Computerized Interpretation Bias Training for Disruptive Mood Dysregulation Disorder: A Fast-Fail Study

Abstract

Objective:

Probing targeted, mechanism-based interventions is the next generation of treatment innovation. Biased threat labeling of ambiguous face emotions (interpretation bias) is a potential behavioral treatment target for anger, aggression, and irritability. Changing biases in face-emotion labeling may improve irritability-related outcomes. Here, we report the first randomized, double-blind, placebo-controlled targeted trial (RCT) of interpretation bias training (IBT) in youth with chronic, severe irritability.

Method:

Patients with current Disruptive Mood Dysregulation Disorder (DMDD; N=44) were randomly assigned to complete four sessions of active (n=22) or sham (n=22) computerized IBT training within a one-week period. The first and last trainings were completed onsite, and two trainings were completed at home. We examined the effects of active IBT on labeling bias, primary outcome measures of irritability, and secondary outcome measures of anxiety, depression, and functional impairment. Follow-up assessments were completed immediately after the intervention and two weeks following the intervention.

Results:

We found that active IBT engaged the behavioral target in the active relative to the sham condition, shown by a significant shift toward labeling ambiguous faces as happy. However, there was no consistent clinical improvement in active IBT relative to the sham condition, immediately after, or two weeks following training in either primary or secondary outcome measures.

Conclusion:

While this RCT of IBT in youth with DMDD engaged the proposed behavioral target, there was no statistically significant improvement on clinical outcome. Identifying and changing behavioral targets is a first step in novel treatment development; these results have broader implications for targeted-based intervention development.

Clinical trial registration information:

Psychological Treatments for Youth With Severe Irritability; https://clinicaltrials.gov/; NCT02531893.

Introduction

Identifying and changing targeted, mechanism-based deficits has been proposed as the next generation of treatment innovation ^1,2. Irritability is a common pediatric psychiatric problem ^3,4 with very few non-pharmacological, mechanism-based, empirically supported treatments ⁵. Recent work reveals a key behavioral target that provides a promising avenue for treatment. Specifically, irritability and its related constructs are associated with a form of biased face-emotion labeling ^6-10, which can be altered through cognitive training ^11,12. In the present study, we report the first randomized controlled trial (RCT) of a face-emotion cognitive training as a potential targeted treatment for pediatric irritability.

A number of psychosocial, evidence-based interventions exist to treat externalizing behaviors such as aggression, oppositionality and impulsivity across various pediatric diagnoses. Parent Management Training ¹³, Cognitive Behavioral Therapy ^14,15, Dialectical Behavior Therapy ¹⁶ and Interpersonal Therapy ¹⁷, either address forms of anger and temper outbursts characteristic of DMDD, or have been adapted to include these symptoms and behaviors as targets. While these interventions have shown some efficacy in treating anger and behavioral outbursts ¹⁸, they can be costly and time-intensive; here we test a computer-delivered training program to complement existing efforts to develop comprehensive psychosocial and pharmacological options.

Hostile interpretation bias (HIB), the proneness to interpreting ambiguous social cues as threatening, is thought to enhance threat salience, increase anger, and motivate reactive aggression ^6,8,19-23. Given this potentially causal role for threat biases in anger and externalizing behaviors, research on face-emotion labeling may provide an avenue for novel therapies ²⁴. Training protocols have been devised to investigate whether these biases can be modified and whether doing so decreases anger and aggression. Specifically, interpretation bias training, “IBT”, is a computer-based training protocol aiming to change participants’ labeling of ambiguous face-emotions towards more benign (“happy”) interpretations.

Several studies have examined the effects of computerized cognitive training programs targeting hostile attributions and interpretations in youth with internalizing and externalizing difficulties ^21,25-28. An open label trial established feasibility and acceptability of IBT for youth with impairing irritability i.e., a lifetime diagnosis of DMDD ⁶. An early study found that IBT in youth at risk of conduct problems resulted in decreases in self- and blinded staff-rated anger and aggression ²¹. Change in HIB in the experimental condition was related to change in state anger, lending support to the hypothesized mechanism. Similarly, an open label trial in youth with irritability showed decreased symptoms alongside change in HIB and a relationship between cognitive and symptom change. A small number of more recent studies have failed to find clinical improvement for social anxiety, anger, or aggression despite target engagement, i.e., a shift towards labeling ambiguous face-emotions as happy ^27,29-31. Hence, across studies, active IBT successfully shifts labeling of face-emotions through personalized corrective feedback. Training effects appear to be sustained beyond training sessions across samples. Findings regarding the effects of this shift on symptoms and behavior are less consistent, but studies especially in developmental populations are limited, with divergent samples and outcome measures.

Many new treatments are ineffective or suboptimal, which has led researchers to propose the “fast-fail approach” ¹ to accelerate the pace of therapeutic discovery. In a fast-fail approach, a small trial is performed with an identified treatment target to establish proof of mechanism thought to mediate the clinical effect. Both target engagement and predefined efficacy, i.e., evidence for the specific therapeutic target effect rather than non-specific improvements, are required before a larger, definitive trial is conducted. If either efficacy or target are not supported, the fast-fail trial should provide information for treatment development. The fast-fail approach was developed for small-scale studies of medications ^32,33; this approach has yet to be applied systematically to cognitive training. Given research linking malleable forms of biased face-emotion labeling to irritability, here we apply this fast-fail approach to a computer-based training. Such an application is particularly needed in youth, given concerns about pediatric psychopharmacology, especially potential endocrine and metabolic effects in certain classes of psychotropic medications ³⁴, and poorly understood effects on the developing brain.

In summary, prior data lay the groundwork for an RCT targeting biased face-emotion labeling to change irritability. Here, we employ a fast-fail approach via an RCT in youth with chronic, severe irritability; we test the hypothesis that active relative to sham cognitive training produces greater changes in biased labeling of ambiguous face-emotions and decreases symptoms of irritability.

Method

Sample

The study was conducted at the National Institutes of Health with active enrollment from November 2015 to December 2018. Study procedures were approved by the National Institute of Mental Health (NIMH) Institutional Review Board (IRB). This trial was registered under NCT02531893 on clinicaltrials.gov on August 25, 2015 (https://clinicaltrials.gov/ct2/show/record/NCT02531893), and the proposed study protocol for this trial is published ³⁵. Participants were recruited nationwide via IRB-approved mailings and websites.

Patients were eligible if they met full, current Disruptive Mood Dysregulation Disorder (DMDD) criteria as assessed by a structured clinical interview Kiddie-Schedule for Affective Disorders and Schizophrenia Present and Lifetime Version; K-SADS-PL; ³⁶ including a supplementary module to assess the presence of DMDD ³⁷. Additionally, to be eligible to enroll, patients’ response to their current community treatment was assessed by the research team to be no more than minimal. This was formally operationalized by a CGI-S score (1-month version) of ≥3 focusing on DMDD symptoms ³⁸. Parents and youth provided informed consent and assent, respectively. Forty-four patients were enrolled in the study (see Table 1 and 2 for demographics and clinical characteristics of the sample and Table S1, available online, for additional measures of anxiety, depression, and functional impairment by training group). Participants were randomized to receive four IBT training sessions (active vs. sham training) in a double-blind design over four days. Participants completed clinician- and child-report ratings pre- and post-training and two weeks following training. The endpoint for the primary outcome measures was the first follow-up visit. See Figure S1, available online, for a CONSORT diagram detailing the flow of participants through the study as well as Supplement 1, available online, for additional information on recruitment, eligibility criteria and randomization procedures.

Table 1.

Participant Characteristics N=44

	DMDD
	N=44
Age (years) M (SD)	11.96 (2.11)
Sex % male	57%
Race N (%)
Black or African American	1 (2)
White	39 (89)
Multiple races	3 (7)
Unknown	1 (2)
Ethnicity N (%)
Latino or Hispanic	6 (14)
Unknown	1 (2)
IQ M (SD)	111.4 (12.5)
K-SADS diagnoses, current N (%)
Any anxietya	27 (61)
ADHD	32 (73)
DMDD	44 (100)
MDD	0 (0)
Medicationsb N (%)
Antidepressants	20 (46)
Antipsychotics	10 (23)
Antiepileptic	6 (14)
Stimulants	24 (55)
Psychotherapyc N (%)
Individual therapy	17 (39)
Group therapy	1 (2)

Note: ADHD = attention-deficit/hyperactivity disorder; DMDD = disruptive mood dysregulation disorder; K-SADS = Kiddie Schedule for Affective Disorders and Schizophrenia for School-Aged Children; MDD = major depressive disorder

^aAny anxiety disorder includes generalized anxiety disorder, social phobia, specific phobia, and/or separation anxiety disorder

^bData was not available from 2 participants.

^cRegularity varied from weekly to monthly.

Table 2.

Pretraining Demographic and Clinical Characteristics by Training Group

	Active (n=22)	Sham (n=22)
Age (years)	11.48 (2.05)	12.44 (2.11)
Sex (% male)	16 (72.73)	9 (40.91)
IQ	113.91 (11.64)	108.95 (13.19)
CL-ARI	58.33 (11.15)	61.62 (13.95)
ARI-C	3.05 (2.32)	4.64 (2.77)
ARI-P	8.73 (1.88)	8.05 (3.05)

Note: n = 1 missing Clinician Affective Reactivity Index (CL-ARI). ARI-CL = Affective Reactivity Index-Clinician; ARI-C = Affective Reactivity Index-Child; ARI-P = Affective Reactivity Index-Parent.

Measures

Primary and secondary outcome measures were completed at baseline, post-training, and follow-up. Ratings were administered by trained masters- or doctoral-level clinicians. Primary outcome measures assessed changes in irritability using the Clinician Affective Reactivity Index (CL-ARI) ³⁹, Clinical Global Impressions Scale-Improvement for DMDD (CGI-I) ³⁸, and parent- and child-report Affective Reactivity Index (ARI-C/P) ⁴⁰. Secondary outcome measures were ratings of anxiety, rated by clinicians (Pediatric Anxiety Rating Scale, PARS ⁴¹) and the parent/child (Screen for Childhood Anxiety Related Emotional Disorders-parent/child, SCARED ⁴²), as well as measures of depression assessed via clinician- (Children’s Depression Rating Scale, CDRS ⁴³) and child-report (Children’s Depression Inventory, CDI ⁴⁴).

Additionally, clinician-rated measures of severity (Clinical Global Impressions Scale-Severity for DMDD, CGI-S ⁴⁵) and impairment (Children’s Global Assessment Scale, CGAS ⁴⁶) were collected. For additional measures assessed at baseline see Table S2 and Supplement 2, available online.

Interpretation bias assessment and training

The Interpretation Bias Task (IBT Task; see Figure 2 for task schematic) estimates and trains happy-angry interpretations of face-emotions ⁴⁷. The stimulus set for assessment and training consisted of 15 face-emotion pictures linearly morphed at equal intervals from unambiguously happy to unambiguously angry facial expressions, created from a happy and angry composite face from the Karolinska Directed Emotional Faces database ⁴⁸. Stimuli were first generated by Penton-Voak and colleagues ²¹ using established procedures designed to maintain relevant features and textures of facial expressions ⁴⁹.

Figure 2. Balance Point by Training Group and Session.

Note: Error bars show standard error.

Assessment Task. Each trial consisted of a fixation cross (1500-2500 ms), face presentation (150 ms), visual noise mask (250 ms), and response prompt (no time limit). Participants entered a forced-choice response of “happy” or “angry” via keyboard button press. Face-morphs were presented in random order. The Assessment Task measured participants’ “balance point” (i.e., the morph along the continuum at which the participant shifts his or her judgment from predominantly happy to predominately angry). The balance point is estimated as the proportion of happy responses to total responses multiplied by 15 during the 45 trials of the assessment task. A low balance point indexes a hostile interpretation bias, a balance point of 8 indicates no bias (middle of morphs 1 to 15). Assessment tasks were completed before and after each training and at the two follow-up assessments.

Training Task. The training task was identical to the assessment task, but feedback was provided to the participant on every trial after the participant’s response. Feedback consists of a message saying “Right!/Wrong!”. In sham IBT training, feedback was based on the participant’s balance point, as measured using the assessment task before training (i.e., the feedback is consistent with the participant’s pre-training interpretation bias). In the active training, feedback was designed to shift the participant’s balance point two face-morphs along the face-morph continuum, so that two ambiguous morphs that were rated angry at pre-training would be rated as happy post-training. The maximum training threshold was face-morph 12, as positive feedback associated with the three most extreme angry morphs was always congruent with an angry interpretation. There were 6 training blocks of 30 trials each.

Sample size and statistical analysis

Power calculations were based on Stoddard and colleagues (2016) who completed an open, uncontrolled arm pilot trial in youth with DMDD which found a large effect of IBT on outcomes (d=1.2 for CGI-I for symptoms of DMDD). However, randomized, controlled designs of similar computer-based approaches have reported more moderate effect sizes ⁵⁰, which likely relates to the presence of a control condition in these studies. Thus, we adjusted our expected effect size to d=0.9 for the primary outcome. At an accrual target of 20 per group and an alpha of 5%, this study has 85% power to detect an effect of this size. Dropouts were replaced for a minimum N=40 of participants who completed the trial. Data were analyzed according to the intention-to-treat (ITT) principle and included all patients who underwent randomization. One participant randomized to active IBT withdrew assent after the first training session. For this participant, where baseline scores were available, these were carried forward and imputed at all time points for any corresponding outcome measures. Interpretation bias and outcome measures were analyzed as a function of training (active vs. sham) and session (pre-training, post-training, one week and two weeks follow-up) using a linear-mixed effects model with participant as a random factor and group and session as a fixed factor. Missing datapoints were rare and not imputed (<2.5%, <1.5%, and <5% of primary outcome, secondary outcome, and interpretation bias measures, respectively).

In secondary analyses, we used Spearman correlations to examine i) associations between pre-training balance point and baseline measures of irritability and ii) associations between shifts in balance point pre- to post-training and irritability-related clinical improvement (CGI-I) or reductions in irritability at the primary outcome time point (i.e., 1-week post-training). Additionally, we used a mediation analysis to test whether change in balance point from baseline to post-training (i.e., first pre-training assessment at day 1 to post-training assessment at day 4) mediated clinical improvement (CGI-I) at the primary outcome time point by training group. All statistical analyses were performed using R statistical software (Version 3.5). The package lme4 was used for linear mixed-effects modeling ⁵¹ and lavaan for mediation analyses ⁵².

Results

Balance point (interpretation bias)

The Session-by-Training interaction was significant (F(3,121.92)=6.33, p<.001, see Figure 2). As expected, there was no difference in mean balance point between IBT conditions pre-training (Active: b[SE]=7.42[0.50], Sham: b[SE]=7.18[0.50], t(41.98)=0.49, p=0.63, d=0.15). However, at post-training assessment, patients in the active IBT condition had a significantly higher balance point (i.e., more happy judgements on ambiguous morphs) compared to the sham condition by b[95%CI] =2.51[0.90, 4.12] face-morphs (t(40.57)=3.16, p=0.003, d=0.96). The active IBT group also had a significantly higher mean balance point at both the 1-week (by b[95%CI] =1.99[0.44, 3.56] face-morphs, t(40.31)=2.59, p=0.01, d=0.79) and the 2-week follow-up (by b[95%CI] =2.33[0.82, 3.83] face-morphs, t(38.17)=3.12, p=0.003, d=0.96), indicating maintenance of balance point change post training completion. Sham training balance points remained stable across all four training sessions (ICC[2,1]=.72).

Primary outcome measures

The main effect of Training and the Session-by-Training interaction term were not significant for clinician-rated improvement scores (CGI-I scores all F<0.90, ps>.34; see Figure 3). For both training groups, CGI-I scores post-training were in the “slightly improved” range (Active: M[SE]=3.95[0.32], Sham: M[SE]=3.67[0.32], t(38.04)=0.60, p=.55, d=.19). CGI-I scores at 1- and 2-weeks post-training were similarly in the “slightly improved/improved” range for both training groups (1-week follow-up: Active: M[SE]=3.71[0.32], Sham: M[SE]=3.52[0.32], t(36.92)=0.44, p=.66, d=.14; 2-week follow-up: Active: M[SE]=3.95[0.32], Sham: M[SE]=3.48[0.32], t(39.99)=1.10, p=.28 d=.34).

Figure 3. Irritability Outcome Measures.

Note: (A) Clinician- and (B) parent-rated irritability by group and session. (C) Improvement scores relative to baseline by group (active vs sham interpretation bias training [IBT]) and session (post-training and 1- and 2-weeks following training; 5 = No Improvement). Right panel: Error bars show standard error.

The Session-by-Training interaction was not significant for either clinician- or parent-rated irritability (ARI-CL: F(3,123)=0.67, p=.55; ARI-P: F(3,122.92)=1.11, p=.35). Training was associated with a reduction in clinician- and parent-reported irritability symptoms across both training groups (ARI-CL: F(3,123)=15.94, p<.001; ARI-P: F(3,122.92)=7.06, p<.001). Pre-training ratings of clinician- and parent-reported irritability symptoms were ARI-CL M[SE]=60.01[1.92], and ARI-P M[SE]=8.39[.29]. Ratings reduced significantly, on average ARI-CL: M[95%CI] =16.83[11.61, 22.04] and ARI-P: M[95%CI] = 2.11[1.21, 3.02] points respectively by 1-week post training (ARI-CL: t(42)=6.52, p<0.001, d=.93; ARI-P: t(43)=4.70, p<0.001, d=.74). Child-rated irritability showed a significant Training-by-Time interaction (F(3,123.03)=3.18, p=.03). Post-hoc analyses indicated marginally significant decreases in child-rated irritability over time in the sham condition (F(3,61)=2.72, p=.05), but not in the active IBT condition (F(3,62)=1.40, p=.25). However, further analyses showed that training group differences in child-rated irritability were only significant at baseline, with marginally higher irritability reported by the sham group (baseline: t(40.7)=2.07, p=.05, d=.62).

Secondary outcome measures

None of the secondary outcome measures showed improvements specific to the active IBT training group, see Supplement 3, available online, for detailed analyses of secondary measures.

Associations between balance point and outcome

We examined associations between pre-training balance point and clinician-, parent- and child-report irritability. We did not find significant correlations between baseline irritability measures and pre-training balance point (all rs<−.22, ps>.15) or shifts in balance point and irritability reductions (all rs<−.08, ps>.60) or irritability-related clinical improvement (r=.05, p=.73). An additional mediation further investigated the relations between target engagement and outcome. The indirect effect of balance point change as a mediator in the relation between training group status and irritability-related clinical improvement was not significant (z=−.57, p=.57).

Discussion

This RCT of computer-based IBT engaged the behavioral target in the active relative to the sham condition, shown by a significant shift toward labeling ambiguous faces as happy; however, active IBT produced no detectable clinical improvement over and above improvement in the sham condition. Thus, while the target was engaged, and the effect size of this change was large (d=.96 after 4 days of training, d=.79 1-week post training), there was no consistent clinical improvement associated with target engagement.

Fast-fail approaches aim to accelerate the pace of discovery via small scale, iterative, rigorous testing of target-based treatments ^32,33. The negative result is consistent with the mixed findings in prior work on IBT. While the sample size in this trial was small, the change in the behavioral target was large; considering this change in the behavioral target, we would expect differential clinical improvement or associations between target engagement (change in labeling) and improvement even with this modest sample size. More consistent efficacy might accrue following refinements in specific recruitment features by first identifying individuals who might most likely benefit from face-emotion labelling training.

IBT is a basic category learning task, where both sham and active IBT train towards reliable judgments of facial affect at a specific target balance point, or category boundary. At this boundary, participants are trained to shift from one response abruptly and reliably to another. Without an active comparator, for example, training individuals on nonsocial features of objects, it is not clear whether the general improvement across arms is attributable to category learning on facial affect or common reasons for improvement participation in clinical trials, such as regression to the mean or demand effects.

The study suggests that behaviorally training a shift in balance point does not lead to clinical improvement in youth with irritability post-training or within the two weeks follow-up period. It is possible that shifts in face-emotion labeling need longer to consolidate before impacting symptom reports ⁵³. Findings support the need to integrate multiple features (i.e., behavioral deficits, neural correlates, phenotypic specificity) when developing novel targets. Behavioral deficits alone may not be sufficient as a treatment target as there may be multiple neural pathways to a common behavioral deficit. Lessons may be learned from RCTs of attention bias modification training (ABMT) as a treatment for anxiety. Similar to attentional processes ⁵⁴, biases in interpretation related to anger appear only moderate in size ²², based on the very limited data available. Thus, a significant proportion of patients may present without this bias toward ambiguous social cues. Future work could select participants based on their initial bias and ability to incorporate trial-by-trial feedback to update their categorical labeling. Selecting participants based on these features may create a more homogenous treatment group more likely to benefit from this targeted cognitive training.

Several limitations need to be noted. First, the RCT is underpowered to detect small effects or examine efficacy in subgroups of youth with more specific characteristics (e.g., current anxiety disorder, specific age range). Second, we used the same face-emotion stimulus set across all training units; a composite face averaged across different male actors with different apparent races and features. Hence, youths were trained on specific face-stimuli, rather than on face-emotions more generally. However, recent work has provided first evidence that training effects on specific faces generalize to identities and facial configurations beyond those included in the training ⁵⁵. Future work should use several different sets of face-emotion stimuli, which may reduce fatigue, and increase transfer to new social material. Third, while community treatment was kept stable throughout the duration of the RCT, participants had a variety of concurrent treatments including psychotropic medication.

Results from this and other fast-fail RCTs might discourage further investments in targeting of behavioral biased face processing. However, IBT involves complex processes such as attention orienting, updating, and decision-making, and each of these parameters could be further explored; post-hoc analyses on data from this report may be informative. Work on other computer-based interventions suggests that it is important to fully couple behavioral dysfunction with specific neural targets ⁵⁴.

Aberrant threat interpretation has been suggested as a cognitive pathway to emotional problems across internalizing and externalizing phenotypes. An RCT of IBT in youth with DMDD found no significant differential improvement on clinical outcome.

Figure 1. Interpretation Bias Training (IBT).

Note: (A) Illustration of a training trial. (B) Active IBT was designed to shift the balance point two morphs along the continuum of face-morphs. Hence, by the end of the training session, participants were expected to rate two ambiguous morphs as happy that were rated angry pre-treatment. Sham IBT provided feedback consistent with the pre-training balance point.