Anxiety spectrum disorders are characterized by excessive and uncontrollable worrying about negative events that may occur in the future (1, 2). The basic psychopathology of anxiety spectrum disorders is related to an attentional imbalance in processing threat vs. safety signals in the environment. In particular, the encoding, maintenance, and retrieval of information could be affected by anxiety spectrum disorders in a way that leads to excessive worrying (3–5). A significant body of literature suggests that individuals with anxiety spectrum disorders exhibit different cognitive dysregulations that could underlie and exacerbate their symptoms (2, 4, 6, 7).

Working memory (WM) includes different subcomponents including encoding, maintenance, and retrieval of information (8). The expression of various anxiety symptoms could be attributed to attentional bias toward task-irrelevant thoughts (e.g., worry/threats) and their subsequent processing (3, 4). Some studies have investigated the interaction between anxiety symptoms and WM in healthy individuals. For example, it was reported that increasing WM load reduces anxiety given the diversion of executive resources toward the increased task difficulty (9–11). This was attributed to the recruitment of top-down executive control mechanisms to override the potential involvement of the threat-detection system (4, 12). Other studies suggested no general WM deficits in healthy individuals with trait anxiety (13). Rather, healthy individuals with high trait anxiety failed to filter threat-related information from WM (14, 15). With conflicting results, previous research did not offer sufficient dissociation between WM components in the modulation of anxiety symptoms in healthy individuals.

Studies investigating WM in patients with anxiety disorders focused on the effect of WM load on anxiety symptoms under safety vs. threat. Najmi et al. and Vytal et al. showed that patients with GAD outperformed healthy individuals in WM tasks that manipulated load under threat, but not in a safe, context (16, 17). Conversely, Park et al. reported that patients with GAD were impaired in WM duration and maintenance tasks (18). Similarly, patients with social anxiety disorder (SAD) performed more poorly than healthy individuals in a WM load task that tested the recall of words based on social and threat inference. However, patients with SAD performed better with the social threat condition compared to the neutral condition, thus demonstrating a potential processing bias for social-relevant information (19). Aside from manipulating the WM load, complexity of the task, or threat/safety, previous studies did not delineate the WM mechanisms in anxiety states (in healthy individuals) or traits (patients with anxiety spectrum disorders). In part, this could be attributed to the inconsistency of the WM tasks and domains that were evaluated in previous studies. This is further compounded by the lack of mapping of different components of WM onto different components of anxiety in different disorders.

To address these disparities, we used three computer-based tasks that target various domains of WM in three major types of anxiety spectrum disorders: GAD, SAD, and panic disorder (PD), as well as matched healthy individuals. We assessed WM duration and maintenance using the short-delay WM and long-delay WM tasks (20). We also used the N-Back task to examine WM load (21–23). We predicted that patients with different types of anxiety spectrum disorders (GAD, SAD, and PD) would exhibit different cognitive and neuropsychological impairments in different domains of WM. We also examined the accuracy of relying on WM components to differentiate patients with anxiety spectrum disorders from healthy subjects.

In this study, we explored the different domains of WM in anxiety spectrum disorders. Patients with SAD and GAD showed impairment in WM duration (long-delay WM task) and load (N-back task) compared to patients with PD and healthy individuals. Importantly, PCA of the WM variables allowed for the uncovering of a single principal WM component which drives this difference (PC-1). We utilized the different principal components in the uncorrelated WM data matrix to significantly predict the severity of anxiety symptoms across the different groups. Finally, we showed the clinical utility of the WM principal components in identifying anxiety spectrum disorders using logistic regression and ROC analysis. To our knowledge, this is the first study to evaluate the principal components of WM and their utilization as potential screening measures in anxiety spectrum disorders.

The novel approach of our study lies in the utilization of uncorrelated/whitened WM components using PCA to identify hidden data structures in raw WM variables. We used the uncorrelated data components from six different WM tasks/questionnaires evaluated within-subjects to make inferences to anxiety spectrum disorders and symptom severity. Common WM measures produce highly correlated results. Although factor analysis was first developed to measure IQ based on different psychometric scores by Spearman in the early 20th century (35), in recent decades, little research utilized orthogonal feature extraction, such as Hotteling transform, on cognitive data to eliminate covariance and examine the collective underlying data structure across different tasks. Instead, previous research focused on defining principal components to explain some, but not all, the variability in cognitive samples. As a result, the reliance on raw cognitive measures or a handful of components to understand psychopathology, as in anxiety spectrum disorders, can overlook the principal mechanisms. In our study, we just transformed the raw data from five different WM measures to an uncorrelated matrix with a similar number of components using PCA without truncating the explained variability. The underlying data structure clearly shows that if we were to utilize dimensionality reduction of PCA, we will only retain the first two components with Eigenvalues that are larger than 1. However, and given that we approached the current problem without a priori hypotheses about WM or anxiety spectrum disorders, we chose to retain all components that pertain to the data and explore the different WM mappings onto anxiety spectrum disorders.

The only significant difference between the various groups (GAD, PD, SAD, HC) was in WM PC-1. With 90% loadings on the N-Back and WM Long-Delay tasks, PC-1 could represent the principal measure of WM maintenance and/or retrieval from the pooled covariance of different mechanisms underlying WM function, including duration, load, maintenance, and retrieval. The GAD and SAD groups exhibited significantly lower performance than PD and HC in PC1. Several studies reported that patients with GAD did not differ from healthy individuals in WM load accuracy in a safe context in the N-Back task (16, 17, 36, 37). However, in these papers, patients with GAD did not exhibit higher anxiety-induced startle or anxiety ratings compared to healthy individuals (17). In an earlier experiment, Vytal et al. suggested that high-load WM in the N-Back task reduced anxiety levels in healthy individuals (10). In patients with various anxiety spectrum disorders, MacNamara et al. reported WM impairment with increased load and PD symptoms in the N-Back task (38). All of these previous studies used a safe/threat version of the N-Back task to study WM load. Moreover, Park et al. reported that patients with GAD were impaired in WM duration and maintenance tasks (18). In line with Eysenck's attention control theory (4), and given the nature of the psychopathology in GAD and other anxiety spectrum disorders, it is expected that any threat encounter will create a “negative context” even for neutral conditions. This is confirmed by earlier findings in a neutral version of the N-Back task, similar to the one we used, where patients with GAD showed significant impairment in high-load conditions only (39). Further confirmation comes from other studies of WM capacity using neutral/threat stimuli. Patients with SAD exhibited conflicting results across several studies, with better performance under threat (19), or no difference from healthy individuals in WM capacity (40). Our results replicated previous findings in the neutral experiments both in our raw (Figure 1) and covariance-free WM variables across anxiety spectrum disorders (Figure 3). We showed that a WM maintenance and/or retrieval deficit exists only in GAD and SAD but not in PD. There were no other differences between anxiety spectrum disorders and healthy individuals in any of the other WM components. Our findings alongside the shortcomings of raw WM measures could explain the conflicting results in previous literature addressing WM in anxiety (41–43). These cross-diagnostic findings based on principal WM components shed much needed attention on the mapping of anxiety spectrum disorders onto WM mechanisms.

Eysenck's attention control theory suggests that anxiety impairs cognitive function, such as WM, by increasing stimulus-driven attentional processing at the expense of goal-directed attentional processing (4). This theory is supported by various empirical findings supporting anxiety-related impairment in verbal reasoning (44), spatial reasoning (45), reading comprehension (42), verbal working memory (46), and sustained attention (47). Collectively, these findings support the involvement of WM mechanisms in anxiety states and anxiety spectrum disorders (48). However, it will be impossible to dissociate the different WM mechanisms without conducting a within-subject evaluation of various facets of WM in multiple anxiety spectrum disorders. Our study utilized verbal, numerical, and sequential evaluations of WM to cover most, if not all, possible facets of the WM data space. Following PCA, it was evident that only two WM principal components sufficiently represent the data space; PC-1 representing maintenance, and PC-2 representing encoding. We intentionally avoided naming the WM principal components. Based on previous studies on uncertainty and error processing in anxiety spectrum disorders (49), we decided to analyze all components to account for all the variance in our dataset.

We found that different WM principal components significantly predict anxiety symptom severity across different anxiety spectrum disorders and healthy individuals. The explained variance in anxiety symptoms by WM principal components outperformed that from WM raw variables. Nevertheless, the predictive power of WM components did not exceed 50% of variance in anxiety symptom severity. Lukasik et al. found negative correlations between WM measures and anxiety, but not stress (48). However, they did not explore the different WM principal components or mechanisms in their dataset despite using 10 WM tasks. Moreover, Berggren et al. showed that self-reported state anxiety negatively correlated with WM performance in healthy individuals (50). One study also found that WM capacity and efficiency predicted symptoms of anxiety and distress in healthy individuals (51). In our study, we report for the first time how different WM components can predict the severity of anxiety symptoms in different anxiety spectrum disorder diagnoses as well healthy individuals. Ultimately, thorough examination of cognitive function across anxiety spectrum and other psychiatric disorders could lead to the useful insights about the fortification of current diagnostic systems.

We examined the ability WM components, not raw measures, to identify anxiety spectrum disorders. Based only on WM PC-1 and PC-2, we could differentiate patients with anxiety spectrum disorders from healthy individuals with 83.7% accuracy (Figure 4). Thus, performing two 5-min computer-based cognitive tasks could be up for further evaluation especially if the current accuracy holds in larger samples with different diagnosis categories. As opposed to lengthy clinical interviews (52), such cognitive tools would offer swift screening results even if the current signal only denotes a non-specific presence of a mental illness. Considering the serious concerns about the validity and reliability of psychiatric diagnoses not exceeding 60% (53), adding cognitive profiles to the diagnostic process for anxiety disorders could significantly improve the outcomes. We build this proposal on previous studies that consistently link cognitive dysfunction with various mental disorders using objective, language-independent, culture-independent cognitive paradigms that are based on animal behavior research (54–57). Future studies with larger datasets ought to provide more evidence to support cognitive screening, and possibly diagnostics, for mental disorders.

The de-identified data supporting the conclusions of this article can be made available by the authors upon request.

The studies involving human participants were reviewed and approved by Ethics committee at Ain Shams University School of Medicine. The patients/participants provided their written informed consent to participate in this study.

OA contributed to data analysis and wrote the manuscript. MH conducted the data analysis, wrote the results, and wrote and finalized the manuscript. AM designed the study, collected the data, and edited the manuscript. EA collected the data and edited the manuscript. HK formatted the data, contributed to data analysis, and edited the manuscript. All authors contributed to the article and approved the submitted version.

This study received funding from Arab Palestinian Investment Company (APIC), Hikma Pharmaceuticals LLC., and GMS Holdings. The funders were not involved in the study design, collection, analysis, interpretation of data, the writing of this article or the decision to submit it for publication. The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.