Combining Unguided Web-Based Attentional Bias Modification and Affective Working Memory Training to Decrease Anxiety: A Randomized Controlled Trial

Anxiety disorders are the most common group of mental disorders with a 12-month prevalence of approximately 14% in Europe (Wittchen et al., 2011). The course of anxiety disorders is quite persistent (Bruce et al., 2005) and many individuals still experience symptoms after cognitive-behavioral therapy (Springer et al., 2018). Despite the burden of anxiety, many individuals fail to seek treatment (Olfson et al., 2000). Web-based interventions have the potential to break down barriers as they are easy to access, suitable for wide-scale implementation, and have low financial costs. Interventions that can be easily transferred to a web-based format are computerized cognitive trainings (Gober et al., 2021). The goals of this study were twofold: (1) to investigate whether web-based computerized attentional bias training and affective working memory training would be effective in reducing trait anxiety and (2) to investigate moderators of training effectiveness.

Cognitive models of anxiety propose several individual difference factors that jointly contribute to a cognitive vulnerability to anxiety, including strong threat-related associations and low levels of cognitive control or working memory capacity (Mathews & Macleod, 2005; Mogg & Bradley, 2016; Ouimet et al., 2009). It has been put forward that anxiety disorders are likely to arise when the strength of emotional activation exceeds the capacity for control (Mathews & Macleod, 2005). When threat-related and non-threat-related stimuli are present in the environment, there is enhanced orientation towards the threatening stimulus, increased attentional engagement and impaired attentional disengagement with the threatening stimulus in anxious individuals (Ouimet et al., 2009). This selective attention to threat could be overcome by high levels of cognitive control, such as working memory capacity. Thus, individual differences in two kinds of cognitive processes (attentional information processing biases and control processes) and their interplay are hypothesized to play a role in the development and maintenance of anxiety disorders (see also Mogg & Bradley, 2016). This may suggest that in order to reduce anxiety, interventions should simultaneously reduce selective attention to threat as well as increase the strength of control processes.

With respect to attentional processes, studies have demonstrated that individuals with elevated levels of anxiety are prone to selectively attend to threatening stimuli (Bar-Haim et al., 2007; Cisler et al., 2010) and theoretical models argue that this attentional bias (AB) for threat plays a role in the onset and maintenance of anxiety (Mathews & Macleod, 2005). Subsequent experimental studies manipulated AB and demonstrated effects on anxiety (for reviews see Macleod & Clarke, 2015; Van Bockstaele et al., 2014); providing evidence consistent with a causal role of AB in anxiety. Clinically interesting is the possibility to reduce ABes to threat with the aim of reducing anxiety. Attention Bias Modification (ABM) training is a computerized training in which participants practice allocating attention to neutral or positive stimuli over negative stimuli. While some meta-analyses reported non-significant effects from ABM on anxiety symptom reduction compared to control conditions (Cristea et al., 2015), others reported more promising, significant effects on anxiety (Fodor et al., 2020, when post-traumatic stress trials were excluded; Linetzky et al., 2015; Price et al., 2016 depending on specific conditions). Generally, there is heterogeneity in ABM’s effect on anxiety (see also Vrijsen et al., 2024) and there are large prediction intervals in relation to ABM’s effect on anxiety when compared with waitlist or sham training (Fodor et al., 2020), and it has been suggested that future studies should include larger sample sizes and examine relevant moderators.

One potential moderator might be context of training as meta-analyses suggested that lab-based ABM studies yielded larger effect sizes than web-based or home-based studies (Cristea et al., 2015; Fodor et al., 2020; Martinelli et al., 2022). A possible explanation for this difference is that participants likely feel more anxious and aroused in the lab compared to their own home. Given that moderate levels of arousal have been associated with better learning performances (Yerkes & Dodson, 1908), increased anxiety and arousal might enhance training effects (Nuijs et al., 2020). In line with this idea, Kuckertz et al. (2014) provided preliminary evidence that ABM combined with anxiety induction was more effective in reducing AB and social anxiety symptoms than ABM without a mood induction. Therefore, the aim of the current study was to examine the effectiveness of ABM training in changing anxiety and innovatively test whether naturally occurring increased levels of state anxiety prior to training sessions enhance ABM’s effects.

Whereas many studies investigated whether training AB can impact anxiety symptoms, less studies investigated whether improving control processes can impact mood and anxiety symptoms (Iacoviello & Charney, 2015). Working Memory (WM) functioning seems a promising target for anxiety-related interventions as in general, cognitive control deficits (Paulus, 2015) and specifically WM functioning impairments (Moran, 2016) have been observed in individuals with elevated levels of anxiety. As such, WM training might be a valuable intervention for anxiety by improving control processes and thereby possibly decreasing anxiety symptoms. However, initial findings from meta-analyses regarding WM training concluded that there was no convincing evidence for far transfer effects (thus beyond the trained process) when compared with an active control condition (Melby-Lervåg et al., 2016). More recently, the crucial distinction between WM in neutral versus affective contexts has been highlighted (Schweizer et al., 2019). That is, WM performance of individuals with mental health problems was specifically impaired by negative, affective material compared to neutral material; suggesting that specifically affective WM capacity plays a role in poor mental health. Relatedly, when directly comparing the effectiveness of affective versus neutral WM training (Minihan et al., 2021), findings suggested that affective WM training resulted in greater reductions in test anxiety than neutral WM training. In sum, the evidence regarding the effectiveness of neutral WM training is mixed, while findings with respect to affective WM training seems more promising in the context of anxiety.

The current study investigated the effectiveness of combining web-based ABM and affective WM training in decreasing trait anxiety and the potential role of state anxiety in ABM training effects. Adults with heightened trait anxiety were randomized to ten training sessions in which they received ABM (either active or control) and affective WM training (either active or control) in a 2 × 2 design. The active version of ABM consisted of a visual search paradigm where participants had to repeatedly find one positive emotional (i.e., happy) face in a 4 × 4 grid with 15 negative emotional faces. In the control version of ABM, participants had to find the only 5-petaled flower in a grid with 15 7-petaled flowers (de Voogd et al., 2014). Visual search training may offer a promising alternative to other ABM paradigms as the training instructions are more explicit concerning the valence of the stimuli (i.e., ‘to search for the positive face’) and hence may facilitate goal-directed processing (Mogg & Bradley, 2018). The active version of affective WM training consisted of an adaptive WM Training with Emotional Distractors (WMT-ED; De Voogd et al., 2016a) and the control version of a non-adaptive version of the same training. For unguided, web-based training, it is important to keep participants engaged with the training. Therefore, before the start of training, participants were asked to formulate personal anxiety-related goals that they aim to achieve by participating in the training and to formulate their most important reasons to change. In addition, they could earn points based on performance during training.

The pre-registered primary outcome was trait anxiety which was assessed at pre-, mid-, and post-training and follow-up at one, two, and three months. The cognitive outcome measures were AB and WM capacity which were assessed at pre-, mid- and post-training and 3 months follow-up. We hypothesized that the double active training condition (active ABM & active WMT-ED condition) would outperform the other conditions and that the single training conditions (active ABM & control WMT-ED; control ABM & active WMT-ED) would outperform the double control condition in improving trait anxiety (primary outcome) and the cognitive outcomes of AB and WM capacity (secondary outcomes). Secondly, we hypothesized that higher levels of state anxiety prior to training would result in stronger effects of ABM (pre-registered moderator) on trait anxiety, AB, and WM capacity. In addition, we exploratively examined whether higher AB and lower WM capacity at baseline would result in stronger training effects.

The aim of this study was to examine the effects of unguided web-based Attention Bias Modification training (ABM) combined with an affective Working Memory training in adults with heightened anxiety. The final sample consisted of 433 adults with elevated levels of trait anxiety as well as borderline to clinical levels of other symptomatology. Approximately 40.6% (n = 176) completed all ten training sessions and only 28.9% (n = 125) completed FU3. It was hypothesized that active ABM, active WMT, and the combination would outperform the other training combinations in reducing trait anxiety. However, the data did not support this hypothesis, as an overall decline in trait anxiety was observed, irrespective of ABM and WMT training combinations. Concerning effects on cognitive outcome measures, stronger reductions in AB in the ABM training compared to ABM control were observed. With regard to improvements in WM capacity, we observed that ABM combined with WMT (double active) outperformed two other training combinations (ABM + WMT control and WMT + ABM control), but surprisingly not the double control condition. With respect to moderation of training effects; higher levels of state anxiety prior to training did not result in stronger ABM effects, nor did baseline levels of AB or WM capacity moderate effects.

Contrary to expectations, the combination of ABM and affective WMT did not outperform the single training conditions, nor the double control condition in reducing trait anxiety. While theoretical models regarding cognitive vulnerability to anxiety propose that strong AB and low levels of WM capacity jointly determine anxiety (e.g., Ouimet et al., 2009), the current findings suggest that simultaneously modifying both processes does not add anything in anxiety reduction compared to modifying only one of the processes. One possible explanation might be related to the finding that each separate training (either active ABM or active WMT) did not result in stronger reductions in trait anxiety compared to the double control condition. While synergetic effects could occur when combining two types of training, it could be argued that the full potential of combining both training paradigms can only be optimally tested when the individual training-procedures are successful in improving their targeted cognitive processes and anxiety.

With respect to ABM, the finding that ABM training did not result in stronger anxiety reductions than control ABM is in contrast with findings from meta-analyses (Fodor et al., 2020; Linetzky et al., 2015; Price et al., 2016). A recent network meta-analysis (Fodor et al., 2020) showed that when excluding post-traumatic stress trials, ABM significantly reduced anxiety compared to waitlist and control training. There are several potential explanations for this discrepancy. First of all, in almost 90% of the ABM studies a dot probe paradigm has been used (Martinelli et al., 2022), while the current study employed a visual search paradigm. Thus, the type of ABM training might play a role in ABM’s effectiveness in reducing anxiety (see also Vrijsen et al., 2024). Secondly, it has also been shown that delivery setting plays a role as stronger effects have been observed in lab-based or offline studies (Cristea et al., 2015; Fodor et al., 2020; Heeren et al., 2015; Martinelli et al., 2022; Linetzky et al., 2015), while the current study was web-based. Thirdly, previous visual search ABM studies only observed effects on clinician- or parent-rated assessments of symptoms and not on self-report measures (Waters et al., 2013, 2015). This is consistent with meta-analyses indicating larger ABM effects with clinician- versus self-report outcomes (Linetzky et al., 2015; Price et al., 2016). As clinician-rated assessments also include general functioning, these might be more sensitive in capturing changes in daily functioning compared to self-report questionnaires focusing on symptoms only. Or that it takes longer for an individual to recognize change in one-self compared to a trained rater. An interim conclusion might be that web-based, visual search ABM with self-reported anxiety symptoms as an outcome measure might be sub-optimal to achieve and test reductions in anxiety.

A fourth aspect that is relevant to discuss is the impact of ABM on AB. Consistent with previous studies (De Voogd et al., 2016b, 2017; Waters et al., 2015), this study observed a larger decrease in AB after visual search ABM compared to ABM control. However, this decrease in bias was not accompanied by a decrease in trait anxiety. This result is in line with findings from other web-based studies which observed a general decline in anxiety symptoms irrespective of training condition (Boettcher et al., 2012; Carlbring et al., 2012; De Voogd et al., 2016b, 2017). In addition, there was no correlation between change in bias and change in anxiety from pre- to post-training. These results contradict the conclusion of Grafton et al. (2017) who found that if a bias is successfully modified, this is accompanied by a decrease in anxiety. It could be that the observed change in bias could merely reflect a practice or task-specific effect as the ABM training was identical to part of the AB assessment. In line with this idea, some studies suggested that training effects of visual search ABM do not transfer to other assessment tasks (Kruijt et al., 2013; De Voogd et al., 2016b; but see Dandeneau et al., 2007). Together with recent reviews on ABM in anxiety (Mogg et al., 2017), these findings highlight the need for including different bias assessment tasks in order to delineate task-specific versus general changes in AB. Additionally, complementing RT measures with eye-tracking measures might give a more detailed account of changes in bias as RT measures give little insight into the time course before a response and often have low reliability (Waechter & Stolz, 2015).

With respect to affective WMT, this training did not outperform the double control condition in either anxiety reduction or increase in WM capacity. This result is similar to a study using the same WM paradigm with emotional distractors in a general population of adolescents (De Voogd et al., 2016a). However, these findings are in contrast to a recent meta-analysis (Wang et al., 2023) that concluded that WM training had a small, but significant effect on anxiety. There might be several reasons why the affective WMT in the current study might not have affected anxiety. First, given that the working mechanism of WMT is increasing WM capacity, it is not surprising that when the training failed to systematically change WM capacity, no effects on anxiety were observed. Secondly, the performed sub-group analyses in the meta-analysis (Wang et al., 2023) revealed a moderating role of type of anxiety and training setting. While WM training alleviated test anxiety and social anxiety, the effects on trait anxiety (the primary outcome in the current study) were not significant. In addition, offline training resulted in significant effects on anxiety, and effect of web-based training (in the current study) was inconclusive. Thirdly, it should be noted that the current study employed a WM training-variety with emotional distractors, while the assessment did not contain emotional distractors. Hence, the assessment task can be considered a far-transfer task as it contains the same visuospatial aspects but does not contain the emotional context. Perhaps a more near-transfer emotional WM task would have been more sensitive in detecting affective WM capacity changes (see for example Schweizer et al., 2011). Another possible reason why WMT did not outperform its control version is that even though only the active training was adaptive, both training varieties contained an emotional distractor. Thus, in both trainings inhibition of negative emotional stimuli was practiced. Moreover, the WMT was not adaptive with respect to the number of emotional distractors. As inhibition of task-irrelevant information is proposed as a deficit in anxiety (Berggren & Derakshan, 2013), training effects on WM capacity and anxiety might be improved by increasing the number of negative emotional distractors which facilitates inhibition of task-irrelevant stimuli. Lastly, the evaluation questions indicated that the majority of participants rated the active WMT as difficult, not very credible, and with an unclear aim. It might be that the WMT was too challenging and in turn may have limited the training effects by demotivating participants to improve their training performance. As participants were aware that there was a control version, the low credibility might have resulted in making participants less motivated and engaged with the training as they might have thought they had the control version. Engagement with the training can possibly be improved by providing a rationale prior to the training but also provide information on the relevance of the training during the intervention.

Contrary to our expectations, trainings effects were not moderated by pre-session state anxiety, baseline AB, or WM capacity. The absence of moderation effects might be explained by our selected sample of high anxious individuals that might have limited the variability in each of these moderators. Although there was adequate variability in baseline AB and state anxiety, there was little variability in baseline WM capacity. Furthermore, since state anxiety levels were not elicited or manipulated before training (Kuckertz et al., 2014), state anxiety levels were quite low on average which might impede the effects of higher versus lower levels of state anxiety. In order to better evaluate how state anxiety might affect ABM, future studies should consider increasing state anxiety during training (e.g., either by training in an anxious situation or by inducing state anxiety) (e.g., Nuijs et al., 2020) and include assessments of state anxiety during training.

This study has several limitations that need to be addressed. First, the drop-out between the pre-assessment and mid-assessment was quite high (41.7–52.6% per condition). While common in web-based studies (De Voogd et al., 2016b, 2017), it might have compromised the interpretability of the final models. The main reason for quitting the training was a low satisfaction with the web-based training which might be best explained by the negative evaluation of the WMT in terms of its enjoyability and rationale. One possible solution could be to offer participants more therapist guidance during the intervention, which is suggested to be an important success factor for web-based interventions (Palmqvist et al., 2007; Spek et al., 2007). Besides the possible benefit of web-based guidance from a therapist, some studies suggest that the user’s experience of cognitive training paradigms can be improved by using virtual reality technology (Otkhmezuri et al., 2019) or more engaging gamified protocols (Notebaert et al., 2018). Second, there was large variability in when participants finished training sessions and completed the following assessments. Hence, participants might have completed post-assessments months after they had finished the training which impedes differentiating between short-and long-term training effects. Although web-based training has advantages as it can be provided at low cost and offers possibilities for wide-scale implementation, it offers little experimental control with respect to the training conditions. Therefore, the findings should not be interpreted as a lack of effectiveness of ABM and affective WM training in more controlled settings such as lab- or clinical contexts (Wiers et al., 2018).

In sum, this study suggests that combining visual search ABM with affective WMT (using a chessboard training paradigm with emotional distractors) in a web-based format is not effective in reducing trait anxiety as we observed general improvements in anxiety over time irrespective of training condition. Consistent with meta-analytic findings, this might be related to the setting of training (web-based versus offline), type of ABM (visual search versus dot probe training), type of anxiety (trait anxiety), or type of anxiety assessment (self-report versus clinician-rated). As web-based interventions that include therapist contact are suggested to be more successful than interventions with no or little therapist contact, testing the effectiveness of cognitive training paradigms in a blended care format might be the way forward.