The gold standard treatment for anxiety is cognitive-behavioral therapy (CBT). Central to CBT is the behavioral component called exposure. Exposure in vivo, facing the fear in a real-life situation, has proved effective (Rodebaugh et al., 2004); however, this may be costly, time-consuming, and could involve situational elements that are difficult to control (Bouchard, 2007; Bouchard et al., 2014). Few individuals with anxiety-related disorders seek treatment or only do so after a long delay (Wang et al., 2005; Stein and Stein, 2008; Bandelow and Michaelis, 2015; Kampmann et al., 2016). Additionally, whereas CBT has a success rate of approximately 70% in research trials (van Dis et al., 2019), the actual success rate is lower because exposure is not used in 70%–80% of cases treated in routine care settings, due to barriers such as practicability and logistical challenges, negative beliefs about exposure and therapist distress (Pittig et al., 2019). To overcome these difficulties, as described thoroughly below, researchers and clinicians have combined exposure with virtual reality (VR) (Bouchard, 2007; Bouchard et al., 2014). Using biomarkers with VR-based exposure may enable clinicians to assess anxiety levels objectively and regulate interventions in real-time (Lindner et al., 2017). This review investigates using biomarkers in VR-based exposure to treat anxiety disorders.

VR is defined as the use of computer and behavioral interfaces to simulate the behavior of three-dimensional (3D) entities that interact in real-time with each other and with a user immersed via sensorimotor channels (Fuch et al., 2011). VR content is generally created using 3D computer graphics or 360° videos (Gregg and Tarrier, 2007; Flores-Arredondo and Assad-Kottner, 2015; Stupar-Rutenfrans et al., 2017). Some studies refer to setups displayed on a monitor as VR (Costa et al., 2018). However, such simulations are not included in this review, as they do not include an immersive experience.

VR has been used for clinical assessments (Emmelkamp et al., 2020) and to treat anxiety disorders. Studies have shown that VR may be effective in treating various phobias (e.g., agoraphobia and fear of flying) (Gregg and Tarrier, 2007) and anxiety disorders (Parsons and Rizzo, 2008; Powers and Emmelkamp, 2008; Opris et al., 2012; Morina et al., 2015; Kampmann et al., 2016; Chesham et al., 2018; Carl et al., 2019). A study by Garcia-Palacios et al. (2007) on patients with various anxiety disorders (including specific phobias, claustrophobia, and social anxiety disorder [SAD]) showed that 76% of participants preferred to receive exposure via VR, rather than in vivo (Garcia-Palacios et al., 2007). In addition, exposure via VR has several advantages compared with traditional exposure methods. For example, although VR-based exposure scenarios can be felt as very realistic (Gutierrez-Maldonado et al., 2015), situational factors can be controlled and regulated (Levin et al., 2015). Furthermore, VR-based exposure can occur in the privacy and safety of the psychotherapist’s office (Gutierrez-Maldonado et al., 2015). Additionally, although they know the VR experiences are not real, patients’ minds and bodies respond to the VR experiences as they would to authentic experiences (Freeman et al., 2017). Hence, patients can face challenging situations more easily in VR and may find it easier to adapt their behavior. Furthermore, VR sessions are less expensive, less time-consuming, and can be planned more flexibly than in vivo sessions (Bouchard et al., 2017). However, VR has disadvantages; these include determining costs, hardware dependency, and VR-induced side effects such as cybersickness (Martirosov and Kopecek, 2017).

Some researchers have sought to add biomarkers to augment VR-based exposure treatments. Using biomarkers with VR-based exposure may enable researchers to assess anxiety levels objectively, regulate interventions in real-time, and collect valuable data (Lindner et al., 2017).

Anxiety disorders are associated with heightened physiological arousal (Bond et al., 1974; Hoehn-Saric and McLeod, 1988), expressed as elevated muscle tension (Barlow and Ellard, 2018), respiration rate (Wilhelm et al., 2001), skin conductance level (SCL) (Globisch et al., 1999), and heart rate (HR) (Cuthbert et al., 2003), as well as decreased heart rate variability (HRV) (Bandelow et al., 2017). These physiological measures are similar to preferred stress measurements, as they are indicators of the larger response of the Hypothalmic Pituitary Adrenal (HPA) axis and Autonomic Nervous System (ANS) (Dammen et al., 2022) Such biological measures can be referred to as biomarkers. A biomarker is a measurable indicator of a biological state or condition. In a psychiatric context, biomarkers can be used to detect, assess, or predict the development of psychiatric disorders, personality traits, or behavioral characteristics and to evaluate emotional or cognitive capacities (Perez et al., 2014). Furthermore, biomarkers may be used to inform treatment decisions (Perez et al., 2014). Biomarkers that are analyzed at different time points to monitor the status of a disease or medical condition and to provide feedback on the response to an intervention are called monitoring biomarkers (Group, 2016; Garcia-Gutierrez et al., 2020). Behavioral markers may also be considered biomarkers. Examples of behavioral markers include information from trackers attached to the feet, information provided by motion capture technology, body movements that are associated with depression (Alessi and Huang, 2001), changes in tone of voice (Place et al., 2017), and eye-tracking data (Shic, 2016). However, these behavioral markers are not considered biomarkers in this review because they are not direct measures of a biological state or condition.

Biomarkers may also be used to provide biofeedback. Biofeedback is a noninvasive procedure with the operating principle of conscious registration of otherwise unconscious bodily procedures (Gaume et al., 2016). The patient’s physiological outputs (i.e., biomarkers) are detected, monitored, and processed electronically and then returned as feedback in different forms (e.g., visual and auditory) to the same individual (Gaume et al., 2016; Fernández-Álvarez et al., 2020). Thus, positive physiological outcomes (e.g., reduced physiological arousal) may be generated by constant positive feedback (Frank et al., 2010).

Biofeedback has been used for stress reduction (Dillon et al., 2016), facilitating relaxation (Poleszak et al., 2019), and conditioning (Duits et al., 2016), and it has been effective in the treatment of an array of medical and psychological conditions. Anxiety disorders are among the most frequently treated psychiatric conditions (Schoenberg and David, 2014), and HRV biofeedback has proved effective in treating anxiety (Goessl et al., 2017; Fernández-Álvarez et al., 2020). Biofeedback is frequently used for exposure-based therapy. By convention, exposure-based treatment and biofeedback-based interventions take different approaches. Exposure-based treatment focuses on learning principles that range from habituation to inhibitory learning (Craske et al., 2008; Craske et al., 2014) and on achieving extinction of the fear response by facing situations that induce anxiety. Biofeedback-based interventions focus on practicing anxiety regulation skills informed by physiological information concerning the arousal response (Lin et al., 2019). While the term “habituation” often refers to changes in underlying psychological, neurophysiological, and social processes, in this review “habituation” is considered an indicator that fear extinction learning is taking place (Craske et al., 2008).

Although biofeedback has been effective in treating anxiety disorders, some means may be more relevant to VR-based exposure than others, such as those more closely associated with autonomous nervous system activity (Tolin et al., 2020). Clinicians must identify the best methods for combining targeted physiological processes and VR-based exposure (Repetto et al., 2013; Zotev et al., 2013).

Combining biomarkers with VR allows researchers to collect data in a controlled setting, facilitates objective anxiety assessments, and means that interventions can be regulated (Lindner et al., 2017). Fernández-Álvarez (2020) noted that VR enables physiological processes to be represented by virtual stimuli. These stimuli are transmitted using biosensors, strengthening user engagement and potentially augmenting the effectiveness of the interventions (Fernández-Álvarez et al., 2020). A range of studies has demonstrated the effectiveness of this approach in healthy and clinical populations (Repetto et al., 2013; Zotev et al., 2013). For example, in a study by Repetto et al. (2013), patients with generalized anxiety disorder (GAD) were treated with VR relaxation simulations in which the environment (a campfire, beach, waterfall, or gazebo) was adapted in real-time by therapists based on HR biofeedback: a reduction in HR reduced the intensity of the stressful stimuli (e.g., fire intensity, wave movement, or stressful image/video).

Combining real-time biofeedback and real-time adaptable VR may facilitate learning relaxation techniques (Repetto et al., 2013; Bossenbroek et al., 2020). For example, in the VR biofeedback game DEEP, players wear a stretch sensor that measures the expansion of the diaphragm associated with breathing: the slower and deeper players breathe, the better they can move around in a VR underwater world (Bossenbroek et al., 2020).

A recent systematic review of randomized controlled trials (RCTs) investigating biofeedback for anxiety disorders found 21 articles eligible of inclusion (Tolin et al., 2020). However, only one of these articles combined VR with biofeedback (Repetto et al., 2013). Few studies have investigated using biofeedback to regulate interventions for anxiety-related disorders (Repetto et al., 2013; Lin et al., 2019). However, more studies have investigated using biomarkers with VR for other purposes, such as assessing treatment effects (Peperkorn et al., 2014; Gujjar et al., 2018; Kaussner et al., 2020).

Results from studies using biomarkers have frequently been compared with results from questionnaires to asses symptom reduction (Tolin et al., 2020). The terms synchrony and desynchrony (Hodgson and Rachman, 1974; Rachman and Hodgson, 1974) refer to the consistency in changes over time between physiological and subjective measures. Findings indicate that questionnaires and biomarkers are not always in synchrony (Craske and Craig, 1984). This inconsistency may be due to psychological factors, such as expectations and attention to bodily sensations, leading to distortions in perceived anxiety (Hoehn-Saric and McLeod, 2000). Furthermore, biomarkers such as HR and respiratory rate often peak and return to their resting state faster than subjective self-report measures (Lande, 1982). Interestingly, increased arousal is associated with greater consistency between questionnaire and biomarker data (Gerew et al., 1989), and greater consistency may also be observed for phobic anxiety (Schaefer et al., 2014).

Combining VR and biomarkers has intriguing and promising potential, allowing for clinical advantages during exposure, such as real-time monitoring and adaption. However, no systematic review has investigated using biomarkers with VR-based exposure to treat anxiety. Here, we investigate synchrony and desynchrony between biomarkers and self-report questionnaires in completed trials.

Journal articles were identified using searches in the scientific databases PubMed PMC, PsychInfo (Ovid), and the Cochrane Library. Potential cross-references were sought; however, this yielded no additional studies for screening. Search criteria included all publications from the Diagnostic and Statistical Manual of Mental Disorders-IV released in 1994 until 2 October 2023. The search strategy was the same for each database and involved searching for the following keywords in the title and abstract: “virtual reality” AND (“galvanic skin response” OR “heart rate” OR “biofeedback” OR “biometric*" OR “biomarker*" OR “EEG” OR “eye tracking” OR “physiological”) AND “exposure” AND (“anxiety” OR "*phobia” OR “posttraumatic stress disorder” OR “ptsd” OR “obsessive compulsive disorder” OR “ocd”) AND “adult".

Studies published in peer-reviewed journals were included if the following criteria were met: 1) written in English, 2) only included participants aged 18 years or older, 3) included participants diagnosed with an anxiety-related disorder, based on DSM-IV (American Psychiatric Association, 1994) or DSM-V criteria (American Psychiatric Association, 2013), 4) included exposure to anxiety by using VR in at least one arm of the study, 5) included therapy with the intention of reducing symptoms of anxiety, 6) included at least one biomarker, and 7) included immersive VR equipment (e.g., HMDs) with immersive properties. Meta-analyses, reviews, single-case studies, and study protocols were excluded.

The initial search of titles and abstracts yielded 2,292 results in Pubmed PMC, 39 results in PsychInfo (OVID), and 78 results in the Cochrane Library, totaling 2,409 studies. Next, 12 duplicates were identified and removed, leaving 2,397 abstracts for screening. Reference manager (Thomson Reuters, Toronto, ON, Canada), Mendeley (London, UK), and Covidence systematic review software (Veritas Health Innovation, Melbourne, Australia) were all used to identify duplicates and review the articles. Titles and abstracts were independently reviewed by two authors (MTE and ER) and coded as to whether they met the inclusion criteria. Disagreements were resolved by discussion until a consensus was reached or were decided by the senior author. After the initial screening of abstracts, 53 studies remained. Full-text articles were obtained for these abstracts, and all were independently reviewed by two authors (MTE and ER). Of the 53 articles selected for full-text review, 27 articles (Gujjar et al., 2019; Shiban et al., 2017; Shiban et al., 2015; Diemer et al., 2013; Malbos E et al., 2013; Donahue et al., 2009; Norrholm et al., 2016; Levy et al., 2016; de Quervain et al., 2011; Cote and Bouchard, 2005; Wiederhold et al., 2002; Tardif et al., 2019; Hur et al., 2021; Kaussner et al., 2020; Loucks et al., 2019; Mühlberger et al., 2001; Borghese et al., 2022; Diemer et al., 2023; Gramlich et al., 2021; Hong et al., 2017; Premkumar et al., 2021; van ’t Wout-Frank et al., 2019; Wechsler et al., 2021; McGlade et al., 2023; McInerney et al., 2021; Price et al., 2015; Reitmaier et al., 2022), representing 1046 participants, were retained for final inclusion (Figure 1).

Two assessment tools were utilized to evaluate Risk of Bias (RoB). For RCTs, the Cochrane tool for RoB in RCTs (Higgins et al., 2011), which categorizes RoB into three groups (High, Low, and Unclear), was used. For cohort studies, the Newcastle–Ottawa Assessment Scale (Stang, 2010), which categorizes each study into one of three groups (Good quality, Fair quality, or Poor quality), was used. The RoB assessment was carried out by three authors (MTE, JHN, and ER), and any disagreement was resolved by the senior author (LC). The scoring groups for the cohort studies and the RCTs were combined in an overall RoB assessment. RoB for each study is shown in Figure 2 and Table 1.

The following data were retrieved from the included studies: authors, publication year, number of participants, gender distribution, age, anxiety-related diagnosis, study design, biomarkers, time of biomarker measurement, anxiety-related questionnaires, and function of biomarkers. In this review, “over time” includes data collected for a single session as well as data collected at several time points within a single session.

Data on identification of group differences and changes within and between sessions were retrieved for both biomarkers and questionnaires. Group differences were defined as differences in score at a particular point in time, if these data were available, and reported as significant or non-significant. Changes within or between sessions were categorized as significant increases, significant decreases, or not significant.

Furthermore, synchrony between biomarkers and questionnaires was assessed. Findings were considered synchronous if both biomarkers and questionnaires showed a significant increase, a significant decrease, or no significant change. The findings were considered synchronous if a particular biomarker and at least one questionnaire exhibited similar results. When summarizing the synchrony over several studies, “levels of synchrony” is descriptively used as instances of positive biomarker detection as a percentage of the total number of instances. The authors agreed upon the following definition of synchrony levels: Low levels of synchrony were categorized as <30%, moderate levels of synchrony as 30%–80%, and high levels of synchrony as >80%.

Data were extracted independently by three authors (MTE, JHN, and ER) using a predefined information sheet. Any discrepancies were resolved by discussion and consultation with the senior author until a consensus was reached.

The 27 studies included a total of 1046 participants with a mean sample size of 38.7 (range, 6–116). Differences in sample sizes across studies were not considered; thus, each study was given equal weight. The majority of participants were female (59.1%), ranging from 0% to 96% (based on 26 studies), and the (weighted) mean age was 32.9 years (range of mean age, 20.6–44.5 years).

Six studies were on social anxiety disorder (SAD), five on post-traumatic stress disorder (PTSD), five on arachnophobia, five on acrophobia, three on aviophobia, one on dental phobia, one on agoraphobia, and one study was on driving phobia. Twenty-three of the studies were RCTs, and five were cohort studies. Sixteen studies had follow-up assessments, and the follow-up periods ranged from 0.5 to 12 months (Table 2).

Table 3 shows that all 27 studies used biomarkers as outcome measures. Thirteen of the studies (Gujjar et al., 2019; Shiban et al., 2017; Shiban et al., 2015; Diemer et al., 2013; Malbos E et al., 2013; Donahue et al., 2009; Kaussner et al., 2020; McGlade et al., 2023; Borghese et al., 2022; Reitmaier et al., 2022; Gramlich et al., 2021; Premkumar et al., 2021; van ’t Wout-Frank et al., 2019) also used biomarkers as measures of inter-session or intra-session habituation. These thirteen studies measured physiological responses at the beginning and/or during the exposure and/or at the end of each exposure session, or they measured changes in physiological responses between specific scenarios during the exposure. The changes were compared with subjective measurements to determine whether anxiety levels had decreased and habituation had occurred. Only two studies provided participants with visual feedback, based on physiological responses, during the exposure in VR (Wiederhold et al., 2002; Hong et al., 2017). None of the studies used biomarkers to automatically and/or dynamically adapt VR-based exposure via computer algorithms.

Biomarkers were applied in a total of 94 instances across all the studies (Table 4). These included 40 instances of HR, 33 instances of SCL, four instances of HRV, four instances of startle response, three instances of systolic blood pressure (SBP), three instances of diastolic blood pressure (DBP), three instances of peripheral skin temperature (SKT), two instances of functional magnetic resonance imaging (fMRI), and two instances of cortisol levels. In this review, the term “instance” refers to an occasion on which a measurement was used; for example, in the study by Donahue et al. (2009), HR data were collected for group differences, as well as within sessions and between sessions, totaling three instances of HR use.

Table 4 shows that HR was the most frequently used biomarker, and HR was reported in 20 of the 27 studies. The next most frequently used biomarker was SCL, which was reported in 15 studies. Only two studies did not use HR or SCL. Eight studies used other biomarkers. Two studies used HRV, two used the startle response, one used SBP and DBP, one used fMRI, one used peripheral skin temperature, and one used cortisol levels. Several studies measured more than one biomarker, and there were a total of 94 instances of biomarker investigation across the 27 studies (Table 4).

Biomarkers identified 15 significant differences among 35 instances that investigated group differences. Biomarkers identified 15 changes within sessions from a total of 24 instances investigated and 24 instances of changes between sessions from a total of 35 instances investigated. Biomarkers were used in 21 studies to investigate changes between sessions. Of these, 15 studies reported significant differences for at least one biomarker.

Regarding specific biomarkers, HR identified group differences in 5 of 13 instances investigated (38%), and SCL identified group differences in 3 of 13 instances investigated (23%). Moreover, HR identified changes within sessions in 9 of 12 instances investigated (75%), and SCL identified changes within sessions in 5 of 9 instances investigated (56%). Finally, HR identified changes between sessions in 9 of 15 instances investigated (60%), and SCL identified changes between sessions in 6 of 11 instances investigated (55%; Table 5).

Table 5 shows that the findings for biomarkers were synchronous with the results from questionnaires in 20 of 35 instances for group differences (57%), 11 of 24 instances for changes within sessions (46%), and 18 of 35 instances for changes between sessions (51%). For group differences, the findings for HR were synchronous with the results from questionnaires in 6 of 13 instances (46%), and the findings for SCL were synchronous with the results from questionnaires in 11 of 13 instances (85%). For changes within sessions, the findings for HR were synchronous with the results from questionnaires in 6 of 12 instances (50%), and the findings for SCL were synchronous with the results from questionnaires in 4 of 9 instances (44%). For changes between sessions, the findings for HR were synchronous with the results from questionnaires in 7 of 15 instances (47%), and the findings for SCL were synchronous with the results from questionnaires in 4 of 11 instances (36%).

This review has several strengths. The literature search was exhaustive; therefore, our conclusions are based on a comprehensive assessment of the work published within this research field. In addition, the majority of the studies that were included were RCTs. The rating process was thorough, with several researchers providing independent ratings before reaching a consensus. The studies were homogeneous regarding VR content (render-based content) and publication year, with all studies being published within the last 23 years.

However, the review also has limitations. While studies were homogenous in terms of VR content, the quality of included content may vary significantly between studies due to content development design choices and resources, and technology development throughout the years. Indeed, HMD and computing capabilities have increased since the earliest included studies, however, even the earliest included studies were able to detect differences in biomarkers and subjective measurements. Changes in biomarker levels were not the primary outcomes of most of these studies, and the findings described in this review were not always reported in a suitable format by the authors of the studies that were included, resulting in variations in the quality and quantity of data that could be used to address our secondary aims. This is particularly true for the findings for biomarkers that were used to assess changes within sessions, which are based on a total of only 24 of 94 instances. Furthermore, because there is no consensus on standard practice for collecting SCL data, the included studies applied different approaches, which may have influenced our analyses. Additionally, few of the included studies focused on comparing the findings for biomarkers with the results from questionnaires, and many of the studies did not report concurrent measures for biomarkers and questionnaires. Except for HR and SCL, few biomarkers were comprehensively assessed; therefore, conclusions regarding other biomarkers are relatively weak. We did not differentiate between experimental tasks, which may vary widely based on diagnosis and scope. Tasks including a higher degree of movement may have affected biomarker outcomes, as these are commonly sensitive to movement.

Using the results from questionnaires to validate findings for biomarkers has limitations. As described above, previous studies (Craske and Craig, 1984) have reported low levels of synchrony between the findings for biomarkers and the results from questionnaires. Furthermore, using more than one questionnaire may generate conflicting results, with only some results agreeing with the findings for biomarkers. Although all questionnaires assessed anxiety levels, some focused on the subjective symptoms of anxiety rather than the physiological symptoms of anxiety that are most closely linked to biomarkers.

A majority of the participants were female (59.1%). However, this observation is consistent with the lifetime prevalence of anxiety among the general population, which is also higher among females (33%) than among males (22%) (McLean et al., 2011). The mean age range was narrow (approximately 21–44 years), perhaps reflecting the recruitment strategies used. Thirteen studies reported their recruitment procedures (Wiederhold et al., 2002; Cote and Bouchard, 2005; Donahue et al., 2009; de Quervain et al., 2011; Diemer et al., 2013; Shiban et al., 2015; Hong et al., 2017; Shiban et al., 2017; McInerney et al., 2021; Premkumar et al., 2021; Wechsler et al., 2021; Borghese et al., 2022; Diemer et al., 2023). These studies mainly used public advertisements for recruitment, and seven of the 13 used advertisements at universities, schools, and colleges, which might explain the narrow mean age range. Younger adults may exhibit more technological literacy than older adults, which may result in younger adults being more likely to participate in relevant research, including research involving VR. Therefore, we may find it more difficult to obtain data on older adults’ psychological reactions to VR-based exposure. One non-VR study found that older adults reacted more calmly than younger adults when anxiety was induced unless a physical threat was involved (Teachman and Gordon, 2009). However, whether older adults react more calmly than younger adults to VR-based exposure is not known.

Finally, our risk of bias assessment showed a “high risk” tendency in one-third of the studies, mainly due to missing outcome data. This could compromise our findings because it is unclear whether the missing data are linked to particular anxiety levels or perhaps to publication bias. Thus, it is uncertain whether particular findings are valid or are influenced by the quality of studies or other confounding variables.

Based on the findings of this review, we cannot provide strong recommendations regarding the use of monitoring biomarkers in clinical practice. However, we can present recommendations for further research based on theoretical and logical considerations and implications.

In this review, we did not differentiate among diagnoses. Therefore, our findings provide valuable information across diagnoses. However, there may be diagnostic differences in the physiological reactions to exposure, or there may be particular diagnoses for which the physiological reactions are particularly sensitive to treatment. Variations in the response of particular biomarkers may be associated with particular anxiety disorders. For example, individuals with phobic anxiety exhibit an increased HR in response to phobic imagery compared with control imagery; this response is greatest for simple phobias, intermediate for SAD, and smallest for agoraphobia (Cook et al., 1988; Hoehn-Saric and McLeod, 2000). Furthermore, baseline differences in HRV have been reported for panic disorder, GAD, and SAD (Pittig et al., 2013). Finally, whereas individuals with anxiety show decreased HRV, no significant decrease in HRV is observed among individuals with obsessive-compulsive disorder (Chalmers et al., 2014). Thus, important physiological differences exist among diagnoses. In the future, VR will be used more widely, and biometric data will be available for more diagnoses. Therefore, understanding the data associated with different diagnoses may enable clinicians to provide the most appropriate treatments.

Frequently used biomarkers, such as HR and SCL, may provide valuable information concerning anxiety levels experienced by patients. However, they do not provide the critical information necessary to decide how to proceed during exposure. Indeed, arousal plays a small role in assessing the optimal rate of progression for exposure. Arousal is essentially an indicator that emotional processing and inhibitory learning are occurring, but it is a poor predictor of outcomes (Craske et al., 2014). Therefore, clinicians should consider collecting data from other biomarkers, such as electroencephalography (EEG) and body temperature, as well as subtle behavioral markers such as voice tone or pitch, facial muscle contractions, and behavioral avoidance patterns. Information from several biomarkers may be needed to draw meaningful conclusions. Whereas individual biomarkers may show low levels of synchrony with questionnaires, combining several biomarkers may increase levels of synchrony. Biomarkers have different characteristics. For example, SCL is particularly sensitive to novel inputs and is associated with the sympathetic nervous system, whereas HR may be used as a biomarker for diverse diagnoses (Craske et al., 2014). Combining several biomarkers may provide a better indication of the anxiety levels actually experienced by sampling diverse patterns of sympathetic nervous system activity. Furthermore, different individuals may react differently across modalities: some may react primarily by exhibiting changes in SCL, whereas others may exhibit changes in cardiovascular activity. Additionally, some individuals may exhibit low SCL when faced with severe anxiety. Therefore, using several modalities can avoid the pitfall of relying on unsuitable markers. Future studies should identify information that most accurately reflects anxiety states and changes and how to use this information most effectively.

Our review suggests that procedures that involve real-time monitoring of biomarkers are associated with low validity in VR-based exposure therapy. However, real-time application of biomarkers during therapy may provide clinicians with insight and inform decision-making, or it may be used as a tool for post-exposure functional analyses. Indeed, one novel approach has involved using biomarkers during therapy sessions to inform therapists and provide a basis for developing artificial intelligence systems that can automate treatment and replace human therapists (Rizzo and Bouchard, 2019; Šalkevicius et al., 2019). This research has the potential to improve therapy by providing detailed guidance for therapists, as well as by enhancing patient interoception and self-regulation. Information may be provided by personalized virtual or conversational agents. Personalized conversational agents (i.e., a virtual character that patients can modify and engage with in verbal conversations) can increase motivation and commitment to psychotherapy (Briskilal.J et al., 2020). These agents may be adjusted by real-time biomarker data and artificial intelligence technology to improve the effectiveness of guidance and therapy, providing valuable information to patients and clinicians. Moreover, such agents may eventually be implemented in everyday practice to bridge the gap between innovative research and the practitioner’s office.