

Patients received 12 weekly sessions of individual exposure-based CBT administered by a licensed clinical psychologist who specializes in youth anxiety. Clinicians received additional, specific training on the CBT protocol^27,28 during a one-day training at the Yale Child Study Center by W.K.S. Clinicians also received bi-weekly reviews with the study team principal investigator to discuss treatment progress, PARS ratings, and ensure quality and ongoing protocol adherence. CBT sessions lasted between 40–60 minutes, followed by additional home practice assignments to build on exposures completed during therapy. The first three sessions consisted of introduction to and psychoeducation of CBT and self-monitoring. Children were instructed to complete their first out-of-session exposure after the third week, followed by in-session exposures and cognitive restructuring starting at week four.

Attention Bias Modification Training.

Prior to beginning CBT in-session exposures (week four), patients were randomly assigned to receive nine weeks of either active or sham ABMT. ABMT was administered on a laptop (1440×900 pixel display) by trained research staff prior to each CBT session to avoid potential interference with in-session CBT exposures. Both active and sham ABMT conditions were embedded in a computer game consisting of the dot-probe and visual search game (10–16 minutes) described above.

In the dot-probe-based ABMT, during three 80-trial blocks, patients viewed vertical face-pairs with either threat and neutral faces (60 angry-neutral pairs), or two neutral faces (20 neutral-neutral pairs) for 500 milliseconds (ms). Next, a left- or right-facing arrow probe (< or >) appeared in the location of one of the faces, prompting participants to respond with left/right button press to indicate the arrow direction. Responses were followed by the presentation of a fixation cross for 500ms. Participants were instructed to respond as quickly and accurately as possible. In the active condition, the arrow always appeared behind the neutral face to train participants to direct attention away from angry-face cues. In the sham condition, arrow probes appeared equally behind angry and neutral faces. Faces were selected from the TAU-NIMH Attention Bias Measurement Toolbox–Picture Set A.

The ABMT visual search puzzle game included a level-structure comparable to real-life digital games designed to be engaging and motivating for youth (Figure 2). Participants completed two runs of this game at each session. Participants were instructed to find hidden numbers within visual scenes in a limited timeframe (i.e., Level 1: 10 numbers in 60s; Level 2: 15 numbers in 75s; Level 3: 20 numbers in 90s). The visual scenes depict six youth engaging in school-related activities. In the active condition, four youth display neutral expressions, and two youth display angry expressions. Target numbers were hidden distal to angry youths and proximal to neutral youths. In the sham condition, all six youth display neutral expressions.

Figure 2: Association of Posttreatment Anxiety Outcomes with Baseline Left Amygdala Seed Connectivity as a Function of Attention Orientation Condition.

Note: A) Left amygdala seed; B) Bilateral precuneus. PARSpost = Pediatric Anxiety Rating Scale (PARS) posttreatment anxiety rating at CBT session 12, centered on 0.

ABMT training included a performance-based point system for both active and sham conditions. Participants could earn up to 80 points in each dot-probe block with correct responses, and additionally could earn “time tokens” for quick correct responses. These tokens could be redeemed in the visual search for an additional 10s of search time. Additional detail are provided in Linke and colleagues.¹⁰

Functional Magnetic Resonance Imaging (fMRI) Attention Bias Assessment – Dot-Probe Task.

Patients completed a dot-probe fMRI task (240 trials), similar to the sham ABMT pre-treatment (baseline) and post-treatment. Healthy youth completed this task at the same 12-week time interval. Each trial began with a 500ms fixation cross, followed by two faces presented vertically for 500ms, and then a left- or right-facing arrow presented until participant response. The task includes three conditions in a random order: 1) 80 threat-congruent trials: the arrow is presented in the previously vacated location of the angry face; 2) 80 threat-incongruent trials: the arrow is presented in the previous neutral face location; and 3) 80 neutral trials: both faces were neutral. The primary contrast of interest was congruent vs. incongruent trials as a measure of attention bias. Faces were selected from the TAU-NIMH Attention Bias Measurement Toolbox–Picture Set B.

fMRI Data Acquisition Parameters.

Functional neuroimaging data were acquired on two identical 3T GE scanners (Waukesha, Wisc.) with 32-channel head coils. Functional image volumes were collected with a T2*-weighting gradient-echo pulse sequence (TR=2300ms, echo time=25ms, flip angle=50°, FOV=24 cm, matrix=96×96, 2.5×2.5×3mm resolution). A high-resolution T1 scan was acquired (MPRAGE: NEX=1, TE = min full; TI = 900ms, FOV=25.6cm, matrix = 256×176, bandwidth = 25.00Hz per 256 voxels). Full parameters for functional and T1 scans are in Figure S1, available online.

fMRI Data Preprocessing for Seed-based Functional Connectivity.

All preprocessing and group-level scripts are available online at GitHub (https://github.com/NIMH-SDAN/abmt-study). Functional neuroimaging data were preprocessed with a state-of-the-art pipeline using Analysis of Functional NeuroImages (AFNI, version 24.0.02)²⁹ software; see Supplement 1, available online. Seventy-one anxious and 87 healthy youth had useable imaging data for the first scan after applying motion and behavioral thresholds; 51 anxious and 45 healthy youth had useable imaging data for both scans.

Generalized psychophysiological interaction (gPPI) methods³⁰ were used to assess amygdala seed-based functional connectivity as a function of task condition. The Harvard-Oxford atlas was used to define right/left amygdala seeds (50% atlas probability – a common threshold for the amygdala³¹), which were resampled and intersected with a whole-brain mask (voxels where >90% participants had data). Separate models were created for the left and right amygdala. PPI modelled the product of the detrended, demeaned seed regressor and each psychophysiological event type (i.e., threat-congruent, threat-incongruent and neutral events; errors and omissions were coded on a separate regressor). First-level models included head motion censoring, task regressors, PPI terms for each event types, seed mean timeseries, baseline drift, and six head motion regressors.

Statistical Analysis

Clinical Data Analysis: Efficacy of ABMT.

As preregistered, intent-to-treat analyses evaluated all participants who enrolled in the study, regardless of study completion. Given prior comparable research demonstrating an enhanced effect of CBT combined with ABMT on youth anxiety (Cohen’s d = 0.51),⁴ the sample was powered to observe a medium effect size. Posttreatment PARS total scores were examined in linear mixed-effects models with ABMT group (active/sham) as a fixed between-participant factor and time (pre-, mid-, post-treatment; i.e., CBT week 3, CBT week 8, CBT week 12, respectively) as a within-participant factor. Participant was a random effect. Age and sex were covariates. Pretreatment PARS differences between conditions were examined using a t-test to determine whether baseline anxiety should be a covariate in alternative models. ABMT efficacy was determined based on the group-by-time interaction and specifically, the post-hoc contrast of group differences in posttreatment PARS. Additionally, the number treatment responders (CGI-I score ≤ 3) and non-responders (CGI-I score >3) at the end of treatment were compared between ABMT conditions using a chi-square test. Additional analyses exploring treatment response on secondary clinical outcome measures are contained in Supplement 1, available online.

fMRI Data Analysis: Reliability of Seed-based Connectivity Measures.

First, we assessed test-retest reliability of right and left amygdala PPI in healthy comparison youth over the 12-week interval. Voxel-wise intraclass correlation coefficients were calculated by applying linear mixed-effects models with a random effect for participant and a fixed effect for time (ICC[3,1]). A Bayesian framework was utilized that modeled participant with Gamma priors (shape=2, rate=0.5) for each task contrast (incongruent vs. congruent; congruent and incongruent vs. neutral; congruent, incongruent, and neutral vs. baseline). The ICC[3,1] formulation assesses consistency in the rank of values, which accounts for systematic changes over time through repeated exposure to the task. An ICC value <0.4 is indicative of poor reliability, between 0.4–0.6 moderate, 0.6–0.75 good, and >0.75 excellent.³²

fMRI Data Analysis: Treatment Effects.

Analyses examined how amygdala connectivity at baseline: 1) differed between patients and healthy comparisons, 2) associated with patient treatment response, and 3) differed as a function of ABMT condition. In an exploratory analysis, we 4) examined treatment-related change in amygdala-seed connectivity from pre- to post-treatment. Exploratory analyses are in Supplement 1 (Table S1–S2, available online) alongside a region-of-interest (ROI) approach testing for significant results in the insula to directly test replication of prior findings,⁴ as well as explorations of other attention-related ROIs. Each whole-brain, voxel-wise analysis used a voxel-wise p<.001 threshold, cluster corrected to a p<.05 familywise error rate resulting in a cluster size of k ≥ 23, per 10,000 Monte Carlo simulations (3dClustSim -acf).

Pre-treatment Baseline Differences.

Baseline connectivity differences between patients and healthy controls were examined with a linear mixed-effects model using AFNI’s 3dLME program, with group (patients vs. healthy controls) as a between-participant variable and task condition (threat-congruent, threat-incongruent, neutral) as a within-participant variable. Age and sex were included as covariates.

Associations with Treatment Response.

Analyses utilized AFNI’s 3dMVM program to test whether pre-treatment amygdala connectivity associates with treatment response in patients. Post-treatment PARS ratings were included as a continuous variable, ABMT group (active/sham) as the between-participants variable and task condition as the within-participants variable. Age, sex, and pre-treatment PARS ratings were included as covariates. Post-hoc analyses used general linear tests (Wald X²) of specific contrasts or fixed effects of any variable while adjusting for all others [phia package³³ in R]. In-scanner behavioral data (reaction time) were analyzed using the same models as the imaging analyses. Data cleaning was consistent with prior work.^4,34

RESULTS

Efficacy of ABMT: Primary Clinical Outcomes

See Table 1 for sample sociodemographic and clinical characteristics. Of the 121 youth who were randomized, 48 completed the allocated active condition and 49 completed the sham condition. There was relatively little missing or incomplete data for the PARS (pretreatment=5.0%; mid-treatment=6.6%; posttreatment=7.4%). On average, patients completed 7 out of 9 ABMT sessions. There was no difference in the number of sessions completed between active vs. sham ABMT groups, t(119) = −0.7, p = .50. At baseline, there were no differences between active and sham groups in baseline age, t(119)=0.93, p=.36; or sex, X² =0.02, p=.89. Pre-treatment PARS ratings did not differ between groups at baseline, t(116) = −1.32, p=.19, and thus, was not included as a covariate in these models.

Table 1:

Sample Characteristics and Descriptive Statistics of Baseline Clinical Variables

	Total Anxiety Sample (N = 121)	Active ABMT (n = 59)	Sham ABMT (n = 62)	HV Sample (N = 90)
Demographics
Age (years) [M (SD)]	12.83 (2.64)	13.05 (2.55)	12.63 (2.70)	13.62 (2.71)
Sex, n (% female)	73 (60.33)	36 (61.02)	37 (59.68)	54 (60.00)
Race, n (%)	-	-	-	-
American Indian/Alaskan Native	1 (0.83)	1 (1.69)	0 (0.00)	1 (1.11)
Asian	5 (4.13)	1 (1.69)	4 (6.45)	7 (7.78)
Black or African American	4 (3.31)	3 (5.08)	1 (1.61)	18 (20.00)
Multiple Races	21 (17.36)	11 (18.64)	10 (16.13)	12 (13.33)
White	87 (71.90)	41 (69.49)	46 (74.19)	41 (45.56)
Unknown	3 (2.47)	2 (3.39)	1 (1.61)	11 (12.22)
Ethnicity, n (% Hispanic/Latine)	17 (14.04)	9 (15.25)	8 (12.90)	4 (4.44)
Clinical Variables
Baseline PARS [M (SD)]	14.23 (3.93)	13.70 (4.10)	14.73 (3.74)	N/A
End-of-treatment PARS [M (SD)]	9.78 (4.58)	9.67 (5.02)	9.89 (4.13)	N/A
CGI-I, n (% treatment responders at end-of-treatment)	80 (66.12)	37 (62.71)	43 (69.35)	N/A
SCARED-Child [M (SD)]	31.18 (15.43)	31.71 (14.19)	30.66 (16.64)	8.58 (8.30)
SCARED-Parent [M (SD)]	29.37 (11.83)	28.41 (10.87)	30.47 (12.86)	4.70 (5.17)
STAI-C [M (SD)]	39.22 (8.73)	39.90 (7.86)	38.60 (9.47)	26.59 (6.05)
CGAS [M (SD)]	54.39 (7.11)	54.39 (6.91)	54.39 (7.36)	N/A

Note: Baseline clinical values presented, unless otherwise noted. ABMT = Attention bias modification training; CGAS = Children’s Global Assessment Scale; CGI-I = Clinical Global Impressions – Improvement score; HV = healthy volunteers; M = mean; PARS = Pediatric Anxiety Rating Scale; SCARED = Screen for Anxiety and Related Disorders; SD = standard deviation; STAI-C = State-Trait Anxiety Inventory for Children.

PARS scores decreased significantly over the course of CBT for all anxiety patients, F(2, 214)=82.84, p<.001, η_p² =0.44. There was no main effect of ABMT condition (active/sham), F(1, 118) = 1.43, p=.23, η_p² =0.01, and no significant condition-by-time interaction, F(2, 214) = 0.61, p=.54, η_p² <.01. Likewise, there were no significant ABMT condition differences in the number of treatment responders compared to non-responders post- CBT, X² =.76, p=.38.

Reliability of Amygdala-Seed Connectivity

Among n=45 healthy control youth with usable data at two scans, poor test-retest reliability was observed for all three task contrasts (voxel-wise whole-brain ICCs (3,1) < 0.4), with minimal clusters containing an ICC value >0.4.

Pre-treatment Baseline Connectivity and Behavioral Differences

Baseline amygdala-seed connectivity was compared between groups (n=71 anxiety; n=87 healthy). No clusters surpassed the whole-brain-corrected p<.001 threshold on any task conditions for the left or right amygdala. Behaviorally, there were also no significant differences between patients with anxiety and healthy volunteers at baseline in attention bias reaction time scores during the fMRI task, F(1, 154) = 0.12, p = .73, η_p² < 0.01, adjusting for age and sex.

Associations with Treatment Response

Significant results, including post-hoc statistical tests, are summarized in Figure 2 and Tables 2–3. Lower levels of post-treatment anxiety (i.e., more improvement, controlling for pretreatment anxiety) were associated with weaker left amygdala connectivity with the bilateral precuneus and right cuneus on threat-congruent, but not incongruent trials. No significant clusters emerged for right amygdala connectivity (Table 2).

Table 2:

Summary of Significant Results for the Anxiety-by-Condition Interaction Assessing Overall Treatment Response^a

PARS posttreatment * Condition
LEFT AMYGDALA
k	mm3	CM RL	CM AP	CM IS	Mean	SEM	F	df	p	Con- InCon: Wald X2	Con- Neutral: Wald X2	InCon- Neutral: Wald X2	Locationb
48	750	−8.6	71.3	38.9	190.94	2.50	13.93	118	<.001	12.48***	26.66***	2.66	R precuneus, R cuneus, L precuneus
RIGHT AMYGDALA
k	mm3	CM RL	CM AP	CM IS	Mean	SEM	F	df	p	Con- InCon: Wald X2	Con- Neutral: Wald X2	InCon- Neutral: Wald X2	Location
--	--	--	--	--	--	--	--	--	--	--	--	--	No significant clusters

Note: AP = anterior-posterior direction; CM = center of mass; Con = congruent; InCon = incongruent; IS = inferior-superior direction; k = number of cluster voxels; mm3 = cluster volume; PARS = Pediatric Anxiety Rating Scale; RL = right-left direction; SEM = standard error of the mean.

^aCluster-corrected voxel-wise linear mixed-effects model results are presented, summarizing regions showing an anxiety-by-condition interaction.

^bLocation of clusters with >5.0% overlap listed, according to CA_ML_18_MNI atlas.

^***p<.001.

Table 3:

Summary of Significant Results for the Two-Way Interactions Assessing ABMT-Specific Treatment Associations^a

PARS posttreatment * Condition
LEFT AMYGDALA
k	mm3	CM RL	CM AP	CM IS	Mean	SEM	F	df	p	Con- InCon: Wald X2	Con- Neutral: Wald X2	InCon- Neutral: Wald X2	Locationb
89	1391	−13.1	66.7	35.1	138.38	3.59	17.53	114	<.001	18.46***	32.23***	1.91	R precuneus,R cuneus
23	359	31.6	−54.2	−4.1	133.83	6.36	14.08	114	<.001	0.27	23.35***	18.63***	L middle orbital gyrus, L superior orbital gyrus
RIGHT AMYGDALA
k	mm3	CM RL	CM AP	CM IS	Mean	SEM	F	df	p	Con- InCon: Wald X2	Con- Neutral: Wald X2	InCon- Neutral: Wald X2	Location
--	--	--	--	--	--	--	--	--	--	--	--	--	No significant clusters
Condition * ABMT Group
LEFT AMYGDALA
k	mm3	CM RL	CM AP	CM IS	Mean	SEM	F	df	p	Con: Active Sham: t-statistic	InCon: Active Sham: t-statistic	Neutral: Active Sham: t-statistic	Locationb
30	469	−12.4	65.3	33.3	184.3	6.10	12.97	114	<.001	0.09	3.52***	5.33***	R precuneus
RIGHT AMYGDALA
k	mm3	CM RL	CM AP	CM IS	Mean	SEM	F	df	p	Con: Active Sham: t-statistic	InCon: Active Sham: t-statistic	Neutral: Active Sham: t-statistic	Location
--	--	--	--	--	--	--	--	--	--	--	--	--	No significant clusters

Note: ABMT = attention bias modification training; AP = anterior-posterior direction; CM = center of mass; Con = congruent; InCon = incongruent; IS = inferior-superior direction; k = number of cluster voxels; mm³ = cluster volume; PARS = Pediatric Anxiety Rating Scale; RL = right-left direction; SEM = standard error of the mean.

^aCluster-corrected voxel-wise linear mixed-effects model results are presented, summarizing regions showing an anxiety-by-condition or group-by-condition interaction.

^bLocation of clusters with >5.0% overlap listed, according to CA_ML_18_MNI atlas.

^***p<.001.

No significant three-way interaction of task condition by ABMT group by posttreatment PARS emerged for the left or right amygdala. A significant two-way condition by PARS rating interaction emerged in several cortical regions (Table 3). Post-hoc analyses revealed that lower levels of post-treatment anxiety were associated with weaker left amygdala connectivity with the right precuneus and right cuneus on threat-congruent compared to incongruent trials, and stronger left amygdala connectivity with the left middle orbital gyrus and left superior orbital gyrus on neutral trials. A significant two-way interaction of ABMT group by condition surpassed the defined threshold in a cluster in the right precuneus, which is not discussed given these are baseline group differences. No significant clusters emerged for right amygdala connectivity for these two-way interactions. Behaviorally, there were no significant differences between ABMT groups on posttreatment attention bias reaction time score, F(1, 54) = 0.25, p = .62, η_p² < 0.01, adjusting for age and sex. All imaging data for primary analyses are available online at OpenNeuro (https://openneuro.org/datasets/ds006577).

DISCUSSION

This study produced three main findings. First, although patients showed improvement, active versus sham gamified ABMT did not enhance CBT reductions in anxiety. Second, at study entry, patients did not differ from healthy comparison youth in behavioral reaction times or amygdala connectivity during the dot-probe task. Third, across ABMT conditions, greater improvement in posttreatment anxiety related to weaker baseline left amygdala connectivity with several regions on threat-congruent task trials. However, reliability of connectivity in healthy youth was poor, and baseline amygdala connectivity in patients did not moderate active ABMT effects on posttreatment response.

Clinical and Behavioral Differences

Given the heterogenous nature of threat responses,³⁵ ABMT was designed through multiple studies to target implicit components of threat processing that are less targeted by CBT alone.⁷ However, initially supported findings from this literature were not observed in this current study. Research on the clinical efficacy of adjunctive ABMT among youths is mixed, ranging from no ABMT effect³⁶ to small-to-medium significant effects.^{37–40 16} White et al.¹⁶ observed a trend for group differences in the primary outcome measure, PARS, with a medium effect size, informing the pre-registration for the current study.¹⁰ Nevertheless, no such differences emerged in this current study using a novel, gamified ABMT, contradicting the pre-registered hypothesis. Similar to White and colleagues¹⁶, no group differences were observed in the primary categorical outcome measure, CGI-I. Lack of significant findings also align with previous randomized controlled studies of ABMT in pediatric populations,^7,41 which found no specific effect of training group on clinician-rated anxiety symptoms.

Of note, all patients in the current study received individual CBT, which demonstrates moderate-to-large effect sizes (Cohen’s d = 0.63–0.85) for reductions in pediatric anxiety symptoms.⁴² As expected, in the current study, a large and clinically significant effect of CBT treatment on PARS ratings was observed. CBT has also been shown to produce threat-related changes in attention bias on its own.⁴³ Indeed, explicit components of CBT, such as exposure exercises and cognitive restructuring, are designed to test beliefs and promote new learning gains. Hence, robust clinical effects of CBT on symptom outcomes may have reduced sensitivity to any additional small effect size of treatment augmentation between the active ABMT and sham group. Yet, mean post-treatment PARS scores remained relatively high, suggesting that room for improvement remained. Thus, while CBT might reduce sensitivity, it did not create a floor effect in the current study. Indeed, one three-armed randomized controlled trial has shown that both active and placebo ABMT significantly improved treatment response compared to CBT-alone.⁷ Although there was no comparison group receiving CBT-alone in the current study, the lack of between-group differences could reflect expectancy effects associated with receipt of treatment. However, a strength of the current design is that it maximizes the sample size in two treatment cells equated for expectancy effects. Regardless, the absence of findings for both primary (clinician-report) and secondary (child- and parent-report) outcome measures suggests that the gamified ABMT did not augment the CBT response in the current pediatric anxiety sample.

There were no baseline group differences in reaction time between the healthy control and anxiety groups or between the active and sham ABMT groups. This result aligns with White and colleagues.¹⁶ The reliability of the reaction time-based attention bias score has been called into question for the dot-probe task.^44–46 The dot-probe task demonstrates, at best, poor-to-modest reliability.^41,44,45 Behavioral manifestations of reaction time can arise from many factors, such as response selection latency, that do not pertain to an underlying attentional threat bias neurocognitive construct.⁴⁴ Moreover, ceiling effects may have obscured behavioral results given the dot-probe task is not highly cognitively demanding. Such limitations may partially explain between-group differences in neural responding despite no such behavioral differences between groups.

Amygdala Connectivity

Disentangling underlying neural dysfunctions of attention bias using imaging approaches can better clarify the intersection of attentional bias and anxiety. Neural indices are often thought to be more reliable and sensitive than behavioral reaction-time based measures of attention bias,⁴⁷ although some prior sample sizes examining test-retest reliability have been small (e.g., n = 12).⁴¹ Results showed that, at study entry, patients with anxiety did not differ from the healthy control group in amygdala connectivity during the dot-probe. Disparate findings in the current study, compared to other literature,^4,48,49 are likely due to psychometric reliability limitations of the short event-related design of the dot-probe paradigm. Indeed, in the current study, test-retest reliability of all three neural contrasts was poor among healthy volunteers. These findings suggest lack of promise for the current approach in assessments of event-specific connectivity. Despite the common practice of performing functional connectivity analyses with brief events, trial-specific functional connectivity is more appropriately modeled in tasks with longer events; this can reduce the chance that results reflect undue effects of time signal outliers.^30,50 Functional connectivity examines correlation of BOLD response between two time-courses in two discrete locations. Accordingly, short trial events reduce the possibility of detecting robust, reliable signal correlations. As with other recent findings, the current results highlight the need to pay greater attention to psychometric properties of neural markers.

Regarding associations between amygdala connectivity and posttreatment response among patients, several findings were observed. Greater improvement in posttreatment clinician-reported anxiety symptoms was associated with weaker left amygdala connectivity on threat-congruent, but not incongruent, trials with the bilateral precuneus, and right cuneus. It is possible that these connectivity patterns may have relevance for predicting CBT response among youth with anxiety disorders and may be worthy of further exploration in future studies. However, considering the lack of significant clinical effects across ABMT groups, together with the aforementioned psychometric limitations of the dot-probe paradigm, this specific pattern of neural findings remains of questionable relevance for future work.

In summary, the current study found no ABMT group differences in clinician-rated or questionnaire measures of anxiety symptoms, thus limiting the clinical relevance of imaging results. The study also found no behavioral group differences in attention bias to threat, thus failing to demonstrate any relevance in behavior evoked by the dot-probe task.

This study has several key limitations. Importantly, the novel, gamified ABMT version used in the current study has not been previously tested or validated as either an augment to manualized CBT or a standalone treatment for anxiety. Namely, the study combined a modified dot-probe task¹⁵ and visual search task, as opposed to a traditional dot-probe-only-based ABMT.¹⁶ Although these elements were pre-registered and intended to maximize treatment engagement among a pediatric sample,¹⁰ methodological differences between this novel paradigm and a standard dot-probe ABMT may account for the failure to replicate past work.⁴ Other limitations involve generalizing findings to various patient groups. This includes patients receiving medication or suffering from conditions that excluded participants from the current study, such as mood-related or attention deficit comorbidities. Racial/ethnic diversity was somewhat limited, particularly in the patient group, further reducing generalizability. Given that ABMT was not able to be offered virtually, findings may be limited for those without accessibility to in-person treatment. Additionally, all patients received CBT plus the active or sham ABMT, precluding a comparison of results for those not receiving any treatment (i.e., waitlist control) or those receiving CBT-alone (i.e., a third-arm comparison). Treatment expectancy was not assessed at baseline and may be an area of investigation for future work. Given the pre-registered sample size, analysis of different subtypes of specific anxiety diagnoses was not possible, thus somewhat limiting specificity. Most notably, the reliability of the connectivity measures derived from the fMRI dot-probe task was poor, as mentioned above, further limiting any meaningful interpretation of significant imaging findings.

Despite these limitations, this randomized controlled study design had important strengths and contributes to the broader attention bias literature. This was the largest randomized controlled trial to-date of anxious youth that utilized fMRI to assess attention bias. The sample size was adequately powered to detect a medium effect size. Further, objectives, hypotheses, analyses, and methodology were publicly pre-registered and published,¹⁰ increasing transparency and reducing the potential for bias. Moreover, randomization allocation was double blinded to participants and clinicians. Notably, there was relatively little missing data or treatment dropout. Finally, the use of independent clinician-based assessments of anxiety as the primary outcome measure, rather than self-report, further reduces reporting bias.

Importantly, future research should continue to improve the reliability of attention bias indices adapted for fMRI. Researchers should strive to use reliable imaging approaches to delineate relevant neural and behavioral anxiety-related mechanisms to contribute to the development of personalized treatment selection. Use of alternate imaging methods, such as task-based activation or magnetoencephalography,^51–53 could be approaches to explore in future research. Moreover, threat sensitivity and attentional control are ABMT targets worthy of further exploration, in line with prior research.^54,55 Overall, current findings suggest further opportunities to refine and rigorously test novel gamified versions of ABMT.⁵⁶ Future work should continue to research youth-friendly ABMTs that provide additive value as an adjunct to a treatment as robust as CBT. Given the current results, moving beyond the tasks used in the current study appears necessary to advance research on psychopathology-related attention biases. One such promising approach that has yet to be explored as an augmentative treatment to CBT is gaze-contingent music reward therapy (GCMRT). Recent data has suggested this treatment may be more engaging than reaction-time-based tasks and better capture different attentional components related to anxiety.^57–59 Notably, GCMRT utilizes an eye-tracking task as a direct measure of attention bias that has been shown to be reliable in youth samples.⁶⁰ Considering inconclusive findings in the current study, further work is needed to more fully describe the predictive relation between pretreatment attentional threat bias, amygdala connectivity, and treatment response to ABMT.

In summary, the main clinical hypotheses were not supported and did not align with the clinical findings of White et al.¹⁶ – youth receiving active ABMT+CBT did not differ from the sham group on anxiety symptom improvement. Lack of clinical group differences merits caution in subsequent interpretation of imaging-related findings. Moreover, reliability of the fMRI attention bias paradigm was poor in the current study, highlighting the importance of ensuring reliable measurement of neural markers. Future research should continue to examine the reliability of attention bias tasks adapted for fMRI and explore the efficacy of alternate, reliable treatments.

Data Sharing:

In accordance with the NIH Data Management and Sharing Policy, participant’s individual data will be shared in a repository for those participants that consented to share their data publicly. Given not all participants consented to share their data, exact replication of analyses with the open dataset will not be possible. Data will be available immediately following publication and there will be no end date. Additionally, all code used for analyses for this data will be shared on GitHub.