Craving, defined as the urge to consume a psychoactive substance, is a criterion for diagnosing alcohol dependence (AD) and alcohol use disorder according to the International Classification of Diseases, 11th Revision (ICD-11) and the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) [1,2] and a predictor of relapses [3]. One explanatory model, the “conditioned drug-like model,” highlights the role of substance-associated stimuli that trigger psychological and physiological responses, promoting drug-seeking behavior [4]. In cue exposure therapy (CET), a component of cognitive behavioral therapy, patients are exposed to alcohol-associated stimuli to identify risk situations and practice coping strategies, potentially reducing conditioned responses over time [5]. The effectiveness of CET for AD remains debated because of insufficient evidence, with a small number of studies and small-to-moderate effects on outcomes, underscoring the need for more trials using innovative methods such as virtual reality (VR) [6]. VR-CET has gained attention because of its advantages: realistic and immersive simulations, individualized content, increased practicability, and better standardization [7,8]. These benefits could reduce the financial and time demands of traditional in vivo CET in clinical practice [8]. Accurate craving assessment is critical for developing effective VR-CET scenarios.

Based on self-report data, Ghiţă et al [9] identified common triggers for alcohol craving, such as being in social settings or specific environments such as bars, restaurants, or home, and found beer, red wine, and whiskey to be most commonly associated with craving. A recent study based on this research, which implemented these VR scenarios, showed no significant differences in relapse rates between the VR-CET and control groups but did show a tendency toward lower subjective craving and anxiety in the VR-CET group. Furthermore, several important limitations of the study were highlighted, underscoring the need for further research [10]. Beyond self-reports, studies have shown that physiological responses such as pupillometry, respiration rate (RR), heart rate (HR), and electrodermal activity (EDA) can also indicate craving [11-14]. These methods help address the limitations of self-reports, including social desirability bias, defined as the tendency to give socially desirable responses [15], and recall difficulties when using retrospective reports [16]. In line with this, correlations between subjective craving and physiological responses are rather weak [17,18], but studies are few and the relationship remains ambiguous, meriting further investigation. VR may help clarify this relationship and identify physiological biomarkers relevant for diagnostics and therapy, including biofeedback applications in VR-CET in the long term.

Several studies have demonstrated that VR exposure paradigms can effectively induce craving [19]. However, research on physiological cue reactivity in VR scenarios is limited. One study included electroencephalography as a potential correlate of craving during VR-CET and showed a decrease in craving for alcohol parallel to an increase in frontal α activity over the course of the therapy [20]. One recent study by Zhang et al [12] focused on the changes in HR, EDA, and RR after VR-CET and found significantly lower increases in HR and subjective craving after cue exposure in the group receiving VR-CET. Apart from these studies, to our knowledge, further research on craving induction through virtual reality cue exposure (VR-CE) in AD populations has relied on subjective craving measurements [19].

In this study, we first investigated the induction and progression of craving through VR exposure using both subjective (questionnaires) and objective physiological measures (EDA, heart rate [variability] [HR(V)], brightness-corrected pupil diameter [BCPD], RR). We hypothesized that exposure to alcohol-associated VR scenarios would significantly increase (1) subjective craving and (2) physiological indicators of craving.

Additionally, we aimed to explore the relationship between subjective and objective measures, hypothesizing that (3) subjective and physiological measures of craving would show significant correlations. On an exploratory basis, we also examined the correlations between the individual psychophysiological parameters.

This prospective, exploratory, single-arm clinical study assessed subjective and physiological craving parameters in patients with AD. It followed a previously published study protocol with details on the study design, measures, and materials [21].

The study was approved by the Charité–Universitätsmedizin Berlin Institutional Review Board (EA1/190/22; May 23, 2023) and was preregistered on ClinicalTrials.gov (NCT05861843). All participants provided written informed consent and received compensation of €30 (approximately US $34.80) for participation.

All study data were pseudonymized, and identification through images or supplementary material was not possible.

Inpatients and outpatients diagnosed with AD were recruited at Charité–Universitätsmedizin Berlin and Salus Clinic Lindow, Germany. Patients were screened for eligibility, and they provided written informed consent and received compensation. Recruitment and data collection took place between September 23, 2023, and June 30, 2024.

Patients aged between 18 and 65 years who were diagnosed with AD according to ICD-10 (F10.2) based on clinical consensus were included in the study. Further inclusion criteria were completion of inpatient withdrawal treatment during the last 3 months and a history of alcohol craving.

Exclusion criteria included other substance dependencies (except nicotine), less than 7 days of abstinence or current intoxication (randomly tested using breath alcohol concentration), severe neuropsychiatric disorders (eg, schizophrenia and bipolar disorder), cognitive impairment, acute suicidality or risk of endangering others, serious conditions affecting brain or heart function, current use of AD-targeting medication (eg, benzodiazepines and acamprosate), medications significantly affecting HR, or inability to understand study procedures.

All analyses were performed using R software (version 4.3.2, Eye Holes; R Foundation for Statistical Computing) [37] and RStudio (version 2024.4.0.735, Chocolate Cosmo; Posit) [38]. Demographic and baseline characteristics are reported as mean, SDs, and ranges.

EDA, HR, BCPD, and RR data were divided into 5 equal segments for each 5-minute exposure period (baseline: B1, B2, B3, B4, and B5; risk scenario: R1, R2, R3, R4, and R5). HRV parameters were calculated across each entire VR scenario (5 min; baseline: B, risk scenario: R). Each exposure combined a baseline and a risk scenario into 1 data frame. Because each participant underwent 2 exposures, each participant produced 2 data frames (N=61 participants, N=122 data frames). Because of recording problems, some data were unavailable: one participant lacked all physiological recordings, resulting in 2 missing data frames for BCPD, RR, EDA parameters, HR, and HRV. In addition, 4 BCPD data frames from 3 participants were missing, leaving 116 data frames for BCPD analysis. For HRV, 1 additional data frame with partial missing values (segments R4 and R5) was excluded, resulting in 119 data frames for HRV analysis. NS-SCR frequency was calculated from 116 data frames, with 2 participants excluded as nonresponders.

Subjective craving, measured using VAS, was assessed during first (B1 and R1), third (B3 and R3), and fifth (B5 and R5) segments, with one missing data frame, resulting in 121 data frames.

The main analyses were conducted using linear mixed-effects models fitted with the nlme package (version 0.3.2) [39] to estimate the effects of VR-CE. Models were specified with a categorical variable for craving assessment as a fixed effect, indicating exposure to baseline (B1-B5) and risk scenarios (R1-R5). A random intercept was included to account for individual variability, whereas a fixed slope was used to estimate the overall effect across all participants. The mean baseline value (B) was included as the reference category. Accordingly, the estimated fixed effects (β) from the linear mixed-effects models are reported as model-based mean differences between each craving assessment and the baseline assessment, as estimated by the nlme package. Models are reported in adjusted form, using covariates with influence on the psychophysiological parameters identified a priori. These covariates were selected individually based on guidelines and prior research and were centered within cluster. Covariates for EDA included gender, age, ambient temperature, and medication [27]. In addition to these covariates, BMI, smoking status, and health status (Charlson Comorbidity Index score) were incorporated for HR, HRV [40], and RR [41-45]. Medication and age were also included for BCPD [46-51]. Ambient temperature for 1 participant was interpolated because of a missing value. Post hoc calculations of t statistics and P values were performed using the Satterthwaite approximation [52]. Missing data were handled using linear mixed-effects models, which incorporate all available data from each participant into the analysis. Effect sizes were estimated as standardized mean differences, that is, approximations of Cohen d based on mixed-effects modeling following Pustejovsky et al [53], using the lmeInfo (version 0.3.2) [54] and nlme (version 0.3.2) [39] packages. Plots of the estimated model effects were generated using the effects package (version 4.2.4) [55]. Correlations between subjective and physiological craving parameters were analyzed using Spearman correlations on calculated difference scores of the parameters between baseline and risk scenarios (first segment: R1-B1; second segment: R2-B2; third segment: R3-B3; fourth segment: R4-B4; and fifth segment: R5-B5). Correlation analyses could only be conducted using complete data frames, leading to 111 data frames for analysis. To correlate subjective and physiological parameters at the same time points, physiological measures during the first, third, and fifth segments of the exposure were selected, during which subjective craving was assessed. Additional correlation analysis was conducted for psychophysiological craving parameters (EDA, HR, RR, and BCPD), excluding HRV because of the longer time frames for analyses. Effect sizes are reported as Cohen d, with 0.2 considered as small, 0.5 as moderate, and 0.8 as large [56]. The predetermined α level was .05. This study was reported in accordance with the American Psychological Association Reporting Standards for quantitative research [57].

Baseline demographic and medical characteristics are presented in Table 1. Participants (49/61, 80% male and 12/61, 20% female) had a mean age of 48.16 (SD 9.82; range 31‐65) years, with a mean duration of 20.64 (SD 22.42; range 7‐109) days since their last alcohol consumption on the day of the experiment.

Mean craving was 9.03 (SD 17.96) during the VR-CE baseline and 22.77 (SD 26.49) during the risk scenario, as assessed using a VAS ranging from 0 to 100 (Figure 1). Cybersickness, including any sign of discomfort, nausea, and dizziness, during the VR-CE was reported by 67.2% (41/61) of patients on the Fast Motion Sickness Scale.

Figure 1 shows subjective craving measured by VAS during VR-CE baseline and risk scenarios, displayed using a combination of boxplots and spaghetti plots. The boxplots display the mean (dotted line), median (horizontal line), IQR (box), whiskers (1.5 × IQR), and outliers marked as crosses.

This study demonstrates that exposure to alcohol-associated VR scenarios significantly induced subjective craving. VR-CE had significant effects on NS-SCR as one measure of EDA, on BCPD, and on RR. No significant changes were observed in SCL, as a second EDA parameter, HR, or HRV. Our first hypothesis was thus confirmed and our second hypothesis partially confirmed. Correlation analyses focusing on simultaneous time segments revealed significant but weak correlations between subjective craving and EDA measures. Subjective craving was not significantly correlated with other physiological parameters. Additionally, significant correlations were found in simultaneous assessments between several physiological craving parameters, leading to a partial confirmation of our third hypothesis.

This study confirms that VR effectively induces subjective craving, consistent with positive results from previous studies [9,58,59]. A recent scoping review from our group once again confirmed that the majority of studies report significant changes in craving [60]. The current study builds on a prior feasibility study [61], adding comparisons to baseline scenarios and physiological measures. Observed effect sizes (d=0.56‐0.72) indicate moderate effects, higher than those reported in the meta-analysis by Carter and Tiffany [62] (d=0.53). This may be due to the higher ecological validity of VR compared with more conventional stimuli (eg, images, sounds, or in vivo cues) used in prior studies [62].

The results of our study indicate that measures of EDA (NS-SCR), BCPD, and respiration could serve as sensitive biomarkers of craving in patients with AD exposed to high-risk scenarios in VR. These measures provide objective insight into a subjective, bias-prone construct and may support more effective VR-CET, including biofeedback integration. Providing patients with feedback on their physiological data may help them better recognize and manage personal risk cues [63].

To date, we are aware of only 2 other VR craving studies that include psychophysiological measures of craving in patients with AD. However, their designs differed from ours: Lee et al [20] reported an increase in electroencephalography α power and a greater craving reduction through 10 sessions of VR-CET compared with patients receiving cognitive therapy. Zhang et al [12] compared 2 groups (n=23 vs n=21) of male patients with AD, one group receiving treatment as usual and the other group receiving additional VR-CET sessions. Comparing these 2 groups regarding their cue reactivity in VR assessments, the VR group exhibited significantly lower changes in subjective craving and HR, but not in EDA and RR, after treatment [12]. Notably, in the study by Zhang et al [12], skin conductance was averaged across the whole signal and not evaluated separately (eg, SCL, SCR, or NS-SCR), which could artificially elevate the parameter [12,26,27]. Although we employed a different analysis approach, we also found no significant effects of VR-CE on SCL. In contrast to Zhang et al [12], we did not observe significant changes in HR but did note changes in RR. Additionally, we used HRV and BCPD, expanding the physiological scope. While VR-CE had significant effects on BCPD, this was not the case for HRV parameters. Previous studies have reported inconsistent conclusions on this matter: Wang et al [64] observed higher RMSSD in patients with AD during exposure to alcohol cues compared with neutral cues. High-frequency HRV, a marker of parasympathetic activity, has been shown to react to alcohol-associated cues and is associated with relapse in patients with AD [65,66]. It may be worth considering that VR itself may elevate HR and SCL because of sensory stimulation, potentially obscuring craving-specific responses.

Similar to our study, several studies indicated that subjective craving is only weakly associated with physiological parameters. For instance, a meta-analysis conducted by Field et al [18] on craving in various substance-use disorders found low correlations (r=0.19) between subjective craving and attentional bias. Similarly, Witteman et al [67] reported low-to-moderate correlations between HRV and subjective craving (r=0.33, P=.004). Interestingly, Szegedi et al [68] identified 3 groups of patients with AD: one with both subjective and physiological craving responses, another group with only physiological craving responses, and a third group with no response at all. In contrast, Ooteman et al [17] found only a few patients with or without physiological craving responses who also reported subjective craving, suggesting a need to explore possibilities to help patients become more aware of these sensations. Analyses by Bollen et al [69,70] and Field et al [71] showed that patients with AD who report only low subjective craving and high abstinence motivation tend to avoid alcohol-associated cues when presented. Based on these findings, future studies should include fixation analyses as a measure of attentional bias to better understand the relationship between subjective craving and physiological parameters and why these sometimes differ from one another.

The single-arm design of this study does not allow comparisons between patients with AD and healthy controls, which would have shifted the study’s focus but should be reconsidered in further studies. Regarding the physiological measurements, we did not assess nicotine or caffeine consumption directly before study participation, which could have influenced both the baseline and risk scenario results. Additionally, while gender and age were shown to influence physiological parameters such as EDA, our sample’s gender distribution reflects AD prevalence and was adjusted for in the models. Further studies should include additional nonalcohol cue exposure scenarios that are comparable in complexity to the alcohol cue exposure scenarios because the complexity might, for example, influence BCPD. This will strengthen confidence in the ability of specific physiological parameters to indicate alcohol craving. A further control group exposed to alcohol-associated stimuli outside VR (eg, photographs) could be used to correct for the potential sensory stimulation of VR described above.

We demonstrated that subjective craving can be successfully induced by VR-CE in patients diagnosed with AD. NS-SCR, RR, and BCPD were significantly altered, reflecting a physiological craving response, whereas HR, HRV, and SCL did not react significantly to alcohol-associated cues. Correlation analyses indicated significant but weak correlations between subjective craving and EDA. Based on high methodological standards and a wide range of psychophysiological craving parameters, this study complements the ambiguous research situation regarding the induction of subjective and physiological craving and the relationship between subjective and psychophysiological parameters, offering guidance for future research on craving in VR-CE. In the long term, this will inform the development of effective VR-CET, and through further focus on psychophysiological craving signatures, possibly also biofeedback-based VR-CET in patients with AD, a disorder for which innovative treatment options are urgently needed.