Better cognitive efficiency is associated with increased experimental anxiety

Abstract

There is increased interest in the development of cognitive training targeting working memory (WM) to alleviate anxiety symptoms, but the effectiveness of such an approach is unclear. Improved understanding of the effect of cognitive training on anxiety may facilitate the development of more effective cognitive training treatment for anxiety disorders. This study uses an experimental approach to examine the interplay of WM and anxiety following WM training. Previous studies show that increased demand on WM reduces concurrent anxiety evoked by threat of shock (induced anxiety). However, improving WM pharmacologically or via exercise prevents this anxiolytic effect. Conceivably, improving WM frees up cognitive resources to process threat information, thereby increasing anxiety. The present study tested the hypothesis that practicing a high load WM (i.e., increased demand) task would improve WM, and thus, free cognitive resources to process threat of shock, resulting in more anxiety (i.e., greater startle) during a subsequent WM task. Participants were randomly assigned to two training groups. The active-training group (N = 20) was trained on a 1- (low load) & 3-back (high load) WM task, whereas the control-training group (N = 20) performed a 0-back WM task. The experimental phase, similar in both groups, consisted of a 1- & 3-back WM task performed during both threat of shock and safety. As predicted, active training improved WM accuracy and increased anxiety during the experimental 3-back WM task. Therefore, improving WM efficiency can increase anxiety, possibly by freeing WM resources to process threat information.

1 |. INTRODUCTION

Anxiety disorders are the most common mental health disorders (Kessler et al., 2009), and represent a substantial psychological, social, and economic burden (Beddington et al., 2008). In search of novel treatments for these conditions, investigators have focused on underlying cognitive and attentional vulnerabilities (Hallion & Ruscio, 2011; Koster, Hoorelbeke, Onraedt, Owens, & Derakshan, 2017; Leone de Voogd, Wiers, Zwitser, & Salemink, 2016; Sari, Koster, Pourtois, & Derakshan, 2016). Impaired working memory (WM) represents one such vulnerability (Eysenck, Derakshan, Santos, & Calvo, 2007; Moran, 2016; Otto et al., 2016). WM refers to the process of actively maintaining a limited amount of information and protecting this information from interference in the service of cognitive tasks (Baddeley, 2003). It has long been suggested that anxiety interacts with WM, such that anxiety impairs WM, whereas impaired WM increases anxiety by allowing the threat to reach awareness (Moran, 2016; Otto et al., 2016; Stout, Shackman, & Larson, 2013). The latter has led to the suggestion that improving WM could alleviate anxiety (Hotton, Derakshan, & Fox, 2018; Koster et al., 2017; Sari et al., 2016). However, studies of WM training on anxiety have yielded mixed results, with studies showing either reduced anxiety after training or no effect of training (Koster et al., 2017; Leone de Voogd et al., 2016; Sari et al., 2016). Understanding how WM affects anxiety is the first step toward developing improved cognitive treatments. The present study focuses on the effect of WM training on subsequent experimental anxiety induced by the threat of shock during performance of an n-back task.

Most cognitive theories that explain the interaction between emotion and cognition (e.g., Attention Control Theory) assume competition for resources between anxiety-related processes and task-related processes (Eysenck et al., 2007; Moran, 2016). According to this view, anxiety interferes with cognitive performance by diverting limited cognitive resources away from a task and toward threatening stimuli (Eysenck et al., 2007). The application of these theories has generally focused on the effect of anxiety on cognitive performance, whereas the ability to downregulate anxiety by engaging in a cognitive task has been relatively unexplored (Vytal, Cornwell, Arkin, & Grillon, 2012). This latter question is crucial if the aim of WM training is to alleviate anxiety.

Several studies have investigated the interplay between WM and experimental anxiety. These studies used threat of shock to induce anxiety, which is measured with the acoustic startle reflex as a physiological index of anxious arousal (i.e., anxiety-potentiated startle) (Dvorak-Bertsch, Curtin, Rubinstein, & Newman, 2007; Patel et al., 2017; Shackman et al., 2006; Vytal et al., 2012; Vytal, Cornwell, Arkin, Letkiewicz, & Grillon, 2013). Two key findings emerged from these studies. First, there was overall an opposite effect of induced anxiety on accuracy in low and high load WM tasks, with induced anxiety impairing accuracy on low load (less WM demand) tasks but improving performance on high load (greater WM demand) tasks (Patel et al., 2017; Vytal et al., 2012, 2013). Second, and more relevant to the current study, anxiety decreased as WM load increased (Dvorak-Bertsch et al., 2007; Patel et al., 2017; Vytal et al., 2012, 2013). These results suggest that when cognitive tasks do not fully engage cognitive resources, the available cognitive resources can be devoted to process threat information, resulting in increased anxiety (Giambra, 1995). Anxiety is then elevated and interferes with performance. Conversely, more difficult tasks monopolize cognitive resources and limit the processing of threat information. As a result, anxiety is reduced and interference with performance is minimized.

If engaging WM has an anxiolytic effect on induced anxiety, consistent with the rationale for WM training treatment to relieve anxiety, improving WM should enhance this anxiolytic effect. We recently tested this possibility in two studies that attempted to improve cognition either pharmacologically or behaviorally. The first study used methylphenidate (Ernst, Lago, Davis, & Grillon, 2016), a well-established cognitive enhancer that increases extracellular dopamine concentration (Volkow et al., 2001). The second study used exercise (Lago et al., 2019), which has also been shown to improve cognition (Crush & Loprinzi, 2017; Hillman, Snook, & Jerome, 2003; Loprinzi & Kane, 2015). Results were contrary to the hypothesis. Although we expected methylphenidate and exercise to further decrease anxiety-potentiated startle during high load WM tasks, we found the opposite effect. These two treatments increased anxiety-potentiated startle. The most straightforward interpretation of these results is that the active treatments improved the efficiency of WM. Consequently, the high load WM task could be performed with fewer cognitive resources, leaving resources available for threat processing, thereby increasing anxiety. According to this view, other manipulations that improve WM should also increase anxiety during task performance.

One such manipulation is WM training, which, by promoting automatization and reducing the need for attentional control, diminishes cognitive resource demand (Mason et al., 2007; Schneider & Shiffrin, 1977). Consequently, the present study tests the hypothesis that following WM training, anxiety induced by threat of shock would be higher during a WM task. Specifically, we predicted that following 1- & 3-back WM training compared to 0-back training, the 1- & 3-back training group would show improved WM performance and, based on our previous findings (Ernst et al., 2016; Lago et al., 2019), would fail to show decreased anxiety-potentiated startle in the 3-back compared to the 1-back WM load.

2 |. METHOD

2.1 |. Participants

Healthy adults were recruited from the Washington, DC metropolitan area via advertisements and flyers and were randomly assigned either to the active-training group or to the control-training group. Prior data from our lab (Ernst et al., 2016; Lago et al., 2019) indicated an expected moderate effect size for startle potentiation (f = .25). A priori calculations indicated that testing 36 participants at an alpha of .05 would provide power of .95 to detect a significant 2 (group) × Load (2) × Shock (2) interaction. Assuming a 25% attrition rate, we enrolled 45 participants into the study. Of these 45, four were excluded due to: one (from the active-training group) for showing an overall performance less than three or more SD from the mean of the entire sample, and four (two from each group) because of no or small startle responses (see Section 2.6). The final sample included 20 participants (eight male) per group. Participants’ demographic information is shown in Table 1.

TABLE 1.

Demographic information, mean (SEM) values

Practice groups	Age (year)	Trait anxietya	WASIb
Control (N = 20)	27.3 (6.5)ns	30.8 (1.3)ns	113.9 (2.3)ns
Active (N = 20)	26.7 (8.5)	29.7 (1.4)	119.0 (2.6)

Note: ^ns for nonsignificant group difference.

^aSpielberger trait form of the state-trait anxiety inventory.

^bWechsler Abbreviated Scale of Intelligence.

All participants were free from the following exclusion criteria: (a) current or past diagnosis of Axis I psychiatric disorder as assessed by SCID-I/NP (First, Williams, Karg, & Spitzer, 2015), (b) positive toxicology screen or psycho-active medications, (c) medical condition conflicting with safety or design of the study, (d) pregnancy (female). All participants gave written informed consent approved by the NIMH Combined Neuroscience Institutional Review Board and were compensated for their participation.

2.2 |. Experimental design

The study used a between-subject design and consisted of (a) n-back training, (b) placement of electrodes, (c) a startle habituation procedure to reduce initial startle reactivity, (d) a shock workup procedure, (e) the experimental task. During n-back training, participants in the active-training group performed the 1-/3-back WM task (see below) and participants in the control-training group performed a 0-back (i.e., non-WM), task. The experimental task consisted of the 1-/3-back WM task performed during alternative periods of safety and threat of shock.

2.3 |. Stimuli and physiology

Stimuli were presented using the Presentation software package (Neurobehavioral Systems, Berkeley, CA) via a standard 19-in. LCD monitor, and responses were made on a standard mechanical keyboard.

2.3.1 |. Startle

The eyeblink startle reflex was recorded with electromyo-graphic (EMG) electrodes placed under the left eye. It was elicited by 40-ms duration 103-dB bursts of white noise (i.e., startle probes) (near-instantaneous rise/fall times) delivered over headphones. EMG data were digitized (1,000 Hz), filtered (30–500 Hz), rectified, and smoothed using a 20-ms sliding window.

2.3.2 |. Shock

Anxiety was induced with threat of shock (Vytal et al., 2012, 2013). The shock was delivered to the wrist using a 100-ms, 200-Hz train of stimulation delivered by a constant current stimulator (DS7A; Digitimer).

2.3.3 |. Startle habituation and shock workup

After n-back training, participants received nine presentations of the startle noise to habituate the startle reflex. This was followed by a standard shock workup procedure to determine individual shock intensity. The shock was set at a level that was uncomfortable and unpleasant, but not painful as we have done in the past (Vytal et al., 2012, 2013).

2.4 |. N-back task

2.4.1 |. Active-training

Participants were asked to remember a continuous series of letters presented sequentially (in purple circles) on a computer screen, and to indicate by pressing a key on the keyboard whether the letter currently displayed was the same (“s”) or different (“d”) compared to a letter presented one position (1-back) or three positions (3-back) back in the sequence. The task was explained and visually demonstrated to the participants for about 1 to 2 min. Each training session included two successive runs lasting approximately 7 min each. For the active-training group, each run consisted of four blocks of 1-back and four blocks of 3-back presented in random order. Each block consisted of 18 letter trials and began with an instruction screen (e.g., “1-back” or “3-back”). All letters were presented for 500 ms with an interstimulus interval of 2,000 ms.

2.4.2 |. Control-training

The same stimuli were presented to the control-training group, but participants were asked to press the left arrow key every time a letter was presented (in purple circles).

2.4.3 |. Experiment

The 1-/3-back task was explained to the control-training participants, who were shown the task for about 1 to 2 min. All participants were informed that the 1-back/3-back task would be performed during threat and safe periods, as indicated by the colored circle surrounding each letter (blue for safe and orange for threat). They were told that they could receive unpredictable shock during the threat periods but not during the safe periods, and that their performance would not impact the delivery of the shock. They were also told that occasionally they would receive a brief burst of noise (i.e., startle probe). Shocks and startle probes were presented pseudo-randomly (1,000–1,700 ms after the letters to minimize potential prepulse inhibition) and had no systematic relationship with letter presentation. The n-back task was organized in two experimental runs (separated by 3 min) with eight blocks of 18 trials each. Within each run, four safe and four threat blocks alternated. Each block featured a single n-back level (1-back or 3-back), 0–2 shocks and 2–3 startle probes. In total, each WM load was tested four times in safe and threat conditions and participants received 6 shocks and 40 startle probes (10 startle probes per condition (safe/1-back, safe/3-back, threat/1-back, threat/3-back)).

2.5 |. Anxiety ratings, cognitive ability, and workload

During screening, participants’ trait anxiety was assessed with the trait subscale of the Spielberger State-Trait Anxiety Inventory (Spielberger, 1983). Estimate of cognitive functioning were obtained with the Wechsler Abbreviated Scale of Intelligence (WASI) (Wechsler, 2011).

After each experimental run (see below), participants rated their subjective anxiety for each load (1-back, 3-back) and each shock condition (safe, threat) on a paper-based analog scale ranging from 1 (not at all) to 10 (extremely) (e.g., How anxious were you during 1-back under threat).

The subjective level of mental fatigue and workload were measured using the National Aeronautics and Space Administration-Task Load Index (NASA-TLX) (Har & Staveland, 1988). The NASA-TLX was used to assess potential differences in mental fatigue and workload in the two groups. The NASA-TLX is divided into six components with specific subscales as follows: mental demand (How mentally demanding was the task?), physical demand (How physically demanding was the task?), temporal demand (How hurried or rushed was the pace of the task?), performance (How successful were you in accomplishing what you were asked to do?), effort (How hard did you have to work to accomplish your level of performance?), and frustration (How insecure, discouraged, irritated, stressed, and annoyed were you?), which participants rated on analog scales ranging from 1 (low) to 10 (high).

2.6 |. Data analysis

The magnitude of the startle response was calculated as the difference between the peak EMG activities during the 20–120 ms window following the startle probe minus the mean baseline EMG activity −50 to 0 ms immediately prior to the startle probe. Subjects with three or more no blink responses during the startle habituation procedure were excluded. Startle responses (including 0 responses) were averaged within each load and shock condition. To control for interindividual variability in startle reactivity, startle magnitudes during the experimental task were standardized using within-subject T scores ([z scores × 10] + 50) (Patel et al., 2017; Vytal et al., 2012, 2013).

Subjective reports, accuracy, and reaction time of accurate responses (RT) were averaged within each load and shock condition.

These measures were analyzed with a three-way analysis of variance (ANOVAs) with group (control-training, active-training) as a between-subject factor and WM load (1-back, 3-back) and shock condition (safe, threat) as repeated factors. Based on our a priori hypothesis, we expected significant group × WM load × shock condition interactions.

Each individual item of the NASA-TLX was analyzed with a two-way ANOVA with group (control-training, active-training) as a between-subject factor and time (post-training, post-run 1, post-run 2) as a within-subject factor. To avoid sphericity with the 3 times, we used the Huynh-Feldt term of the repeated-measures ANOVA.

We report both partial eta-squared (${\eta }_p^2$) and Cohen’s d to document effect sizes.

3 |. RESULTS

3.1 |. Demographic information

Demographic data are presented in Table 1. The two groups did not differ in age, trait anxiety, or WASI scores (all p > .5).

3.2 |. Mental workload

There was no group main effect for any of the NASA-TLX items. There were group × time interactions for mental load (F(2,76) = 43.1, p < .00009, ${\eta }_p^2=.49$), temporal demand (F(2,76) = 18.7, p < .0009, ${\eta }_p^2=.24$), effort F(2,76) = 9.1, p < .001, ${\eta }_p^2=.23$), and frustration (F(2,76) = 14.7, p < .00009, ${\eta }_p^2=.29$). All of these interactions were due to higher mental load (7.3 versus 2.9, t(38) = 6.6, p < .0009, d = 2.1), temporal demand (5.6 versus 4.0, t(38) = 2.3, p = .03, d = .73), effort (6.7 versus 3.5, t(38) = 4.6, p < .0009, d = 1.15), and frustration (3.9 versus 2.9, t(38) = 2.8, p = .004, d = .99) after training in the active-training group compared to the control-training group, with no group difference postexperimental runs 1 and 2. These results suggest no mental workload difference between the two groups during the experimental phase, in contrast to during the training.

3.3 |. Performance

3.3.1 |. Accuracy

The accuracy scores are shown in Table 2. There was a group main effect (F(1,38) = 5.3, p = .027, ${\eta }_p^2=.17$), due to higher accuracy in the active-training group compared to the control-training group. There was also a WM Load × Shock Condition interaction (F(1,38) = 5.8, p = .020, ${\eta }_p^2=.14$), due to overall higher accuracy in the threat compared to the safe condition during 3-back (+ 3.4%, SEM = 1.1%: F(1,39) = 9.1, p = .004), but no significant effect of threat during 1-back (−.8%, SEM = 1.0%: F(1,38) = .57, ns).

TABLE 2.

Mean (SEM) startle and performance scores

	Control training				Active training
	Safe		Threat		Safe		Threat
	1-back	3-back	1-back	3-back	1-back	3-back	1-back	3-back
Anxiety rating	2.7 (.3)	3.8 (4)	4.2 (.5)	5.6 (.5)	2.2 (.2)	3.1 (.3)	4.0 (.3)	5.8 (.4)
Accuracy (%)	86.2 (3.8)	63.9 (4.7)	85.3 (3.8)	67.3 (4.8)	93.8 (1.7)	74.8 (2.2)	93.2 (1.5)	78.3 (2.6)
Reaction time (ms)	649.7 (39.5)	707.1 (46.3)	636.1 (36.4)	765.3 (50.5)	713.4 (49.0)	825.9 (62.7)	677.8 (42.4)	839.4 (53.1)

3.3.2 |. Reaction time

The RT scores are shown in Table 2. There was a WM Load main effect (F(1,38) = 40.7, p < .0009, ${\eta }_p^2=.51$), due to slower RT in 3-back compared to 1-back. There was also a WM Load × Shock Condition interaction (F(1,38) = 8.5, p = .006, ${\eta }_p^2=.18$) that reflected faster RT in the threat compared to the safe condition during 1-back (−24.6 ms, SEM = 11.4 ms: F(1,39) = 4.6, p = .04), but a trend for slower RT in the threat compared to the safe condition during 3-back (+ 35.8 ms, SEM = 13.6 ms: F(1,39) = 3.9, p = .053). There was no group main or interaction effect.

3.4 |. Startle, T scores

The startle scores are shown in Figure 1. As expected, there was a Shock Condition main effect (F(1,38) = 121.1, p < .0009), due to larger startle magnitude in the threat compared to the safe condition (i.e., anxiety-potentiated startle). However, this effect was qualified by a three-way group × WM Load × Shock Condition interaction (F(1,38) = 4.8, p = .035, ${\eta }_p^2=.11$). This interaction was due to the fact that in the control-training group, anxiety-potentiated startle decreased significantly as WM load increased (F(1,38) = 6.1, p < .02, ${\eta }_p^2=.13$), whereas in the active-training group anxiety-potentiated startle (startle magnitude in the threat condition minus safe condition) was not affected by WM load (F(1,38) = .4, ns, ${\eta }_p^2=.01$). As a result, the magnitude of anxiety-potentiated startle did not differ between the two groups during 1-back (t(38) = .3, ns, d = .08), but it was significantly larger in the active-training group compared to the control-training group during 3-back (t(38) = 2.1, p = .04, d = .65).

FIGURE 1.

Startle (eyeblink) magnitude in the control-training and active-training groups in the safe and threat conditions of the 1-back and 3-back WM loads. * indicates a significant increased (p < .01) from the safe to the threat condition (i.e., fear-potentiated startle). ^ indicates a significant decreased (p = .02) in fear-potentiated startle magnitude from the 1-back condition to the 3-back condition in the control-training group. ns indicates no significant difference in fear-potentiated startle magnitude between the 1-back and 3-back condition in the active-training group. Error bars are SEM

Note that although high load tasks are usually more stressful than low load tasks (see anxiety rating below), the WM load main effect was not significant (F(1,39) = 1.0, ns).

3.5 |. Subjective anxiety

The subjective anxiety scores are shown in Table 2. There were main effects of Shock Condition (F(1,38) = 85.8, p < .0009, ${\eta }_p^2=.69$) and WM load (F(1,38) = 30.7, p < .0009, ${\eta }_p^2=.44$). Like startle, subjective anxiety was increased in the threat compared to the safe condition, but, unlike startle, it was increased in the 3-back compared to the 1-back WM load. There was also a WM Load × Shock Condition interaction (F(1,38) = 13.2, p = .001, ${\eta }_p^2=.26$), which was due to a greater increase in anxiety ratings from the safe to the threat condition in 3-back (+ 2.3, SEM = .24) compared to 1-back (+ 1.6, SEM = .22).

There was also a trend for a group × WM load × Shock Condition interaction (F(1,38) = 2.8, p = .1, ${\eta }_p^2=.07$). We followed up this interaction with additional analyses because of our a priori hypothesis (which called for such a three-way interaction) and because anxiety-potentiated startle showed increased anxiety in the threat/3-back condition in the active-training group compared to the control-training group. To analyze the rating data like the startle data, we quantified subjective anxiety-potentiated ratings as the difference in ratings from safe to threat. Anxiety-potentiated ratings did not differ between the two groups during 1-back (t(38) = .5, ns, d = .16) but tended to be greater in the active-training group compared to the control-training group during 3-back (t(38) = 1.8, p = .08, d = .56).

Because, unlike startle, anxiety ratings were affected by WM load, we attempted to better understand how anxiety ratings related to workload by correlating the anxiety ratings in each condition with the items of the NASA-TLX questionnaire. Results (Table 3) confirm that ratings were affected by the workload of the WM task, especially during 3-back.

TABLE 3.

Correlation (level of significance) between anxiety ratings and NASA-TLX items in the entire sample

	Mental demand	Physical demand	Temporal demand	Performance	Effort	Frustration
1-back/safe	.08 (ns)	.27 (ns)	.02 (ns)	.01 (ns)	−.02 (ns)	.15 (ns)
3-back/safe	.47 (.002)	.38 (.03)	.41 (.008)	.14 (ns)	.34 (.04)	.39 (.01)
1-back/threat	.02 (ns)	.11 (ns)	−.02 (ns)	−.25 (ns)	0 (ns)	.35 (.02)
3-back/threat	.22 (ns)	.37 (ns)	.17 (ns)	−.11 (ns)	.25 (ns)	.41 (.008)

4 |. DISCUSSION

Previous work shows that increasing demand on WM reduces anxiety evoked by threat of shock (Patel et al., 2017; Vytal et al., 2012, 2013) and that improving WM reduces this anxiolytic effect (Ernst et al., 2016; Lago et al., 2019). The latter result was interpreted in terms of cognitive resources, such that improving WM frees cognitive resources to process threat. The aim of the present study was to test the hypothesis that a period of WM training would similarly reduce the anxiolytic effect of increasing demand on WM. To this aim, the training manipulation was designed to improve WM efficiency (Antrobus, 1968; Giambra, 1995; Mason et al., 2007) in the active-training group, but not the control-training group. Subsequent to training, we manipulated orthogonally both WM load (1-, 3-back) and anxiety state (safe, threat). Results were consistent with our hypothesis. In the control-training group, we replicated the previous effect of WM load on anxiety. Specifically, the magnitude of anxiety-potentiated startle was reduced in the 3-back compared to the 1-back WM load. However, this downregulation of anxiety-potentiated startle in the 3-back was not replicated in the active-training group. As a result, anxiety-potentiated startle in the 3-back was higher in the active-training group compared to the control-training group.

The experimental manipulation was successful. The threat condition evoked anxiety: startle was potentiated, and the anxiety ratings were higher in the threat compared to the safe condition. Although startle potentiation was relatively large (about eight T scores in the 1-back condition), the subjective ratings were moderate, which suggests that the effect of practice are evident even with moderate threat. In addition, as documented by the performance data, the 3-back WM task was more difficult than the 1-back. Finally, WM training was successful: performance accuracy was better in the active-training group compared to the control-training group. The mechanism by which WM training increased anxiety is not clear, but it likely involves enhanced availability of processing resources (Antrobus, 1968; Giambra, 1995; Mason et al., 2007). Indeed, although attention to threat information can be fast and unintentional, it relies on WM (Van Dillen & Koole, 2009). The automatic shift of attention toward task-irrelevant processing when resources become available has been shown in studies of mind wandering, and might be a factor at play in this study. Mind wandering is a shift of attention away from ongoing goals toward internally generated thoughts or concerns (McVay & Kane, 2010; Smallwood & Schooler, 2006). Typically, in studies that probe mind wandering during cognitive tasks, individuals report high levels of mind wondering during easy WM tasks, but little mind wandering during difficult WM tasks (Antrobus, 1968; Giambra, 1995; Mason et al., 2007). However, after WM training the level of mind wandering increases. This is reminiscent of the present results where anxiety-potentiated startle was reduced by WM training (Antrobus, 1968; Giambra, 1995; Mason et al., 2007). Taken together, these results suggest that WM training may have promoted mind wandering (i.e., worrying) about the threat of shock, thereby increasing anxiety.

The finding that WM training increases anxiety challenges the proposal that improved WM can alleviate anxiety (Hotton et al., 2018; Koster et al., 2017; Sari et al., 2016). The rationale for treatments seeking to improve WM comes from observations of reduced cognitive control over information in WM in anxious individuals (Kertz, Belden, Tillman, & Luby, 2016; Moran, 2016), which may act as a vulnerability factor (Eysenck et al., 2007; Millan et al., 2012; Otto et al., 2016). Consequently, WM training should improve WM and cognitive control, including the ability to focus on task goals and prevent interference from irrelevant emotional information, and thus, improve emotion regulation (Beloe & Derakshan, 2019; Leone de Voogd et al., 2016). So far, however, the impact of cognitive or WM training is not clear. First, there is no strong evidence to suggest that the effect of cognitive training in one domain (e.g., visual WM) can transfer to another domain (e.g., verbal WM), or that any effect can last over time (Au et al., 2015; Melby-Lervåg, Redick, & Hulme, 2016). If WM training does not generalize across cognitive tasks, then, any effect on negative information processing may be restricted to a subset of cognitive tasks. Second, among the dearth of studies addressing the effect of cognitive training on anxiety or emotional processing, results are inconclusive. Some studies show reduced self-reported anxiety after several days of cognitive training (Beloe & Derakshan, 2019; Hadwin & Richards, 2016), whereas others show mixed (Sari et al., 2016) or no significant (Larsen et al., 2019; Leone de Voogd et al., 2016) effects.

If the function of WM is not only to maintain task goals in mind, but also to protect these goals from external and internal interference, why was the WM active-training group not more protected from threat processing? Some theories of WM emphasize the maintenance of information in the face of simultaneous processing, whereas others focus on control of interfering information (Baddeley, 1992; Engle & Kane, 2003). It is possible that n-back training improved the former but not the latter process. A related possibility is that WM training was conducted without emotional distractors or emotional contexts. Cognitive training with only neutral stimuli may be unlikely to transfer to tasks that contain emotional distractors. For such a transfer to occur, cognitive training requires the presence of emotional information (Iacoviello & Charney, 2015). Indeed, recent studies show that WM training with emotional distractors compared to similar training with no such distractors improves cognitive control over affective information, emotion regulation, and augments the efficiency of the fronto-parietal cognitive control network (Schweizer, Grahn, Hampshire, Mobbs, & Dalgleish, 2013). These results raise the possibility that WM training with emotional stimuli would increase, rather than decrease, the anxiolytic effect of WM load during threat of shock.

The increased anxiety-potentiated startle in the 3-back WM load in the active-training group was not replicated with the subjective anxiety ratings. However, we do not believe that this lack of consistency between these two measures weakens the conclusion that training increased anxiety during the 3-back WM load during threat. It is possible that some of these differences between startle and anxiety ratings were artifactual; the startle probe provided an online measure of anxiety, whereas the anxiety ratings were retrospective and may have been influenced by a host of effects such as task demand, the passage of time and the complexity of the design. However, there was an important difference between the startle reflex and anxiety ratings. Startle magnitude was not affected by the difficulty of the task (i.e., increased in WM load), whereas anxiety ratings were higher in the 3-back compared to the 1-back task (see data in the safe condition). These results indicate that startle was solely sensitive to the threat of shock, whereas the anxiety ratings were sensitive to both the threat of shock and cognitive load. In fact, the correlation analysis between the anxiety ratings and the NASA-TLX items (Table 3) confirm that the anxiety ratings were sensitive to the difficulty of the WM task. These observations are consistent with the view that startle potentiation reflects the activation of defense survival mechanisms to highly arousing stimuli (Davis, Falls, Campeau, & Kim, 1993; Lang, 1995; LeDoux, 2014; LeDoux & Brown, 2017), whereas anxiety ratings may reflect a more general subjective anxiety state evoked by different types of threats. Despite these differences, there was some evidence that like anxiety-potentiated startle, subjective anxiety was higher in the 3-back WM task/threat in the active-training group compared to the control-training group. However, this effect did not quite reach significance (2-tails). Taken together, these results indicate that increasing demands on WM reduces the activation of defensive survival mechanisms activated by threat of shock and that WM training diminishes or prevents this reduction. Whether WM load and WM training have similar effects on subjective anxiety cannot be clearly determined, because the anxiety ratings reflected subjective feelings evoked by the threat of shock and by the WM task. It will be important for future cognition/anxiety interaction studies to attempt to dissociate feelings related to threat versus cognitive stress. Neuroimaging techniques could contribute to such an understanding.

This study has several strengths. A significant strength is the use of a well-validated physiological measure of anxiety, the startle reflex, and a well-established method of evoking anxiety, the threat of shock. Relative to self-report, objective measures are less likely to be influenced by subjective bias and demand characteristics (Levenson, 2007), especially when experimental manipulations cannot be blinded. An additional strength is that our a-priori hypotheses are based on studies that used a very similar paradigm to study the interaction of cognition and anxiety (Ernst et al., 2016; Lago et al., 2019; Patel et al., 2017; Vytal et al., 2013). A limitation of the study is that it focused on n-back task to measure WM. The construct validity of the n-back task has been questioned (Kane, Conway, Miura, & Colflesh, 2007). In addition, WM covers an umbrella of functions (e.g., inhibition, updating, shifting). It is possible that a different type of cognitive or WM training may have led to a different outcome (Miyake et al., 2000). Future studies should seek to identify the fundamental ingredients that make training more efficacious. However, the present study suggests that improving cognitive efficiency on any task may free processing resources, which then may become available for task-irrelevant processing, including threat.

We believe the present results are relevant to the discussion on the reconceptualization of aversive states in humans (Fanselow & Pennington, 2018; LeDoux, 2014; LeDoux & Brown, 2017). According to the “two-system” framework, two classes of loosely related responses are elicited by threats: (a) autonomic and behavioral responses reflecting the activation of a defense survival mechanisms and (b) conscious feeling states relying on higher order cognitive circuits, which are reflected in self-reports (i.e., anxiety ratings) (LeDoux, 2014; LeDoux & Brown, 2017). Relevant to the current study is the assumption that WM, as a higher cognitive process, plays a preponderant role in the conscious experience of fear and anxiety (i.e., subjective feeling) but not in the defense survival system. This assumption is not supported by current findings, or past studies exploring the interplay between anxiety-potentiated startle and WM load (Dvorak-Bertsch et al., 2007; Patel et al., 2017; Shackman et al., 2006; Vytal et al., 2012, 2013). These studies clearly show that anxiety-potentiated startle, an index of the defense survival system, is robustly affected by manipulations that affect WM. Although it is likely that the subjective feelings of fear or anxiety require WM (LeDoux & Brown, 2017), it is also clear from these studies that WM impacts defense survival responses. It is possible that physiological responses and subjective feelings are more closely related (Fanselow & Pennington, 2018) than assumed by the two-system framework and that subjective feelings are partly driven by physiological responses. It is also possible that subjective feelings and survival mechanisms are influenced by different neural circuits, including subcortical circuits (Bolkan et al., 2017; Guo et al., 2017; Peräkylä et al., 2017), which give rise to different aspects of WM. In order to better understand the influence of cognitive control in anxiety, it will be important for future studies to clarify how WM relates to the subjective experience of anxiety and defense survival mechanisms.

In conclusion, the hypothesis that WM training would reduce the anxiolytic effect of higher WM load on anxiety induced by threat of shock was confirmed. We argue that WM training improved cognitive efficiency and allowed simultaneous processing of both the 3-back WM load and threat. We suggest that WM training in an emotional context or with emotional distractors could improve control over task-irrelevant affective information, including information regarding threat. Threat of shock offers a powerful experimental model of anxiety to test this possibility and to generate insights into the key cognitive processes that need to be targeted by cognitive training to boost control over emotional stimuli (Grillon, Robinson, Cornwell, & Ernst, 2019).