Anxiety disorders are the most common class of mental illnesses, with a 31% lifetime prevalence [1]. This has already increased during the COVID-19 pandemic [2-4]. For example, in the United States, the Census Bureau reported that adults were more than 3 times more likely to screen positive for an anxiety disorder in 2020 than they were in 2019 (31% vs 8%), and a cross-sectional survey of people in China in 2020 reported the prevalence of generalized anxiety disorder (GAD) to be 35.1% [2]. A recent meta-analysis of 17 studies (N=62,000) found an average prevalence of anxiety of 32% during COVID-19 [5].

However, anxiety, particularly GAD, is difficult to treat. Current practice guidelines recommend pharmacological and psychological interventions [6], but most patients favor psychotherapy over medications [7]. Benzodiazepines have the risk of tolerance and addiction; the United Kingdom National Institute for Health and Care Excellence guidelines state that benzodiazepines “should not be used routinely to treat anxiety disorders” [6]. First-line treatment for anxiety, such as selective serotonin reuptake inhibitors, can have limitations, including delayed patient responses and adverse effects (eg, gastrointestinal and sexual) [8,9]. The number needed to treat (NNT) for selective serotonin reuptake inhibitors is 5.2; one needs to treat >5 individuals to see a significant response in 1 individual [10].

Cognitive behavioral therapy (CBT), the most commonly used and researched psychological intervention for anxiety, has been associated with reduction in symptoms using measures such as the Beck Anxiety Inventory (small to medium effect sizes), and yet is typically delivered in-person [11]. Given the growing need and shortage of trained therapists [12,13], new mechanistically based treatments that can be delivered at scale and a distance are needed for those unable to attend in person because of barriers such as living in a resource-limited area or unwillingness to go to a mental health clinic because of stigma. In addition, treatments that target more recent mechanistic underpinnings of anxiety are needed.

From a theoretical standpoint, reinforcement learning mechanisms have been suggested to drive anxiety disorders [14,15]. Worry is widely regarded as the central defining feature of GAD and has been shown to be triggered as an avoidance reaction to emotional experiences [16], thus learned as “a negatively reinforced avoidant behavior” [17]. Recent research has linked reinforcement learning with biological mechanisms and clinical symptoms [18]. Worry represents an attempt to engage in mental problem solving on an issue with an uncertain outcome [19]. Although worry is unpleasant, the immediate emotions that are avoided by focusing on worry, such as fear, are often perceived as more unpleasant [20,21]. Therefore, aversive stimuli can trigger worry as an avoidant behavior, which then becomes habitual [22]. Theoretically and mechanistically, worry is learned and reinforced in a manner similar to other types of operantly conditioned behaviors [14]. With cycles of reinforcement learning, when triggered by feelings of anxiety, individuals with GAD learn a maladaptive thinking style that uses worry as a focus on the future rather than the present. As described by Borkovec et al [14,15], when worry becomes habitual, its negative reinforcement pathway can spiral out of control; when the negative emotional experience of worry rises to the same level as unpleasant emotional states that trigger it, it may become its own trigger for more anxiety, which leads to more worry directly driving anxiety habit loops (Gao et al, unpublished data, 2021).

In conventional frameworks, anxiety is conceptualized as an overestimation of danger and an underestimation of one’s ability to cope with it [23]. Cognitive therapies aim to interrupt the cycle of worry by replacing maladaptive cognitions with more constructive ones. For individuals who cycle into anxious worry to a degree that impairs prefrontal cortical function and the ability to use cognitive therapies [24], other strategies are needed. New treatments such as mindfulness training (MT) have shown promise in efficacy and cost, with effect sizes rivaling current treatments [25]. Furthermore, cross-sectional studies of nonclinical populations have suggested a mediating role of worry in the effects of MT on anxiety [26].

Mindfulness can be defined as the awareness that arises when paying attention to the present moment on purpose and nonjudgmentally [27]. The attitudinal quality of not judging and allowing experience to unfold with curiosity targets maladaptive reinforcement learning by helping individuals to simply observe repetitive cycles of perseverative worry rather than to habitually react and reinforce them [28]. MT has been found to mechanistically break key links in the reinforcement pathway for other habitual behaviors such as smoking and emotional eating [29,30], with concomitant changes in related brain regions predicting clinical outcomes [31]. Worry has been shown to activate brain networks associated with self-referential processing, such as the default-mode network, suggesting that the more one is caught up in perseveration about uncertain events, the more this network is activated [32,33]; meditation has been shown to directly deactivate these brain regions [34], and neurophenomenological studies suggest that this may be because of the ability to observe thoughts and emotions rather than being caught up in them [35-37]. Specific to GAD, MT has been associated with changes in the fronto-limbic brain regions involved in emotion regulation with simultaneous improvements in reported symptoms [38].

Regarding treatment delivery, mindfulness-based interventions, such as mindfulness-based stress reduction, are generally delivered in a group format. However, concerns remain regarding the scalability of in-person–delivered treatment [39]. Digital therapeutics (ie, app-delivered interventions) have garnered much attention as a new modality that can deliver high-fidelity treatment at scale and low cost. However, empirically tested apps are not widely used, and widely used apps do not have an evidence base (<0.05%)—a clear “digital research practice gap” [40]. To date, only 1 study has reported the clinical efficacy of a digital therapeutic specifically for people diagnosed with GAD (showing that CBT may be effective when delivered in this format) [41].

To address the digital research practice gap, we designed an app-delivered digital therapeutic program for anxiety that mechanistically targets reinforcement learning using MT to help individuals identify habitual worry thinking patterns and learn not to habitually react to unpleasant emotions (ie, break the worry cycle). In a single-arm study of anxious physicians, we found preliminary evidence for its utility in reducing anxiety (57% reduction in GAD-7 scores after 3 months) [42]. However, randomized controlled studies are required to determine the efficacy and mechanisms of action. We tested the following hypotheses in a randomized controlled trial of individuals with GAD: (1) app-delivered MT would show superior efficacy in reducing anxiety and worry than standard treatment; (2) increases in nonreactivity would mediate reductions in worry; and (3) reductions in worry would mediate reductions in anxiety.

We used a parallel-group randomized controlled trial design with analyses and outcome measures preregistered at ClinicalTrials.gov. Individuals were recruited using Facebook advertisements and screened for eligibility via a phone call by the project director (AR). Textbox 1 details the inclusion and exclusion criteria. These criteria were chosen to mimic, as closely as possible, real-world clinical situations while minimizing potential confounders (eg, a recent change in medication dose may mask treatment effects). Eligible participants attended an in-person research visit at Brown University, where they underwent informed consent procedures with the project director before enrolling in the study. Participants were provided Amazon gift cards worth up to US $80 to complete the questionnaires. This study was approved by the Institutional Review Board of Brown University.

After enrolling participants in the study, the project director, who had previously undergone training and was supervised by a psychiatrist, conducted an in-person, abbreviated version of the Mini-International Neuropsychiatric Interview to confirm a diagnosis of GAD, along with the assessment of other potential comorbid disorders, such as major depressive episode, panic disorder, agoraphobia, social anxiety disorder, obsessive compulsive disorder, and posttraumatic stress disorder [44]. Participants were then asked to complete a web-based questionnaire via Qualtrics [45], which included demographic and self-report items. Participants in both groups received questionnaires via email 1 and 2 months after treatment initiation. Upon completion of the final questionnaire, the treatment as usual (TAU) group received instructions on how to download and install the app.

After completing the baseline questionnaire, participants were given a sealed, opaque envelope (prepared, reviewed, and sealed by individuals independent of the study team) by the project director that contained their group assignment: TAU+app-delivered MT or TAU. The 1:1 randomization scheme was generated by an independent statistician with variable block sizes of 4 and 6. Team members who randomized the participants and carried out the study procedures did not perform the study analyses. The principal investigator and the statistician who conducted the statistical analysis were blinded to the group allocation until all analyses were complete.

The app-delivered MT program, Unwinding Anxiety (version 2), is a Health Insurance Portability and Accountability Act compliant digital therapeutic that teaches individuals to understand how anxious worry develops and perpetuates through reinforcement learning and how to bring mindful awareness to moments of stress and worry so that they can observe feelings of anxiety rather than perpetuate reactive worry thinking. This process helps individuals unlearn or extinguish worry at the core mechanistic level. This experiential education is delivered via a smartphone-based platform, which includes a progression through >30 daily modules of brief didactic and experience-based MT (videos and animations approximately 10 minutes per day; Multimedia Appendix 1), app-triggered check-ins, user-initiated guided meditations (5-15 minutes), and brief (30 seconds) on-demand mindfulness exercises to help disrupt anxiety cycles in vivo (Textbox 2; Multimedia Appendix 2). The content for this intervention was developed based on a combination of clinical experiences and previously developed, in-person and app-delivered MT protocols for habit change that have yielded clinically meaningful outcomes (eg, smoking and overeating) [28-30,46-49]; an open-label pilot study of the app demonstrated a 57% reduction in GAD-7 scores in anxious physicians [42].

Individuals randomized to TAU+app-delivered MT were assisted with the installation of the app on their smartphone and the reviewing of the features. They were instructed to complete 1 module per day over the subsequent 30 days at a time of their choice. In addition, the intervention would check in with them 3 times throughout the day (this could be modified by the user) and offer brief mindfulness exercises. Participants were encouraged to use other app features but were informed that this was not a requirement for the study. The project director sent check-in messages on days 3, 7, 14, and 21 from treatment initiation to help mitigate technical difficulties and encourage engagement. Specifically, participants were asked “how things were going with the app since the last time they received a check-in.” If the participant expressed difficulties, efforts were made to resolve the problem.

As part of TAU, participants were asked to continue the standard care set forward by their clinician or clinicians. This could include pharmacological treatment or psychotherapy. Participants were also provided with a list of local resources.

The primary outcomes were changes in anxiety, as measured by the GAD-7, and emotional reactivity at 2 months after treatment initiation. Secondary outcomes included changes in worry, as measured by the Penn State Worry Questionnaire (PSWQ) and interoceptive awareness.

As there are no prior studies using the intervention, the target sample size of 65 was determined using pilot data. A pilot study of individuals with GAD-7 scores >9 (n=17) demonstrated a 39% reduction in the scores after completing 21 modules. Assuming a 10% reduction in the TAU group, a 1-tailed t test determined that a sample size of 52 would have 80% power with 1-sided 5% type I error to detect a statistically significant between-group difference in GAD-7 scores (Cohen d=0.7). We recruited 65 individuals to account for 25% attrition.

Analyses were conducted using R (version 3.4.1), and a modified intent-to-treat approach was employed. This was defined as all participants who downloaded the app and completed all study assessments, regardless of treatment completion. A robust, mixed analysis of variance (ANOVA) was used to evaluate the primary outcomes, changes in anxiety, and emotional reactivity 2 months after treatment initiation in the TAU+app-delivered MT group relative to those in the TAU group. This and other robust statistical tests were chosen to avoid violating underlying model assumptions, such as normality. Because of the use of these nonparametric tests, the median and IQR were calculated. Post hoc comparisons of single effects were performed using the WRS2 package in R. In addition, Mann-Whitney U tests were performed, and the Hodges-Lehmann estimate of location shift was used to calculate the difference between groups at each time point [58]. Bonferroni correction was used to adjust for multiple comparisons. Effect sizes (r) were calculated by dividing the z score by the square root of the sample size using the Cohen criteria for r, where 0.1 is small, 0.3 is medium, and 0.5 is large [59].

We then conducted exploratory mediation analyses to evaluate (1) model A, whether increased nonreactivity mediated the relationship between MT and reduced worry, and (2) model B, whether reduced worry mediated the relationship between MT and reduced anxiety. To reduce the impact of the unit on the comparison of direct and indirect effects, the variables were standardized to have 0 means and unit SD. To have causal interpretations, the mediation models were built on the longitudinal [60]: In model A, MT was the independent variable, worry was the dependent variable, and nonreactivity was the mediating variable (Figure 1); in model B, MT was the independent variable, anxiety was dependent variable, and worry was the mediating variable (Figure 1). We calculated the direct and indirect effects, for which the SEs and 95% CI were computed using the bootstrap method with 1000 bootstrapped resamples [61].

The NNT, defined as the total number of individuals who need to receive treatment to prevent 1 adverse event, is a standard epidemiological measure used to communicate the effectiveness of a health care intervention [62]. The inverse of the absolute risk reduction was calculated by subtracting the total percentage of individuals who achieved remission (GAD-7 score ≤5, minimal anxiety) in the TAU group from the total percentage of those in the TAU+app-delivered MT group.

Unlike statistical significance, clinical significance has traditionally lacked a consistent definition [63]. To address this need, a reliable change index (RCI) was created to evaluate the reliability of clinically significant changes [63]. We used the method developed by Jacobson and Truax to calculate the RCI for changes in anxiety scores [64]. If the RCI exceeded the z-scored level of significance from –1.96 to +1.96 (P<.05), we evaluated the percentage of participants with clinically significant change who met or exceeded it at 1 and 2 months after treatment initiation [63].

Six areas of potential bias across 7 domains were assessed using the Cochrane Collaboration tool for evaluating the risk of bias [65].

We recruited 65 participants, obtained their consent to participate, and randomized them between May 2019 and October 2019. Baseline demographic characteristics are reported in Table 1. Of the 65 participants, 61 completed the study and were included in the modified intent-to-treat analysis (Figure 2). Before treatment initiation, 30% (19/63), 25% (16/63), and 32% (20/63) of participants reported comorbid anxiety, depression, and anxiety and depressive disorders, respectively (Table 1).

There were no adverse events in the TAU group and 11% (3/28) of adverse events in the TAU+app-delivered MT group (2/3, 66% anxiety and 1/3, 33% back pain).

At 2 months after treatment initiation, we found an NNT of 1.6: 64% (18/28) of the participants achieved remission in the TAU+app-delivered MT group compared with 3% (1/33) in the TAU group.

Using the Cochrane Collaboration criteria for evaluating bias [65], we found a low risk of bias in 6 of 7 categories, including random sequence generation, allocation concealment, blinding of outcome assessment, incomplete outcome data, selective reporting, and other bias. Blinding of participants and personnel was deemed to have a medium risk of bias because the project director and participants were unblinded to group allocation.

Anxiety is a debilitating and difficult-to-treat condition that affects hundreds of millions of people worldwide. Using a theory-based approach (targeting reinforcement learning), we developed a digital therapeutic that demonstrated a significant and clinically meaningful reduction in anxiety in individuals with GAD (NNT=1.6) [66]. This was confirmed by our finding that anxiety decreased with further progression through the program through the completion of more modules. Furthermore, we determined a potential mechanism underlying its effect: increases in mindfulness mediated decreases in worry and decreases in worry mediated reductions in anxiety. These effects were specific to the MT intervention. The mediation effect was significantly higher at 2 months (45%) than at 1 month (15%), which is consistent with participants having more exposure to the treatment (ie, a dose effect), a sleeper effect [67-69], or a combination of the two. These results are in direct alignment with the theoretical underpinnings that anxiety can be perpetuated through negative reinforcement—worry can feed back and perpetuate anxiety by introducing the reward of feeling more in control or temporarily distracting an individual from the aversive feelings of anxiety [14,15,21,42,70].

Anxiety treatment has largely relied on antidepressant medications and psychotherapy (eg, CBT). These have yielded medium effect sizes for anxiety [9]. RCI is increasingly being used in treatment studies to assess whether changes are clinically significant [63]. A longitudinal study of low-intensity CBT using an RCI of ≤5 demonstrated a reliable change in 43.8% (181/439) of the participants [71]; our study found that 64% (18/28) and 54% (15/28) of the participants demonstrated reliable change at 1- and 2-months posttreatment using RCIs of 4.6% and 7.9%, respectively. For GAD, the NNT with antidepressants was 5.15. In this study, we found that specifically targeting a mechanistic pathway yielded large effect sizes with an NNT of 1.6. A previous single-arm pilot study in physicians with comparable levels of anxiety (median baseline GAD-7 score of 11.5) showed a similar magnitude of reduction in anxiety (57% reduction at 3 months) using the same app-delivered MT program [42]. This randomized controlled trial extends previous results and broadens these findings beyond anxious physicians to individuals with moderate to severe anxiety.

Potential mechanisms underlying anxiety have been hypothesized for over a century; yet, refinement in recent decades has opened the door for specific hypothesis testing [14,15]. For example, Mkrtchian et al [18] recently demonstrated avoidance as a part of reinforcement learning pathways in individuals with anxiety disorders. Our findings provide an important extension of these results by showing that targeting worry and avoidance yields clinically meaningful reductions in anxiety. These results and the finding that individuals reported increases in interoceptive awareness (measured by the MAIA) are in line with broader theoretical mindfulness frameworks that suggest that MT helps individuals learn to become more aware of and observe unpleasant emotions with awareness imbued with curiosity [72].

MT may promote decentering, defined as a “metacognitive capacity to observe items that arise in the mind as mere psychological events” [73]. Decentering may help individuals disengage from perseverative worry habit loops that are perpetuated through reinforcement learning [46-48,74]. Our findings show that increases in mindfulness directly mediate the effects of app-delivered MT on reductions in worry. This may be the case possibly because of helping individuals step out of perseverative worry habit loops that are at the core of GAD and, in doing so, reduce their reinforcement. Furthermore, our results show that reductions in worry mediate the effects of MT on reductions in anxiety.

The high prevalence of anxiety “vastly exceeds the capacities of mental health services,” and this gap has only increased over the past several years [75]. App-based digital therapeutics offer a viable and practical route toward augmenting traditional mental health care and, in some cases, serve as a first-line treatment [76]. For example, if a patient in a primary care clinic screens positive for anxiety, an evidence-based digital therapeutic such as the one described in this study can be offered as an augmentation to standard medication treatment, or in some cases, it can be offered if a trial of medications has yielded suboptimal results or if a patient is not interested or willing to try a medication as an alternative. In addition, for individuals who are concerned about the confidentiality of mental health care (eg, feeling the need to ask a boss for regular time off for therapy visits), Health Insurance Portability and Accountability Act compliant digital therapeutics can offer discretion, privacy, and convenience. As integrative care models (eg, embedding psychiatric or psychological services within primary care clinics) gain momentum, one of the primary limitations is the cost and availability of trained therapists. However, because 85% of the US population has a smartphone, digital therapeutics may be able to serve as the mobile component of an integrative care clinic at a low cost, filling in for the lack of physical space and trained mental health clinicians [76]. In addition, in corporate settings where employers are increasingly aiming to meet the mental health needs of employees, it may be possible to quickly and confidentially deploy evidence-based digital therapeutics to help employees with mild or moderate anxiety. For employees with severe anxiety, who may have to wait several months to see their doctor or to obtain a mental health referral, a digital therapeutic may serve as a bridge to therapy or even a first-line treatment.

We chose the GAD-7 as an outcome measure because it is widely used in clinical practice, yielding results that are interpretable in nonresearch settings. We chose TAU because clinicians deliver standard treatment, such as prescribing a medication, and bolster these with an additional medication or recommendation for psychotherapy if a patient does not achieve a reduction in symptoms (ie, the TAU+X add on model). Although far from perfect as a control condition, TAU is standardly used in pragmatic clinical trials for these and other reasons [77,78].

Although the TAU group showed a significant decrease in anxiety symptoms (14%), there may be several reasons why TAU failed to show a greater reduction in symptomatology or achieve remission (3% vs 64%). These include a higher NNT for current medications and current models in which medical practices are designed more as a sick care model, in which acute, physically based issues are prioritized over mental health despite clear advantages of integrating mental health care into primary care settings [79,80]. This study demonstrated a clear proof-of-concept trial of a mindfulness-based digital therapeutic to deliver specific theory-driven and mechanistically based treatment in a clinically relevant setting. Furthermore, we aimed to recruit a real-world population by minimizing the exclusion criteria, such as comorbid disorders. In this study, most of the individuals (84%) presented with comorbid disorders, such as depression, which is consistent with how individuals present in primary care settings and to treatment specialists.

This study has several strengths and notable findings, including designing for real-world applicability (eg, including individuals with comorbid disorders and concomitant medication treatment), accounting for engagement, assuring adequate sample sizes, registering outcomes, and minimizing the risk of bias. However, this study has some limitations. The TAU+ model was chosen to closely match the treatment a patient would encounter in a clinical setting. Whereas standard clinical care is highly variable, the study was designed such that randomization would ensure that this variability was equally distributed between the groups. Future studies using an active comparator (eg, CBT-based app) to control for attentional effects, longer follow-up periods, and incorporating multiple sites are needed to confirm the efficacy of this program. Second, this study sample comprised 90% (57/63) women. Although future sex-balanced studies are needed to determine the generalizability of these findings, women are twice as likely to develop an anxiety disorder and have a higher lifetime prevalence of GAD (7.1% vs 4.2%) than men [81,82]. Third, this study was designed to evaluate anxiety symptoms at 2 months after treatment initiation. Long-term follow-up studies are necessary to establish the long-term effects. Furthermore, although this study identifies potential psychological mechanisms of app-delivered MT, such as increased mindfulness mediating decreases in worry and anxiety, future studies are needed to explore its neurobiological mechanisms. In addition, studies performed in research laboratory settings (eg, National Institutes of Health stages I and II) may select more motivated individuals in general. Although randomization controls for between-group differences, real-world efficacy (National Institutes of Health stage III) is needed as the next step to determine efficacy in clinical settings.

In summary, for a large portion of the world’s population that is affected by moderate to severe anxiety, targeted and mechanistically based treatments are needed. By combining theory and a new field of treatment delivery (digital therapeutics), we found that app-delivered MT significantly reduced anxiety, and its effects were mediated by increases in psychological nonreactivity and reductions in worry, suggesting a specific targeting of reinforcement learning.