Individual Differences in Response to Prediction Bias Training

Abstract

Depressed and dysphoric individuals predict negative outcomes more often than healthy individuals, and more frequently than positive outcomes. However, it is unclear if there is a causal relationship between negative emotional processing biases and depressive symptoms. We examined whether prediction-targeted neurocognitive training changed symptoms, behavior, and physiological reactivity to positive or negative feedback in euthymic and dysphoric undergraduates. Participants were randomized to a positive training intervention or neutral training. Among participants who received the positive training, pupillary reactivity to negative feedback predicted decreased symptoms and decreased after each session. Neither effect was present in the neutral training group. Both groups were more likely than controls to predict positive outcomes after training. Data suggested that change in a physiological mechanism indexing emotional reactivity and change in depressive symptoms during a targeted intervention may be related, and that it is possible, using psychophysiology, to predict which individuals will respond to neurocognitive intervention.

Keywords

neurocognitive training positive affect dysphoria pupil dilation

The past decade has seen increased emphasis on targeting specific aspects of information processing in personalizing psychiatric treatment using neurocognitive training interventions (Siegle, Ghinassi, & Thase, 2007). The underlying principle is that information processing disruptions are not only present in disorders such as depression, but also are causal in their maintenance. Repetitive training exercises geared at changing information biases are hypothesized to address underlying neural substrates of the biases. Here we test this principle in a domain that has not been explored with mechanistic intervention—the tendency to predict negative outcomes. We examined causal associations of information processing with symptoms by assessing (a) whether a neurocognitive training intervention geared to change prediction biases can change both behavioral and physiological expression of the biases and (b) whether observed symptom change is predicted by pretreatment behavioral bias and physiological reactivity. The first question is important in showing causally that change in mechanism could yield change in symptoms. The second, which is less frequently addressed, shows that targeting a mechanism is most important for people with abnormalities of the proposed mechanism prior to intervention.

We specifically targeted the tendency, in depression, to predict negative outcomes more often than healthy controls do and to predict more future negative events than positive events (Dunning & Story, 1991; MacLeod & Byrne, 1996; Siegle, Ingram, & Matt, 2002; Wisco & Nolen-Hoeksema, 2010). This prediction bias may act to maintain negative thinking in depression by conditioning individuals to expect negative outcomes, interpret ambiguous information as negative, and finally attend to negative information as confirmatory. This maladaptive response could play a role in developing and maintaining common depressive symptoms and behaviors, such as expecting negative outcomes, interpreting outcomes as failures, and experiencing feelings of hopelessness (Tucker, Luu, Frishkoff, Quiring, & Poulsen, 2003).

Prediction biases were addressed in this study using neurocognitive training exercises. Such interventions have recently been used to target substrates of other emotional information processing biases. For example, attention bias modification trains individuals who preferentially attend to negative information to disengage from it (e.g., Bar-Haim, 2010). We have used cognitive control training exercises to increase executive control yielding decreases in rumination and improving depressed subjects’ mood (Calkins, McMorran, Siegle, & Otto, 2014; Papageorgiou & Wells, 2000; Siegle et al., 2007; Siegle, Price, Jones, Ghinassi, & Thase, 2014). These studies suggest a promising and plausible mechanistically targeted avenue for treatment or adjunctive interventions for affective disorders that could be coupled with traditional psychotherapy or pharmacotherapy (Siegle et al., 2007). To date, prediction biases have not been targeted with such interventions; here, we addressed this gap.

Neurocognitive exercises designed to target a mechanism are only expected to work for individuals who have abnormalities in that specific mechanism (Siegle et al., 2014). Furthermore, relationships between pretreatment bias and response have rarely been explored, and when they have been explored these relationships often do not hold up (Amir, Taylor, & Donohue, 2011). Rather, physiological indices of information processing have been shown to predict response to validated common interventions that would be expected to target them, such as cognitive therapy (e.g., Siegle, Steinhauer, Friedman, Thompson, & Thase, 2011). Physiological indices of information processing have been associated with response to targeted interventions (Siegle et al., 2014), but such associations have not been replicated. Thus, in addition to considering whether this novel intervention for a specific information processing bias can be successful, we consider whether underlying mechanisms, assessed physiologically, predict its success.

We were specifically interested in dynamic recruitment of cognitive or affective processes underlying emotional judgment as a possible mechanism underlying intervention effects. Toward this end, we examined pupil dilation while participants underwent neurocognitive training. The pupil serves as an index of both cognitive and emotional processing (Beatty, 1982; Beatty & Lucero Wagoner, 2000; Siegle, Steinhauer, Carter, Ramel, & Thase, 2003; Siegle, Steinhauer, & Thase, 2004). Pupillary motility is associated with activity in brain regions subserving both emotional processing such as the amygdala (Koikegami & Yoshida, 1953) and regions involved in emotion regulation and cognitive control such as the dorsolateral prefrontal cortex (Siegle, Steinhauer, Stenger, Konecky, & Carter, 2003). Numerous studies have shown reliable pupillary differences between participants with depressive features and healthy controls in response to emotional information processing (for review, see Graur & Siegle, 2013). For example, in response to emotional words, depressed participants show an increased peak and sustained pupil dilation compared with healthy controls (Johnstone, van Reekum, Urry, Kalin, & Davidson, 2007; Siegle et al., 2011; Siegle, Granholm, Ingram, & Matt, 2001; Siegle, Steinhauer, Carter, et al., 2003). The pupil also responds to changes in task contingencies associated with feedback (Gilzenrat, Nieuwenhuis, Jepma, & Cohen, 2010; Lavin, San Martin, & Rosales Jubal, 2014; Mossbridge, Tressoldi, & Utts, 2012). Assuming prediction training is mechanistically targeted, we expected pretreatment deficits to be related to symptom change. Thus, we selected pupil dilation as a possible predictor of an individual’s response to prediction training. We have successfully used pupillary responses to intervention-relevant tasks as a predictor of treatment response in similar previous studies (Siegle et al., 2011; Siegle et al., 2014).

Nondysphoric controls were assessed to create a “normal” or standard response, both behaviorally and physiologically, that was used to determine what was “abnormal” in the dysphoric participants. We expected that some of our dysphoric participants would display either abnormally high or abnormally low pupillary reactivity, with respect to nondysphoric controls. To promote prediction training as targeting a relevant mechanism, consistent with the National Institute of Health’s goal of “experimental therapeutics” (Insel, 2014), it was important to examine the extent to which prediction training “normalized” such abnormal responses. In addition, nondysphoric controls’ behavioral responses to the novel training intervention served as the benchmark for what could be considered good performance on our novel training task.

In addition to expecting change in symptoms, we expected change in the frequency of predicted positive outcomes. We also considered the extent to which effects generalized, that is, whether changing biases on one task was associated with change in other frequently observed information processing biases in depression, under the idea that an effective training should change the way emotional information is processed, not just the way one responds to a specific task. In particular, we considered if this training could change processing of emotional information in a separate yet similar emotional processing task. We have previously found delayed reaction times in participants with depressive features in response to positive words when compared with reaction times to negative words and to nondepressed controls’ reaction times (Siegle, Ingram, et al., 2002). Therefore we examined both pre- and posttraining effects on a well-established word valence identification task (VID) as well as the prediction training task to explore the effects of prediction bias training on other emotional information processing.

Ultimately, the goal was to reduce symptoms of depression in participants who had a prediction bias at study onset and received the positive training condition of the paradigm. We intended to achieve this through altering both the behavioral and the physiological expression of their prediction bias.

Method

Participants

A total of 420 students in an introductory psychology class voluntarily participated in an online screening for dysphoria using the Quick Inventory for Depressive Symptomatology (QIDS; Rush et al., 2003), for which they received course credit. Students were eligible to be in the dysphoric group if they scored greater than 8, which equates to a score of 11 on the 17-item Hamilton Rating Scale for Depression (HAM-D17) and were eligible to be in the nondysphoric control group if they scored less than 4, which equates to a score of 5 or 6 on the HAM-D17.

Following the online screening, eligible nondysphoric control and dysphoric participants were contacted by email and invited to participate. Students who were interested underwent a brief phone screen, and eligible individuals were scheduled for lab visits no more than one week after phone screening. Subjects with history of head trauma and seizures were excluded. For their participation in this study, students received course credit. This protocol was approved by the University of Pittsburgh Institutional Review Board (IRB # PRO08020308).

The recruited participant set included 43 students, of whom 1 dysphoric was excluded for scoring below the cutoff on the depression screening at the baseline visit, and 1 participant’s data were not saved correctly, yielding a final analyzed set of 41 participants (age M = 18.88 years old, SD = 1.94), 12 of whom were male; 37 identified themselves as Caucasian, 1 identified as African American, 2 as Asian, and 1 listed other. In all, 23 subjects were included in our dysphoric group (age M = 19.26 years old, SD = 2.49), 20 Caucasian, 6 male (QIDS M = 11.83, SD = 2.89). A total of 18 subjects were in our nondysphoric control group (age M = 18.39 years old, SD = 0.61), 17 Caucasian, 6 male (QIDS M = 2.22, SD = 0.73). Data regarding socioeconomic status were not collected in this study. Of the dysphoric participants, 11 received the neutral training condition (age M = 18.36 years old, SD = 0.51), 9 Caucasian, 3 male (QIDS M = 10.55, SD = 1.92), and 12 dysphoric participants received the positive training condition (age M = 20.08 years old, SD = 3.26), 11 Caucasian, 3 male (QIDS M = 13.00, SD = 3.19). Eight nondysphoric controls returned for training and follow-up assessments (age M = 18.25 years old, SD = 0.46), 7 Caucasian, 3 male (QIDS M = 2.13, SD = 0.84), and 10 nondysphoric controls completed baseline measures only (age M = 18.80 years old, SD = 0.63), 10 Caucasian, 3 male (QIDS M = 2.30, SD = 0.68).

Materials and procedure

Participants completed four separate training sessions over the span of 2 weeks. Before beginning tasks on the first and final sessions, subjects completed the Diagnostic Inventory for Depression (DID; Zimmerman, Sheeran, & Young, 2004)) as a measure of depression symptoms and psychosocial functioning. The DID is a valid measure of symptom change and is highly correlated with Structured Clinical Interview for DSM-IV Axis I Disorders interviewer ratings. Participants completed 192 trials of the prediction task on each of three separate visits. A fourth visit, scheduled 2 weeks after the initial visit, served as a follow-up where an abbreviated version of the prediction task was completed (96 trials to assess presence of bias).

Prediction task

In this emotional information processing task, subjects were told that they were to predict the positivity or negativity of statements people might make about them. The task was based on another probabilistic feedback task in which participants learned to predict statements about the weather (Gluck, Shohamy, & Myers, 2002; Knowlton, Mangels, & Squire, 1996; Knowlton, Squire, & Gluck, 1994). The probabilities for each stimulus were explicitly unknown to the participants. Through trial and error, these probabilities should be somewhat understood at the conclusion of the first training session (192 trials).

Participants viewed an initial stimulus (shape card) for 2,000 ms and were asked to make a prediction by pressing a key on a number pad. Feedback was presented for 4,000 ms, which displayed the accuracy of their prediction as well as the emotional statement associated with the trial (see Fig. 1). Each statement began with “You are . . .” and was paired with either a positive or negative adjective. The adjectives used in the statements were rated as highly positive or highly negative in a separate study (Bradley & Lang, 1997). A total of 80 positive and 80 negative words were used in the feedback and were drawn randomly for each trial. Three different shape cards served as stimuli and were associated with a unique probability. By manipulating the probabilities associated with the stimuli, we were able to create two separate conditions. In one condition, the net total of positive and negative statements was equal (75%, 50%, and 25% probability of a positive statement being associated with each individual cue). In another condition, the probabilities were adjusted to create a strong positive skew (70% positive overall; 90%, 70%, and 50% for individual cues). This was designed to decrease a negative prediction bias in the dysphoric group. To increase accuracy, participants had to predict more positive outcomes than negative outcomes. These two conditions established the neutral training condition (equal outcomes) and the positive training condition (positively skewed).

Fig. 1.

Experimental design. Participants first viewed a shape cue for 2,000 ms (left column) and were told to press a key to predict a positive or negative outcome (middle column). After the response, feedback was presented for 4,000 ms (right column), showing both a positive or negative statement and accuracy feedback (thumbs up or down). The rows of the figure represent three different possible configurations of cues and outcomes given that each of the three shape cues was associated with a different probability of a positive outcome, which was unknown to the participant. In the training condition, these were 90%, 70%, and 50%. In the neutral training condition, these were 75%, 50%, and 25%.

Eye tracking and pupil data were collected by an ISCAN RK 464 pupillometer. Pupil size was recorded at a rate of 60 Hz (or 16.7 ms) and transferred digitally from the pupillometer to a computer that stored the raw data along with markers for trial onset, stimulus onset, and reaction time. Stimuli were displayed in dark gray on a light gray computer screen to minimize pupil dilation to changes in illumination associated with stimulus onset and offset.

Valence identification task

To determine if modification of one bias generalized to another, we also administered a valence identification task (as in Siegle et al., 2001) to participants at the initial and final training sessions. Participants completed 60 trials of 12 s each day. The stimuli for these trials consisted of 20 positive words, 20 neutral words, and 20 negative words. Words were balanced for normed valence, length, and word frequency using custom software (Siegle, 1994). Before each word was presented, a cue (dashes above and below a row of xs) was presented on the screen for 1,000 ms, signaling an upcoming word. Words were presented for 200 ms then masked with a row of xs, which remained on screen between trials. Subjects were instructed to press keys on a number pad corresponding to choices of positive, neutral, and negative and to respond as quickly and as accurately as possible. The buttons corresponding to positive, neutral, and negative were counterbalanced across subjects.

Analysis plan

Selection of stimuli for analysis

Trials with reaction times less than 150 ms were discarded as outliers because previous results suggest that reaction times in this range indicate that a response was made without regard for the stimulus (Matthews & Southall, 1991).

Calculation of pupil dilation indices

Data were cleaned using our lab’s standard methodology (Siegle, Ichikawa, & Steinhauer, 2008). Software written in Matlab indentified blinks as large changes in pupil dilation occurring too rapidly to signify actual dilation or contraction. Trials including more than 50% blinks were eliminated from further analysis. Where blinks occurred in the data set, linear interpolations replaced the missing data. Data were smoothed using a 10-point weighted average filter. Then, linear trends in pupil dilation calculated over blocks of 20 trials were removed from pupil dilation data to eliminate effects of slow drift in pupil diameter that were not related to trial characteristics. To calculate an index of pupil dilation differences, pupil diameter was measured as the average dilation over the 1 s preceding the onset of the stimulus, and this value was subtracted from pupil diameter after stimulus.

Primary analyses

Differences in reaction time, pupil dilation, and positive bias between groups over time was assessed using repeated measures tests with Hyunh-Feldt corrections for sphericity. Regressions using pupil dilation as the independent variable and symptom change as the dependent variable were used to understand the extent to which preintervention responses were associated with changes in symptoms. All tests were two-tailed. To detect a negative prediction bias, we examined the proportion of positive guesses in each group. Mixed effects tests with a heterogeneous compound symmetry covariance structure were used to determine if interactions were present between group status, word valence, and session on reaction time and pupil dilation in the valence identification task, given the potentially more complex covariance structure across word valences as well as session. All statistics were computed with SPSS software. An alpha level of .05 was used for each analysis.

Exploratory analyses of the time courses derived from pupil data used statistical tests at each sample along the condition-related mean pupil waveforms. Such a large number of tests requires Type I error control, so as in previous publications (Siegle et al., 2008) we used Guthrie and Buchwald’s (1991) technique. Using Monte Carlo simulations, we estimated the number of consecutive significantly different time points needed to achieve p < .05 significance given the highly autocorrelated waveform. For pupil dilation, windows of 35 consecutive significant tests (0.58 s) at p < .1 have reflected segments of significant differences at p < .05 for our equipment on similar tasks with large samples (Siegle et al., 2008).

Results

Data cleaning

Stimulus selection and data cleaning procedures resulted in the elimination of M = 11.75 (SD = 9.9) trials per subject for prediction training at Day 1, and M = 7.5 (SD = 5.5) trials dropped per subject at Day 4. For the valence identification task at Day 1, M = 1.5 (SD = 3.0) trials were dropped per subject and M = 12.1 (SD = 3.3) trials were dropped per subject at Day 4. Pupil data could not be collected from one control subject on Day 1. For the prediction task at Day 1, one dysphoric participant’s pupil data were not saved. For the prediction task at Day 4, one dysphoric participant was excluded as an outlier. Nine subjects’ (one nondysphoric control’s) reaction time data were not recorded properly for the valence identification task at Day 1, and five subjects’ not recorded at Day 4. One dysphoric participant’s DID was not saved on Day 4.

Predicting remediation of bias and symptoms

Preintervention behavioral biases

We predicted that at study onset, the dysphoric participants would have a negative prediction bias and therefore would choose more negative outcomes than the nondysphoric control group in each task condition. Indeed, dysphoric participants were more accurate, and thus more likely to choose negative outcomes in the neutral training condition (M = 0.48, SD = 0.07) than nondysphoric controls (M = 0.56, SD = 0.10). That said, dysphoric participants were also more accurate and thus less likely to choose negative outcomes in the positive training condition (M = 0.70, SD = 0.08) than nondysphoric controls (M = 0.66, SD = 0.05), Group × Condition F(1, 37) = 5.98, p = .02, η_p² = .14.

Symptom reduction

We predicted that the subjects who experienced training-induced changes in bias and in peripheral physiology would report a reduction in depressive severity, as measured by self-report (DID) administered before and after the training period. Contrary to hypotheses, group and training condition did not interact, F(1, 20.6) = 1.06, p = .32, η² = .05, though there was a significant main effect of session on DID score, F(1, 20.6) = 14.97, p = .001, η² = .42, with symptoms generally decreasing in both training conditions.

Subsequent analyses thus regarded differences between responders and nonresponders based on the mean decrease in symptoms, as measured by the DID (M = 11.30, SD = 13.26, Mdn = 9). Thus, participants whose DID score decreased by more than 11 points were considered dysphoric high responders (n = 10), and those whose DID decreased by 11 points or less (n = 13) was considered dysphoric low responders in further analyses. Had we used the median, 1 participant would have changed responder group; using the mean reflected a slightly more clinically significant cutoff.

Predicting response with physiology

There was an expected stimulus-related response in the pupil to making a prediction and viewing feedback (Figs. 2A, 2B). We expected more reactivity to task feedback in dysphoric high responders because they had a physiological expression of the bias addressed by the positive intervention. We expected that dysphoric participants in the neutral training group would have higher reactions than nondysphoric controls but would not differ based on DID scores because this training was not designed to target the mechanism sustaining a prediction bias.

Fig. 2.

(A) and (B) Response locked pupillary waveforms for negative correct trials on Day 1 as a function of outcome group (nondysphoric control, dysphoric high responders [large change in dysphoric symptoms] and dysphoric low responders [small change in dysphoric symptoms]) for (A) positive training group and (B) neutral training group. Gray shading along the x-axis marks areas of significant difference between the waveforms (light gray corresponds to p < .05; dark gray corresponds to p < .01). (C) Association of dysphoric participants’ mean pupil dilation during 3–4 s of feedback presentation (x-axis) with change in dysphoric symptoms (y-axis) in the neutral and positive training groups. Nondysphoric controls’ pupil reactions to negative correct feedback on Day 1 have been plotted as a reference with no symptom change included.

In the positive training condition, at their initial visit, dysphoric high responders had increased pupillary responses throughout the prediction period (Fig. 2A). This was not the case for the neutral training group (Fig. 2B). To understand this result for individual participants, area under the peak of the pupillary response curve (3–4 s following prediction onset) was compared against symptom change for each participant (Fig. 2C). For participants who received the positive training condition, a decrease in symptoms was correlated with higher initial reactivity to correct negative predictions, as measured by the mean pupil dilation in 3- to 4-s time period of the trial, when the feedback is displayed, r(11) = –.93, p < .0001. In the neutral training condition, there was no correlation of symptom reduction with reactivity, r(11) = –.04, p < .92. Due to the restricted range of reactivity within the neutral training group (Fig. 2C), even with these large differences in effects, a Pupil Reactivity × Task Type interaction was not statistically observed, ΔF(1, 17) = 1.49, p = .24, ΔR² = .06.

Due to our limited sample size, it was important to examine whether these correlations could have occurred by chance. As recommended for examining group differences in clinical trials with small samples (Ludbrook, 1994, 1995) we ran 1,000 simulations permuting group membership. The magnitude of the actual difference in correlations between the positive and neutral training groups (0.89) was achieved only once in the 1,000 permuted data sets, yielding an estimated p = .001 for the magnitude of condition-related differences in pupil-symptom associations.

Changing behavioral biases with training

Though the dysphoric subjects in the positive training condition did not display a statistically significant negative prediction bias at the start of training, their predictions became even more positive, and therefore farther from the ideal proportion of positive predictions, over the four training sessions, F(2.32, 23.24) = 4.61, p = .017 (Hyunh-Feldt corrected), η_p² = .32, reflecting a linear increase with session, F(1, 10) = 6.00, p = .03, η_p² = .38 (Day 4 M = 0.80, SD = 0.09). Dysphoric participants’ predictions in the neutral training condition also became more positive, and again, farther from the ideal proportion of positive choices, on the final training session F(2.06, 18.5) = 9.90, p = .001 (Hyunh-Feldt corrected), η_p² = .52, linear increase F(1, 9) = 15.85, p = .003 (Day 4 M = 0.60, SD = 0.08). The study was not powered to test a Group × Condition × Day interaction. These behavioral changes within both groups are illustrated in Figures 3A and 3B.

Fig. 3.

(A) and (B) Proportions of positive predictions made by dysphoric subjects in each training condition changed over time. Dysphoric participants in the positive training group steadily increased the number of positive predictions with each training session, slowly moving away from the correct proportion of 70% positive predictions, and making more positive choices than would be ideal for accuracy. In the neutral training group, dysphoric participants remained close to the ideal proportion of positive predictions (50%) but increased on the final day of training. For reference, nondysphoric controls’ positive predictions from Day 1 are plotted on the left side of each figure. In each training group, the dysphoric participants are closer to the ideal proportion of positive predictions than the nondysphoric controls are. (C) and (D) Physiological reactions to negative correct trials also changed with repeated training in dysphoric subjects. In the positive training condition, the group mean for pupil dilation to negative correct trials was elevated at Day 1 and decreased with each training session. In the neutral training condition, the mean pupil dilation in response to negative correct trials, was not elevated at Day 1, and the same pattern of decreasing reactivity was not observed. Nondysphoric controls pupillary reactions to negative correct feedback is plotted for comparison in the left side of each figure.

Changing physiological reactivity with training

There was a significant Session × Training condition effect of pupil dilation during negative correct-prediction trials, F(2.29, 50.1) = 3.57, p = 0.03 (Hyunh-Feldt corrected), η_p² = .17, which was driven by decreasing pupil in the positive training group but not in the neutral training group (Figs. 3C and 3D), Condition × Linear Trend contrast F(1, 18) = 10.80, p = .004, η_p² = .37, with a nonsignificant linear increase with day in the neutral training group, F(1, 9) = 3.70, p = .08, η_p² = .29, and a significant linear decrease with day in the positive training group, F(1, 9) = 7.10, p = .03, η_p² = .44.

Generalizability

Finally, we predicted that training geared toward changing prediction biases would lead to modification of other types of cognitive bias. A linear mixed model analysis with a heterogeneous compound symmetry covariance structure, which minimized the Bayesian information criterion estimates, was performed on both reaction time and pupillary data (3–6 s poststimulus corresponding to the peak interval), where valence was a repeated measure and both valence and dysphoric status were factors.

In the reaction time analysis, before treatment there was a main effect of dyphoria, F(1, 31) = 6.86, p < .0001, and of valence, F(2, 62) = 58.72, p < .001, but no Dysphoria × Valence interaction, p = .27, suggesting a general psychomotor slowing in the dysphoric participants (M = 2.27 s, SE = 0.054) compared with nondysphoric controls (M = 2.07 s, SE = 0.055). After training on the prediction task, reaction times in the dysphoric participants decreased, F(1, 56.95) = 26.2, p < .001, to approach controls (dysphoric session M = 2.08, SE = 0.046). Training condition and valence effects and interactions were not significant, all ps > .25. To examine the extent to which this increase could be due to practice effects, the small sample of nondysphoric controls (n = 7) who retook the valence identification task a second time was analyzed and also became faster, F(1, 36.07) = 12.00, p = .001 (M = 1.96, SE = 0.039).

For the pupillary response, there was a significant main effect of dysphoria F(1, 40.5) = 7.95, p = .007, reflecting overall larger responses in dysphoric participants (M = 0.19 mm, SE = 0.02) compared with nondysphoric controls (M = 0.09 mm, SE = 0.03), and word valence, F(2, 52.73) = 9.0, p < .005, reflecting larger responses to neutral, compared with positive (p < .0005) and negative words (p = .001), and a nonsignificant interaction between word valence and dysphoric status, F(2, 52.73) = 2.03, p = .14. After training there was a Session × Training Group interaction within dysphoric participants, F(1, 100.58) = 4.10, p = .045, such that pupil dilation decreased below that of nondysphoric controls for the neutral training group (M = 0.02 mm, SE = 0.03), but to the level of nondysphoric control participants in the positive training group (M = 0.11 mm, SE = 0.03). To examine whether this was possibly due to practice effects, the small number of nondysphoric controls who were retested also displayed decreased pupil dilation at Session 4 (M = 0.049 mm, SE = 0.02).

Discussion

The field of cognitive training for affective conditions has exploded in the past few years, with training paradigms geared toward increasing cognitive function (Owen et al., 2010), decreasing unconscious attention to emotional information (Bar-Haim, 2010), decreasing rumination (Siegle et al., 2007), and changing how people interpret emotional situations to be more positive (Holmes, Mathews, Dalgleish, & Mackintosh, 2006). The promise is that by identifying key features of emotional information processing, these can be targeted directly with rehabilitation-like exercises, rather than exclusively with more general and less mechanistically specific interventions, such as psychotherapy or medications. This study tested the effects of a novel neurocognitive training paradigm on a well-recognized phenomenon in affective disorders—negative prediction biases.

The design had a number of features we believe are critical for next-generation tests of translational interventions based on mechanism rather than possibly antiquated diagnostic schemes. First, to understand the potential contribution of the work for precision medicine, that is, targeting specific disease mechanisms, we considered the extent to which preintervention behavioral and physiological (pupil dilation) manifestations of a prediction bias (a tendency to select negative outcomes) predicted the degree of symptom reduction at the level of individuals. Second, to understand mechanistic specificity, both an active positive prediction training condition and a placebo neutral training condition were used, to allow differentiation between placebo and true cognitive training effects. Third, the sample was dysphoric undergraduates. Although this decision is often a liability in the context of clinical trials geared to understand and change diagnoses (Coyne, 1994), here, the use of an analog sample emphasized associations of symptoms, measured continuously, with the phenomenon, rather than with disorder labels per se. Fourth, we examined generalizability to a task outside the training set. This is critical for knowing whether the training affected task learning or a broader mechanism.

Predicting who will respond

The most prominent finding regarding response was that half of the participants, regardless of group, displayed decreased symptoms and very few increased. This is of interest in that college students are notoriously dysphoric (Coyne, 1994; Kendall, Hollon, Beck, Hammen, & Ingram, 1987) and because the study was recruiting for a period of several months, this was likely not a cohort effect. This result could be due to numerous nonspecific factors (i.e., clinical attention, task mastery, unrelated psychosocial improvements). It could also be that increased positive predictions were a common mechanism for both training conditions, as both conditions yielded an increase in positive prediction rates.

Yet the mechanism of change may have been differential across the groups as improvement in only the positive prediction training group was strongly associated with a physiological correlate of the targeted bias. This physiological reaction could represent a mechanism for maintaining a negative prediction bias. That is, cognitive biases are thought to be maintained by mechanisms that predispose depressed individuals to remember negative experiences more, encode them more deeply, or react more strongly to negative information (Siegle, Steinhauer, Thase, Stenger, & Carter, 2002). One way that a physiological mechanism of bias could work is to respond more intensely to negative information or error feedback (Tucker et al., 2003). High levels of physiological reactivity to correct negative predictions could represent hyperreactivity to negative feedback, consistent with confirming a negative expectation. It is people who have that mechanism that appear to respond to an intervention geared toward modifying the mechanism.

Although there was not sufficient power to detect difference in symptom decrease between the two training groups, there were differences in pupillary reactivity between individuals with higher and lower levels of symptom change within the positive training condition. The small sample precluded an omnibus test of a pupillary reactivity by training group interaction on symptom change. That said, uncertainty of this finding due to the small groups is somewhat reduced as permutation tests suggested that associations of symptom change with pupillary reactivity were stronger in the positive training than the neutral training group. These differences could suggest that positive training is effective for reducing symptoms of depression, but only for the individuals who initially display this initial physiological hyperreactivity to confirmatory negative predictions. In the future, it could be possible to use physiological measurement techniques, such as pupillometry, to identify certain individuals who might benefit from this type of targeted neurocognitive training. If the change in physiology is causal this would suggest that decreased reactivity to expected negative feedback could yield decreased negative thinking; that is, participants learned to react less to negative information with depressogenic thinking.

In addition to common nonspecific factors, the neutral training condition may have acted as an intervention for other nonmeasured reasons. This is likely, above and beyond common mechanisms, as the magnitude of symptom change was comparable across groups. Differential mechanisms would be consistent with other trials of neurocognitive training exercises which have also found that targeted trainings do not reduce symptoms in all participants, but rather only in those who strongly engage with the training task and that treatment-as-usual may better serve other participants who would not engage with the intervention (Siegle et al., 2014).

Changing behavior and physiology with training

Overall, dysphoric subjects were initially more accurate with their predictions when compared with nondysphoric controls, who tended to be more optimistic and make too many positive predictions in the neutral version of the task, but not make enough positive predictions in the positive training condition. These results may suggest that a task with as much positive feedback as participants received in the positive training condition may be too transparent. The version with the equal proportions may have allowed the subjects’ biases to be detected without influencing the number of positive predictions unnaturally.

Potentially because of the initial absence of a negative prediction bias, behavioral biases cannot be said to have “remitted” in the group trained on the positive version of the prediction task. Rather, the positive training condition was associated with the appearance of a positive bias in the dysphoric group, which is normally present in healthy subjects. This “optimism bias” is theorized to have protective effects (Sharot, Riccardi, Raio, & Phelps, 2007).

Together, these data suggest that to change a bias, it may be useful to modify the underlying mechanisms that maintain it. Indeed, hyperreactive dysphoric subjects’ physiological reactions to feedback changed over the course of the intervention. In the positive training intervention group only, pupillary reactions to correct negative predictions decreased over time. As there were no hyperreactive subjects in the neutral training group, it is unclear whether this is due to individual differences in our small sample of subjects, or if it was a function of the different training conditions. The relative scarcity of correct negative predictions in the positive training version of the task could have contributed to the large pupillary reactions we observed, because pupil dilation can increase in response to novel information (Steiner & Barry, 2011).

Training generalizes to separate biases

As predicted, before training, dysphoric subjects were initially significantly slower than nondysphoric controls and displayed larger pupil dilation on a word emotion-identification task. Several explanations might explain these biases which we have observed in depressed (Siegle et al., 2001) and dysphoric individuals (Siegle, Ingram, et al., 2002). It has been suggested that a cognitive bias exists which predisposes them to interpret all incoming information as negative (Beck, 2008). Thus, dysphoric individuals may have longer response latencies and greater pupil dilation as they process negative information in a ruminative and elaborative way. Alternately, participants may be cognitively inefficient with observed biases reflecting general processes such as psychomotor slowing. Training was associated with decreases in these abnormalities, though the extent to which the decreases reflected practice estimates is unclear.

Limitations

The sample for this pilot study was small, and thus, analyses of subgroups were likely underpowered. A replication with a larger sample would help determine if the lack of hyperreactive subjects in the neutral training condition was due to the individuals themselves or an effect of the task. Further replication may also prove useful in determining causality, and may be able to detect reliable differences in symptom change between the positive and neutral training groups. In addition, the extent to which results generalize to a clinical population is unclear as an analog sample of nondiagnosed dysphoric college students was tested. Some nondysphoric control subjects participated in only the baseline assessment and were not trained on the tasks. It would be necessary to gather data on more nondysphoric control subjects to rule out the possibility that the decreased reactivity we find when analyzing pupil dilation over time is not due to a general desensitization to these affective stimuli. Finally, training on the fourth day was shorter than on the other days, yielding a decreased dose that potentially accounts for the lack of change from Day 3 to Day 4 in physiological reactivity in the positive training group (Fig. 3C).

Summary and significance

These limitations notwithstanding, this study revealed several interesting trends and lends validity to the idea that negative biases found in individuals with depressive symptoms, such as a prediction bias, can be a feasible target for neurocognitive trainings like the one tested here. There was a physiological negative prediction bias in dysphoric young adults. Repeated training can modify this bias in those individuals who display it before training. When presented with an artificially positive scenario, these dysphoric young adults do not hesitate to predict positive outcomes more often than not, whereas their healthy counterparts do. Using the valence identification task, we demonstrated another facet of cognitive bias in depression that is slower and less efficient reactions to emotional information, which also appeared to change with cognitive training. This study corroborated its findings with physiological data and suggested that both behavior and physiology can change with training on cognitive tasks. Over the brief duration of this study, we detected the potential for symptom change.

To the extent that these results generalize, it could suggest that prediction biases are not just detectable epiphenomena of depression, but are causal in symptom maintenance. Changing these biases may be associated with symptom change. These data could thus suggest the utility of neurocognitive training interventions targeted at specific physiological mechanisms. A patient presenting at a future clinic could be assessed physiologically for prediction biases, and given a targeted intervention based on these.

Footnotes

Acknowledgements

The authors thank Alexandre Y. Dombrovski and Mark E. Wheeler for comments on an early write-up of this work, Agnes Haggerty for assistance in testing subjects, and the Program in Cognitive Affective Neuroscience members for comments and discussion of this work.

Declaration of Conflicting Interests

The authors declared that they had no conflicts of interest with respect to their authorship or the publication of this article.

Funding

This work was supported by MH082998 and 2R25 MH054318.

References

Amir

Taylor

C. T.

Donohue

M. C.

(2011). Predictors of response to an attention modification program in generalized social phobia. Journal of Consulting and Clinical Psychology, 79, 533–541. doi:2011-13209-001[pii]10.1037/a0023808.

Bar-Haim

(2010). Research review: Attention bias modification (ABM): A novel treatment for anxiety disorders. Journal of Child Psychology and Psychiatry, 51, 859–870. doi:10.1111/j.1469-7610.2010.02251.x JCPP2251[pii]

Beatty

(1982). Task-evoked pupillary responses, processing load, and the structure of processing resources. Psychological Bulletin, 91, 276–292.

Beatty

Lucero Wagoner

(2000). The pupillary system. In Cacioppo

J. T.

(Ed.), Handbook of psychophysiology (2nd ed., pp. 142–162). New York, NY: Cambridge University Press.

Beck

A. T.

(2008). The evolution of the cognitive model of depression and its neurobiological correlates. American Journal of Psychiatry, 165, 969–977.

Bradley

M. M.

Lang

P. J.

(1997). Affective Norms for English Words (ANEW) technical manual and affective ratings. The Center for Research in Psychophysiology University of Florida, Gainsville FL.

Calkins

A. W.

McMorran

K. E.

Siegle

G. J.

Otto

M. W.

(2014). The effects of computerized cognitive control training on community adults with depressed mood. Behavioural and Cognitive Psychotherapy. Advance online publication. doi:10.1017/s1352465814000046

Coyne

J. C.

(1994). Self-reported distress: Analog or ersatz depression. Psychological Bulletin, 116, 29–45.

Dunning

Story

A. L.

(1991). Depression, realism, and the overconfidence effect: Are the sadder wiser when predicting future actions and events? Journal of Personality and Social Psychology, 61, 521–532.

10.

Gilzenrat

M. S.

Nieuwenhuis

Jepma

Cohen

J. D.

(2010). Pupil diameter tracks changes in control state predicted by the adaptive gain theory of locus coeruleus function. Cognitive and Affective Behavioral Neuroscience, 10, 252–269. doi:10.3758/CABN.10.2.252 10/2/252[pii]

11.

Gluck

M. A.

Shohamy

Myers

(2002). How do people solve the weather prediction task? Individual differences in strategies for probabilistic category learning. Learning and Memory, 9, 408–418.

12.

Graur

Siegle

(2013). Pupillary motility: Bringing neuroscience to the psychiatry clinic of the future. Current Neurology and Neuroscience Reports, 13, 365. doi:10.1007/s11910-013-0365-0

13.

Guthrie

Buchwald

J. S.

(1991). Significance testing of difference potentials. Psychophysiology, 28, 240–244.

14.

Holmes

E. A.

Mathews

Dalgleish

Mackintosh

(2006). Positive interpretation training: Effects of mental imagery versus verbal training on positive mood. Behavior Therapy, 37, 237–247. doi:S0005-7894(06)00041-4[pii]10.1016/j.beth.2006.02.002

15.

Insel

(2014). A new approach to clinical trials. Director’s Blog. Retrieved from http://www.nimh.nih.gov/about/director/2014/a-new-approach-to-clinical-trials.shtml

16.

Johnstone

van Reekum

C. M.

Urry

H. L.

Kalin

N. H.

Davidson

R. J.

(2007). Failure to regulate: Counterproductive recruitment of top-down prefrontal-subcortical circuitry in major depression. Journal of Neuroscience, 27, 8877–8884.

17.

Kendall

P. C.

Hollon

S. D.

Beck

A. T.

Hammen

C. L.

Ingram

R. E.

(1987). Issues and recommendations regarding use of the Beck Depression Inventory. Cognitive Therapy and Research, 11, 289–299.

18.

Knowlton

B. J.

Mangels

J. A.

Squire

L. R.

(1996). A neostriatal habit learning system in humans. Science, 273, 1399–1402.

19.

Knowlton

B. J.

Squire

L. R.

Gluck

M. A.

(1994). Probabilistic classification learning in amnesia. Learning and Memory, 1, 106–120.

20.

Koikegami

Yoshida

(1953). Pupillary dilation induced by stimulation of amygdaloid nuclei. Folia Pychiatrica Neurologica Japonica, 7, 109–125.

21.

Lavin

San Martin

Rosales Jubal

(2014). Pupil dilation signals uncertainty and surprise in a learning gambling task. Frontiers in Behavioral Neuroscience, 7, 218. doi:10.3389/fnbeh.2013.00218

22.

Ludbrook

(1994). Advantages of permutation (randomization) tests in clinical and experimental pharmacology and physiology. Clinical and Experimental Pharmacology and Physiology, 21, 673–686.

23.

Ludbrook

(1995). Issues in biomedical statistics: Comparing means by computer-intensive tests. Australia and New Zealand Journal of Surgery, 65, 812–819.

24.

MacLeod

A. K.

Byrne

(1996). Anxiety, depression, and the anticipation of future positive and negative experiences. Journal of Abnormal Psychology, 105, 286–289.

25.

Matthews

Southall

(1991). Depression and the processing of emotional stimuli: A study of semantic priming. Cognitive Therapy and Research, 15, 283–302.

26.

Mossbridge

Tressoldi

Utts

(2012). Predictive physiological anticipation preceding seemingly unpredictable stimuli: A meta-analysis. Frontiers in Psychology, 3, 390. doi:10.3389/fpsyg.2012.00390

27.

Owen

A. M.

Hampshire

Grahn

J. A.

Stenton

Dajani

Burns

A. S.

. . . Ballard

C. G.

(2010). Putting brain training to the test. Nature, 465, 775–778. doi:nature09042[pii]10.1038/nature09042

28.

Papageorgiou

Wells

(2000). Treatment of recurrent major depression with attention training. Cognitive and Behavioral Practice, 7, 407–413.

29.

Rush

A. J.

Trivedi

M. H.

Ibrahim

H. M.

Carmody

T. J.

Arnow

Klein

D. N.

. . . Keller

M. B.

(2003). The 16-item quick inventory of depressive symptomatology (QIDS), clinician rating (QIDS-C), and self-report (QIDS-SR): A psychometric evaluation in patients with chronic major depression. Biological Psychiatry, 54, 573–583.

30.

Sharot

Riccardi

A. M.

Raio

C. M.

Phelps

(2007). Neural mechanisms mediating optimism bias. Nature, 450, 102–105.

31.

Siegle

G. J.

(1994). The balanced affective word list creation program. University of Pittsburgh, Pittsburgh PA. Available web: http://www.pitt.edu/~gsiegle/wordlist/index.htm

32.

Siegle

G. J.

Ghinassi

Thase

M. E.

(2007). Neurobehavioral therapies in the 21st century: Summary of an emerging field and an extended example of cognitive control training for depression. Cognitive Therapy and Research, 31, 235–262. doi:10.1007/s10608-006-9118-6

33.

Siegle

G. J.

Granholm

Ingram

R. E.

Matt

G. E.

(2001). Pupillary response and reaction time measures of sustained processing of negative information in depression. Biological Psychiatry, 49, 624–636.

34.

Siegle

G. J.

Ichikawa

Steinhauer

(2008). Blink before and after you think: Blinks occur prior to and following cognitive load indexed by pupillary responses. Psychophysiology, 45, 679–687.

35.

Siegle

G. J.

Ingram

R. E.

Matt

G. E.

(2002). Affective interference: An explanation for negative information processing biases in dysphoria? Cognitive Therapy and Research, 26, 73–88.

36.

Siegle

G. J.

Price

R. B.

Jones

N. P.

Ghinassi

Thase

M. E.

(2014). You gotta work at it: Pupillary indices of task focus are prognostic for response to a neurocognitive intervention for depression. Clinical Psychological Science, 2, 455–471.

37.

Siegle

G. J.

Steinhauer

S. R.

Carter

C. S.

Ramel

Thase

M. E.

(2003). Do the seconds turn into hours? Relationships between sustained pupil dilation in response to emotional information and self-reported rumination. Cognitive Therapy and Research, 27, 365–382.

38.

Siegle

G. J.

Steinhauer

S. R.

Friedman

E. S.

Thompson

W. S.

Thase

M. E.

(2011). Remission prognosis for cognitive therapy for recurrent depression using the pupil: Utility and neural correlates. Biological Psychiatry, 69, 726–733. doi:10.1016/j.biopsych.2010.12.041

39.

Siegle

G. J.

Steinhauer

S. R.

Stenger

V. A.

Konecky

Carter

C. S.

(2003). Use of concurrent pupil dilation assessment to inform interpretation and analysis of fMRI data. NeuroImage, 20, 114–124.

40.

Siegle

G. J.

Steinhauer

S. R.

Thase

M. E.

(2004). Pupillary assessment and computational modeling of the Stroop task in depression. International Journal Psychophysiology, 52, 63–76.

41.

Siegle

G. J.

Steinhauer

S. R.

Thase

M. E.

Stenger

V. A.

Carter

C. S.

(2002). Can’t shake that feeling: Event-related fMRI assessment of sustained amygdala activity in response to emotional information in depressed individuals. Biological Psychiatry, 51, 693–707. doi:S0006322302013148[pii]

42.

Steiner

G. Z.

Barry

R. J.

(2011). Pupillary responses and event-related potentials as indices of the orienting reflex. Psychophysiology, 48, 1648–1655. doi:10.1111/j.1469-8986.2011.01271.x

43.

Tucker

D. M.

Luu

Frishkoff

Quiring

Poulsen

(2003). Frontolimbic response to negative feedback in clinical depression. Journal of Abnormal Psychology, 112, 667–678.

44.

Wisco

B. E.

Nolen-Hoeksema

(2010). Interpretation bias and depressive symptoms: The role of self-relevance. Behaviour Research and Therapy, 48, 1113–1122. doi:10.1016/j.brat.2010.08.004

45.

Zimmerman

Sheeran

Young

(2004). The Diagnostic Inventory for Depression: A self-report scale to diagnose DSM-IV major depressive disorder. Journal of Clinical Psychology, 60, 87–110. doi:10.1002/jclp.10207