Evaluation of a Virtual Human in Delivering Relaxation Exercises for Wound Healing and Stress Reduction: A Randomised Controlled Trial

Abstract

Objective: Relaxation delivered via audiotapes can reduce stress and improve wound healing. Virtual humans are a promising technology to deliver relaxation, but robust research is needed into their effectiveness. This randomised controlled trial investigated whether relaxation delivered by a virtual human could improve healing and reduce stress after an experimental wound. Methods: A total of 159 healthy adults underwent a tape-stripping wounding procedure and were randomly assigned to relaxation delivered by a virtual human, human audiotape, or a control condition. Skin barrier recovery (SBR) was measured by assessing changes in transepidermal water loss at baseline, post-tape-stripping, and post-intervention. Psychological and physiological variables were measured over the session. Participants’ perceptions of the interventions were assessed. Results: There were no significant differences in SBR between conditions. All conditions experienced significant improvements in the psychological variables, heart rate, and cortisol over time. After controlling for the baseline values, the virtual human and audiotape conditions were significantly more relaxed post-intervention than the control condition (p = 0.005), the audiotape condition had lower post-intervention anxiety than the control condition (p = 0.016), and alpha-amylase was significantly reduced in the virtual human group compared with the audiotape (p = 0.041). The audiotape received the highest satisfaction and engagement ratings, with qualitative results suggesting the appearance and lip-syncing of the virtual human could be improved. Conclusions: Relaxation instructions delivered by a virtual human increased participants’ relaxation levels with similar effects to traditional audiotapes. Furthermore, it reduced physiological stress indices. Further work with other wound types and stressed samples is needed. The voice and interactiveness of the virtual human should be improved to promote greater engagement and satisfaction.

1. Introduction

Despite efforts in the past decade, 30% of the global population still lacks access to essential health services, which are marked by an overwhelming demand, resource shortages, and a deficiency of trained health professionals [1,2,3]. Financial constraints, stigma, health literacy, and the COVID-19 pandemic have further exacerbated the barriers to accessing necessary health services. Since the COVID-19 pandemic, there has been an increased prevalence of poor mental health, including stress, depression, and anxiety [4,5]. As more people are waiting longer for healthcare and mental health services, there is more reliance on self-management through the internet and mobile apps. However, engagement with these digital tools remains low due to poor usability, lack of personalisation, limited interactivity, and reduced social connectedness [6,7,8,9].

One area where increasing engagement and adherence could be beneficial is stress management. Relaxation interventions are widely used to counteract stress by reducing sympathetic activation and increasing parasympathetic activity [10]. By doing so, relaxation aims to reduce stress hormones, such as cortisol and alpha-amylase, which can further affect the immune system and inflammatory processes [11,12,13]. Notably, inflammatory cytokines play a role in wound healing [14], and evidence from several trials shows that relaxation exercises can improve the healing of clinical and experimental wounds when delivered face-to-face or through audiotapes [15]. Given these connections, enhancing engagement and adherence with relaxation interventions could have meaningful benefits not only for stress reduction but also for wound healing.

Relaxation exercises are traditionally delivered in person, but due to resource constraints, audio recordings and videos have become more accessible. These passive formats may not be as engaging or sustain adherence over time. Some research suggests that multisensory experiences enhance relaxation responses. For example, one study investigated the sensory components of a virtual reality (VR) relaxation intervention. Four conditions were compared: auditory and visual input (audiovisual), auditory only, visual only, and no video nor audio. The audiovisual condition resulted in the greatest physiological effects (lower heart rate and blood pressure) and reduced the awareness of real-world distractions [16]. This study highlights that audiovisual components may be beneficial for enhancing relaxation responses in VR.

Virtual humans (VHs) have emerged as a novel approach to delivering psychological interventions due to their multimodal and social capabilities. VHs are embodied conversational agents with realistic and human-like animated appearances and multimodal communicative capabilities. Advanced VHs can respond to users in real time by interpreting the user’s input (facial expressions, speech) to determine their responses (verbal and non-verbal) [17]. This social presence and interactivity have made them promising tools for healthcare applications, including stress and pain management, mental health assessment, medical interviews, and health coaching [18,19,20,21,22,23].

Prior research suggests that human-like qualities, such as facial expressions and body language in robots can enhance engagement and adherence. For example, a study found that participants preferred and adhered more to relaxation instructions given by a physical robot over a tablet [24]. As higher adherence leads to consistent relaxation practice, and therefore, greater health benefits, a VH’s ability to simulate a human guide and provide social presence could improve long-term engagement compared with more passive formats, such as audiotapes.

However, VHs exist on a spectrum of interactivity, ranging from fully autonomous agents to scripted and pre-recorded interactions [18]. While advanced VHs offer the greatest potential for therapeutic use, technical limitations and inconsistent responses could hinder their effectiveness. To establish a foundation for future research, this study focused on the effects of a pre-recorded VH that delivered relaxation exercises against an audio modality. By only changing one variable, this study aimed to isolate the effects of the visual human-like presence.

Therefore, the current study investigated whether a pre-recorded relaxation intervention delivered by a VH could improve the speed of wound healing compared with traditional audiotaped relaxation and a control condition (quiet reading). To explore potential mechanisms underlying this effect, this study examined whether the VH relaxation video could reduce self-reported and physiological stress and induce self-reported relaxation. An experimental acute wound model (tape-stripping) with a healthy population was used to isolate intervention effects and reduce confounding factors. This laboratory model for inducing a non-invasive minor wound was previously used for research on psychology and wound healing [15]. Satisfaction, engagement, and open-ended feedback about the virtual human were also assessed to inform future virtual human design.

It was hypothesised that the virtual human and audiotape conditions would have faster wound healing, as indicated by higher skin barrier recovery (SBR) rates. It was hypothesised that the secondary outcomes (relaxation, psychological stress, heart rate, electrodermal activity, salivary cortisol, and alpha-amylase) would improve in the two relaxation conditions compared with the control condition. The satisfaction and engagement scores were also hypothesised to be higher in the audiotape and virtual human conditions than in the control condition.

2. Materials and Methods

2.1. Design

An RCT with three conditions (audiotape relaxation vs. virtual human relaxation vs. control) was conducted to assess the effects of relaxation on SBR and stress outcomes after tape-stripping. Recruitment occurred from May 2021 to February 2022.

2.2. Participants

A sample of 159 healthy adults was recruited from the local community. Participants were included if they were over the age of 18 and could speak, read, and write fluent English. Potential participants were excluded if they were allergic to adhesive tape, had an inflammatory skin or immunological-related health condition, took medication that affects immune functioning (e.g., prednisone), were pregnant, were over 60 years old, or had hearing difficulties or vision loss. Ethical approval was granted by the Auckland Health Research Ethics Committee (protocol code: AH21981). This study was pre-registered at https://anzctr.org.au (ACTRN12621000442808).

2.3. Power Analysis

The required sample size was calculated using the programme G*Power (v3.1.9.7) [25] for a one-way ANOVA for the outcome of SBR. A power level of 0.8 and a two-tailed significance level of α = 0.05 were chosen [26]. An expected effect size of F² = 0.28 was estimated from a tape-stripping study on changes in SBR after relaxation [13]. The sample size was increased by 20% due to possible errors in transepidermal water loss (TEWL) measurement, giving a total required sample size of N = 159 (53 per condition).

2.4. Procedure

Following standard salivary collection guidelines, participants were asked not to chew gum or drink caffeine, juice, or alcohol 18 h before the session and not to eat or brush their teeth the hour before. Participants were also asked to refrain from applying moisturiser on their arms, exercising, or showering within the hour before their session to reduce any interference in the TEWL measures. All sessions were conducted between 12:00 p.m. and 5:30 p.m. to reduce the effects of diurnal rhythms on salivary cortisol and alpha-amylase [27,28]. Each 90 min session took place in the University of Auckland Clinical Research Centre in ventilated rooms with a dehumidifier, as per the Tewameter probe recommendations, to control humidity levels [13].

The study procedure is provided in Figure 1. After receiving written informed consent, an Empatica (Cambridge, MA, USA) biosensor watch was placed skin-tight on the participants’ dominant wrist, and physiological variables were continuously collected throughout the session. The participants then completed baseline measures, including questionnaires that assessed their demographics, health behaviours, and psychological variables; a saliva sample; and TEWL measures. After the baseline measures, the participants were exposed to a standardised tape-stripping procedure to disrupt the skin barrier (see the tape-stripping section below). Immediately after this, the participants completed the TEWL, psychological, and saliva measurements again. The participants were then randomly assigned to one of three conditions (virtual human, audiotape, or control). The allocation sequence was generated using a random online number generator by a researcher uninvolved in running this study. Until this point, the randomised allocations were concealed from the primary researcher in sealed opaque envelopes. Participants were not told there was a control condition, and therefore, all three conditions believed they were receiving a relaxation intervention to control for potential placebo effects.

The participants were introduced to their randomised interventions and asked to follow the instructions during the 20 min intervention (recovery from the tape-stripping period). Those in the audiotape and virtual human condition were asked to listen and follow along with deep breathing and progressive muscle relaxation exercises through over-ear headphones on a 17-inch laptop screen positioned on a desk approximately 30 cm from the participant. The researcher started the videos before leaving the room, and the participants had some control over the playback (e.g., no pausing but able to adjust the volume). Those in the control condition were instructed to read a selection of magazines provided. The lights were dimmed in the two relaxation conditions but not for the control so they could read the magazines. All participants were asked to follow the relaxation instructions, not use any other electronic devices, not touch their wounds or pull down their sleeves, and not fall asleep. All the participants were video recorded during the intervention to measure whether they adhered to the instructions. After the intervention, the participants completed the final TEWL, psychological, and saliva measurements. They were also asked to provide feedback about their intervention’s acceptability, engagement, and delivery. The participants were given an NZD 30 voucher as compensation for their time. Figure 1 shows the flow of this study.

2.5. Interventions

2.5.1. Virtual Human Condition

Participants assigned to the virtual human condition watched a 20 min video of “Sam”, the virtual human, delivering deep abdominal breathing and progressive muscle relaxation exercises. The pre-recorded video format was chosen to isolate the effects of an additional visual-human-like presence while avoiding possible technical issues that could introduce variability. The virtual human was developed by Soul Machines (Auckland, New Zealand) and is shown in Figure 2. Sam was programmed to display human-like gestures (e.g., swaying her shoulders and nodding her head). Her body movements were accompanied by emotional facial expressions that were programmed through a text-to-speech Emotional Markup Language. This technique allowed her to express specific facial expressions if she said certain phrases; for example, she elicited a smile, warm eyes, and eyebrows if she said positive phrases (e.g., “feel how nice and relaxed you are”). Sam used a conversation engine (IBM Watson Agent, New York, NY, USA) that followed a pre-programmed relaxation script and spoke with a custom text-to-speech voice package based on a female human voice. The final intervention was a video of Sam reading through the relaxation script with piano background music.

2.5.2. Audiotape Condition

Participants allocated to the audiotape condition listened to a 20 min audio recording of a real human talking through the same relaxation script and background music as the virtual human condition.

2.5.3. Control Condition

The control condition was not provided with any relaxation besides a selection of magazines to read. In previous psychological intervention studies, magazines were used as a control condition as a neutral activity to prevent participants from experiencing negative emotions or stress from becoming too bored [15,29,30].

2.6. Outcome Measures

As shown in Figure 1, all wound healing, psychological, and physiological measures were taken at three time points throughout the session: at the baseline (timepoint 1), after the tape-stripping (timepoint 2), and after the intervention period (timepoint 3).

2.6.1. Tape Stripping and SBR Procedure

Tape-stripping is an experimental and standardised wounding technique that removes the outer layer of skin (the stratum corneum) through repeated application and removal of adhesive tape [31]. As a result, the skin barrier is disrupted and loses its ability to regulate water movement in and out [14]. It is a non-invasive, minimally painful, and quick procedure used in previous psychological intervention studies [13,29,30].

A baseline measure of skin barrier function was determined by collecting TEWL measures from a Tewameter TM300 probe (Courage + Khazaka, Cologne, Germany), which measured the evaporation rate (g/m²h) in the air layer next to the skin [30]. The TEWL measures the skin’s ability to prevent water loss; a high TEWL measurement indicates more water evaporating through the skin and a poorer skin barrier function. As the skin barrier heals, the TEWL returns to baseline levels, and this change in the TEWL over time provides a measure of SBR and an estimate of wound healing.

The participants placed their non-dominant arm flat on a pillow, and four 1 cm² sites were marked on the inside of their forearm, 1 cm below the elbow crease. The bottom site was left undamaged as the control site. The Tewameter probe was heated to 35 °C to match the skin temperature and was gently placed against the skin on each of the four sites for 60 s.

Once the baseline TEWL measures were taken, the three test sites were dry-shaven to ensure that hair was not pulled during the tape-stripping procedure. Standardised packaging tape (Scotch Commercial Grade Packaging Tape, 3M, Auckland, New Zealand) was applied with pressure to the three test sites and gently peeled off. After the first 20 strips of tape, the TEWL was measured to determine whether it had reached 15 g/m²h above the baseline. If not, ten more strips were applied, and the TEWL was measured again. The tape was applied until the TEWL was elevated to 15 g/m²h above baseline or until 40 pieces of tape were used. A TEWL measurement for each site was then collected to determine the skin barrier impairment levels immediately following the tape-stripping procedure.

2.6.2. TEWL Analysis

An overall TEWL reading for each site and time point was determined by averaging 20 consecutive measurements with a standard deviation below 0.6. The following formula was used to obtain the final SBR percentage from the TEWL readings [32,33]:

(TEWL_elevated − TEWL_recovery)/(TEWL_elevated − TEWL_baseline) × 100

A higher percentage indicates a higher SBR, indicating faster wound healing. There was a considerable variation in the TEWL data across the sample, possibly because the room temperature could not be controlled. The TEWL data were therefore screened only to include valid data. As shown in Figure 1, the SBR values were excluded if the baseline TEWL readings were more than two standard deviations (SDs) above the mean (n = 5), readings were not complete for each time point (n = 7), the skin was not elevated enough (more than 5 g/m²h above the baseline, n = 11), or the skin was damaged more than two SDs of elevation above the mean (n = 4), as these denoted outliers in the data.

After the initial exclusion, the SBR values of the remaining sites were all averaged if they were all within 10 g/m²h of each other. If not, the two closest sites were averaged if they were within 30 g/m²h, as these indicated unreliable readings (n = 4). After the exclusion, 44 participants were in the control condition, 42 in the audiotape condition, and 43 in the virtual human condition.

2.6.3. Demographics and Health Behaviours

At the baseline, the participants were asked about their general demographics, including age, gender, weight, height, ethnicity, employment status, and education level. The participants were also asked about their alcohol consumption, physical exercise, diet, smoking status, medication, and sleep. These were assessed because they were associated with stress, immune, and tape-stripping outcomes in previous studies [15,29,34,35].

2.6.4. Psychological Measures

The 10-item Perceived Stress Scale (PSS) was administered at baseline to assess how stressful, unpredictable, and uncontrollable the participants perceived their current lives [26]. The participants rated how often they felt or thought a certain way in the last month on a scale from 0 (never) to 4 (very often). The scores were totalled to give a final score between 0 and 40. Higher scores on the PSS indicate that the participants were experiencing a greater level of perceived general life stress than those with lower PSS scores. The scale showed good reliability (α > 0.70) and validity in different ethnicities and clinical samples.

The self-reported stress, relaxation, pain, and anxiety were measured at the baseline, post-tape-stripping, and post-intervention using 100 mm visual analogue scales. The participants dragged a digital slider on a scale to mark an exact number between 0 and 100 to indicate how they were currently feeling (a higher score indicates higher levels of these feelings). The anchors on the scales were as follows: not at all stressed to extremely stressed, not at all relaxed to extremely relaxed, no sensation of pain to the highest sensation of pain imaginable, and not at all anxious to very anxious. Measurement errors (i.e., participants not dragging the slider) were coded as missing data in the final analysis.

2.6.5. Physiological Measures

The participants were given a wearable biosensor (Empatica E4) watch to wear on their dominant wrist to collect biometric data. The watch collected continuous data for the entirety of the session that were time-stamped after each questionnaire was completed. The heart rate and electrodermal activity were collected as two physiological stress response measures that reflect the degree of parasympathetic or sympathetic arousal [36,37,38]. Electrodermal activity was processed using a validated tool that removed confounding artefacts [39].

Saliva samples were taken at the three time points using standardised SaliCaps collection devices (IBL, Hamburg, Germany). These samples were collected to evaluate whether any interventions caused changes in the salivary stress biomarkers (cortisol and alpha-amylase). The participants were instructed to rinse their mouths with water before collecting saliva using the passive drool technique in the SaliCap tube. After the collection, the saliva samples were stored at −20 °C at the University of Auckland before being shipped on dry ice to the University of Vienna, Austria, for analysis. The salivary cortisol concentrations were determined using a commercial enzyme-linked immunosorbent assay (ELISA, IBL, Hamburg, Germany). The salivary alpha-amylase was determined by a kinetic colorimetric test with reagents from Roche (Roche Diagnostics, Mannheim, Germany). Both tests’ intra- and inter-assay coefficients of variance were below 10%.

2.6.6. Engagement, Satisfaction, and Feedback

After the intervention, the participants completed two 100 mm VAS tests to rate how satisfied and engaging they found the session. The anchors were as follows: not engaging at all (0) to extremely engaging (100) and not at all satisfied (0) to very satisfied (100).

Lastly, the participants were asked to provide open-ended feedback about the relaxation session, including, “What did you like about the delivery of the relaxation session?” and “How do you think the delivery of the relaxation session could be improved?”. These questions were adapted from Loveys and colleagues’ virtual human and CBSM feasibility study [40] and were given to all the participants.

2.6.7. Adherence to Instructions

The extent to which participants adhered to the relaxation and wound-healing instructions was coded from the video recordings of the intervention period. This measure was used to control for potential influences in the healing process, as noted in previous tape-stripping studies [29,30]. Each video was split into 5 min blocks, and if at least one block was coded with non-adherence, the participant was coded as overall non-adherence. The data were then converted into a percentage for each condition to indicate the proportion of participants who followed the instructions.

2.7. Statistical Analysis

2.7.1. Quantitative Data

The data were analysed using IBM SPSS Statistics (version 28) [41]. The data were checked for errors, outliers, and violations of the normality assumption before the analysis. Shapiro–Wilk tests were performed to test the assumption of normality, with results of p < 0.001 considered a violation of the normality assumption. An ANOVA was conducted to test the differences in the mean SBR across the conditions. A 3-step hierarchical regression was conducted to evaluate the effect of condition on the SBR, controlling for covariates in the first two steps of the model. Tape-stripping variables known to affect the SBR (number of strips of tape and levels of skin barrier impairment) were added into the first step of the model. Variables that differed at the baseline (sleep hours and relaxation and stress levels) and known covariates from previous studies were entered into the second step of the model (body mass index (BMI), age, and gender) [15,29]. As the condition was a categorical variable, it was first dummy-coded before being added to the analysis to compare the audiotape and virtual human with the control condition.

ANCOVAs were conducted to assess the differences in post-intervention scores of the psychological and physiological measures after controlling for baseline scores to account for any baseline differences between conditions. The cortisol and alpha-amylase data violated the normality assumption and were therefore transformed using a natural log transformation, with logged data used. Mixed ANOVAs were further conducted to test for the main time effects on all secondary outcomes. Kruskal–Wallis tests were conducted to assess the condition effects on satisfaction and engagement ratings, as these measures violated the normality assumption.

2.7.2. Qualitative Data

The responses to the two open-ended questions for those in the virtual human group were analysed using an inductive approach to code and generate the themes and subthemes. All coded data for each theme and subtheme were collated and summarised.

3. Results

3.1. Sample Characteristics

The demographic and baseline characteristics of the sample are shown in

Table 1. Demographic characteristics, health behaviours, and tape-stripping measures across conditions.

	Condition
	Control (n = 53)	Audiotape (n = 53)	Virtual Human (n = 53)	p-Value
Demographics
Age (years), M (SD)	25.8 (9.6)	25.6 (6.4)	26.9 (7.6)	0.758 a
Gender:				0.815 b
Female, n (%)	41 (77.4)	38 (71.7)	37 (69.8)
Male, n (%)	12 (22.6)	14 (26.4)	15 (28.3)
Non-binary/other, n (%)	0 (0)	1 (1.9)	1 (1.9)
Ethnicity:				0.475 b
NZ European, n (%)	20 (38.5)	16 (30)	21 (40)
Māori/Pacific, n (%)	4 (8)	5 (9.5)	1 (2)
Chinese, n (%)	9 (17)	9 (17)	5 (10)
Other, n (%)	19 (36.5)	23 (43.5)	25 (48)
Education level:				0.113 b
High school or less, n (%)	24 (45.3)	17 (32.1)	9 (17)
Trade, n (%)	1 (1.9)	2 (3.8)	1 (1.9)
Undergraduate, n (%)	15 (28.3)	16 (30.2)	21 (39.6)
Postgraduate, n (%)	13 (24.5)	17 (32.1)	21 (39.6)
Employment status:				0.446 b
Full-time, n (%)	9 (17)	15 (28.3)	16 (30.2)
Part-time, n (%)	9 (17)	3 (5.7)	5 (9.4)
Student, n (%)	31 (58.5)	32 (60.4)	29 (54.7)
Unemployed, n (%)	4 (7.5)	3 (5.6)	3 (5.7)
BMI, M (SD)	23 (3.8)	23.6 (3.8)	24 (4.3)	0.758 a
Health behaviours
Exercise (days/week), M (SD)	3.4 (2.1)	4.4 (2.3)	3.7 (1.8)	0.069 a
Sleep (hours/night), M (SD)	7.51 (0.95)	6.49 (2.09)	7.18 (1.34)	0.003 a*
PSS score, M (SD)	18.7 (4.9)	17.2 (3.8)	18.4 (4.3)	0.172 a
Tape-stripping measures
Baseline TEWL (g/m2/h), M (SD)	19.5 (5.1)	18.7 (5.2)	19.8 (5.4)	0.547 a
TEWL impairment (g/m2/h), M (SD)	33.4 (11.4)	30.5 (8.7)	35.0 (12.1)	0.106 a
Strips used, M (SD)	34.9 (7.8)	36.3 (7.4)	33.9 (8.4)	0.287 a

Note: M—mean, SD—standard deviation, %—percentage of participants in that category, ^a—ANOVA, ^b—chi-square test, BMI—body mass index, PSS—perceived stress scale, VAS—visual analogue scale, TEWL—transepidermal water loss, *—significant difference (p < 0.05).

. The age range of the sample was 18 to 58 years (M = 26.1, SD = 7.9). Most of the sample were female (N = 116, 73%) and almost half the participants identified as Asian (N = 75, 47.2%). Table 1 shows no significant differences between the conditions regarding the demographic variables. However, significant differences existed between the conditions regarding sleep hours (p = 0.003), whereby the control had significantly higher sleep hours than the audiotape condition.

3.2. Adherence to Instructions

A total of 130 participants followed the instructions for the intervention and wound healing (82%). There were no significant differences between the conditions regarding adherence.

3.3. Skin Barrier Recovery

The mean SBR rate for each condition is presented in Figure 3. A one-way ANOVA revealed no significant differences in the mean SBR between the conditions (F(2,122) = 0.2, p = 0.818, ηp² = 0.003).

A follow-up three-step hierarchical regression was conducted to analyse the predictive effects of the condition on the SBR, controlling for known covariates. Step 1 of the regression model (containing tape-stripping variables that affect SBR) was significant (R² = 0.04, ΔF (2,121) = 3.55, p = 0.032), where these variables explained a significant 40% of the variance in the SBR. Step 2 of the regression model (containing variables that differed at baseline and known covariates) accounted for an additional non-significant 26% of the variance in the SBR (R² = 0.066, ΔF (6,115) = 1.57, p = 0.161). Adding the dummy codes for the condition into Step 3 of the model did not significantly explain any additional variance in the model (R² = 0.059, ΔF (1,114) = 0.08, p = 0.778). This further demonstrated that the condition allocation did not significantly predict a participant’s level of SBR.

3.4. Psychological Variables

Figure 4 shows the estimated marginal mean scores of post-intervention relaxation across conditions after controlling for the baseline scores. The ANCOVAs revealed a significant difference in the mean relaxation scores between the conditions at the post-intervention time point after controlling for the baseline relaxation scores (F(2,151) = 5.432, p = 0.005, ηp² = 0.07), with a large effect size. The post hoc tests showed that those in the audiotape condition (M = 87.4, SE = 2.13, p = 0.007, Cohen’s d = 0.61) and the virtual human condition (M = 85.7, SE = 2.13, p = 0.038, Cohen’s d = 0.50) were significantly more relaxed than those in the control condition (M = 78.02, SE = 2.13) post-intervention. There were no significant differences in the relaxation levels between the audiotape and virtual human conditions (p > 0.999, Cohen’s d = 0.11), suggesting that both relaxation interventions were similarly effective at increasing self-reported relaxation.

Figure 5 shows the estimated marginal mean of the post-intervention anxiety scores across the conditions after controlling for the baseline scores. There were significant differences in the post-intervention anxiety scores between the conditions after controlling for the baseline anxiety scores (F(2,155) = 4.25, p = 0.016, ηp² = 0.05), with a moderate effect size. The post hoc tests revealed that those in the audiotape condition had significantly lower post-intervention anxiety scores (M = 7.39, SE = 1.56) than those in the control condition (M = 13.86, SE= 1.57, p = 0.012, Cohen’s d = 0.57). There were no significant differences in the post-intervention anxiety scores between the virtual human (M = 10.73, SE = 1.56) and control conditions (p = 0.480, Cohen’s d = 0.27) and between the audiotape and virtual human conditions (p = 0.394, Cohen’s d = 0.29). Figure 5 shows that although the virtual human group had lower anxiety scores than the control group and higher anxiety scores than the audiotape condition with small effect sizes, these differences were not statistically significant.

There were no significant differences in the post-intervention VAS stress scores (F(2,155) = 0.59, p = 0.558, ηp² = 0.01) and pain scores (F(2,155) = 0.0.91, p = 0.407, ηp² = 0.01) between the conditions after controlling for baseline stress and pain, respectively.

Table 2. Secondary outcomes at each time point across conditions.

	Condition	Baseline	Post-Tape-Stripping	Post-Intervention
VAS stress, M (SD)	Virtual human	40.53 (27.95)	22.85 (23.64)	16.66 (23.61)
	Audiotape	33.08 (25.50)	19.53 (21.31)	10.62 (15.07)
	Control	28.66 (21.93)	15.19 (19.06)	11.47 (17.21)
	Total	34.09 (25.56)	19.19 (21.51)	12.92 (19.05)
VAS relaxation, M (SD)	Virtual human	56.63 (27.47)	60.52 (25.22)	83.12 (19.97)
	Audiotape	63.94 (27.80)	69.82 (25.75)	87.47 (13.18)
	Control	70.33 (23.88)	70.40 (23.03)	80.50 (20.38)
	Total	63.63 (27.47)	66.90 (24.95)	83.67 (18.28)
VAS anxiety, M (SD)	Virtual human	24.85 (27.65)	15.94 (20.97)	12.32 (19.43)
	Audiotape	24.74 (23.61)	15.38 (19.11)	8.92 (13.40)
	Control	16.23 (21.96)	13.55 (20.32)	10.74 (19.12)
	Total	21.94 (24.70)	14.96 (20.05)	10.66 (17.48)
VAS pain, M (SD)	Virtual human	9.55 (18.53)	8.51 (12.59)	4.77 (10.89)
	Audiotape	8.58 (13.32)	9.08 (12.84)	3.70 (6.53)
	Control	8.30 (16.35)	9.21 (14.95)	5.49 (11.99)
	Total	8.81 (16.12)	8.93 (13.42)	4.65 (10.05)
Heart rate, M (SD)	Virtual human	82.03 (8.86)	79.70 (10.76)	74.36 (8.90)
	Audiotape	80.67 (11.81)	80.87 (9.73)	71.62 (10.18)
	Control	80.39 (9.33)	82.04 (13.88)	74.60 (10.18)
	Total	81.02 (10.07)	80.88 (11.53)	73.49 (9.78)
Electrodermal activity, M (SD)	Virtual human	0.70 (1.29)	0.51 (0.71)	0.30 (0.26)
	Audiotape	0.26 (0.42)	0.60 (1.23)	0.30 (0.39)
	Control	0.70 (1.52)	0.97 (2.02)	0.50 (0.93)
	Total	0.54 (1.18)	0.69 (1.44)	0.37 (0.60)
Salivary cortisol, M (SD)	Virtual human	0.71 (0.74)	0.51 (0.71)	0.32 (0.70)
	Audiotape	0.63 (0.57)	0.50 (0.82)	0.36 (0.52)
	Control	0.70 (0.58)	0.48 (0.52)	0.27 (0.56)
	Total	0.65 (0.65)	0.48 (0.69)	0.31 (0.59)
Salivary alpha-amylase, M (SD)	Virtual human	3.77 (1.10)	3.86 (0.99)	3.71 (0.99))
	Audiotape	4.04 (0.86)	4.13 (0.84)	4.19 (0.84)
	Control	3.72 (0.95)	3.82 (0.92)	3.74 (0.94)
	Total	4.84 (0.97)	3.90 (0.95)	3.85 (0.97)

Note: M—mean, SD—standard deviation; reported salivary cortisol and alpha-amylase means and SDs are natural log-transformed data.

shows the mean scores and standard deviations for the psychological variables at each time point. These means suggest there may have been significant changes over time for all the groups. Therefore, we performed mixed ANOVAs to investigate the time effects. These showed main time effects for the VAS stress (F(1,156) = 17.69, p < 0.001, ηp² = 0.43), relaxation (F(1,152) = 120.06, p < 0.001, ηp² = 0.44), anxiety (F(1,156) = 80.78, p < 0.001, ηp² = 0.34), and pain (F(1,156) = 19.30, p < 0.001, ηp² = 0.11), which all reduced over time with large effect sizes.

3.5. Physiological Variables

Using one-way ANCOVAs, controlling for the baseline values, there were significant differences in the post-intervention salivary alpha-amylase values between the conditions after controlling for the baseline salivary alpha-amylase values (F(2,153) = 3.27, p = 0.041, ηp² = 0.04), with a moderate effect size. Figure 6 shows the estimated natural log-transformed marginal mean of post-intervention salivary alpha-amylase values across the conditions. The post hoc tests revealed that those in the virtual human condition had significantly lower post-intervention salivary alpha-amylase values (M = 4.04, SE = 0.08) than those in the audiotape condition (M = 3.83, SE = 0.08, p = 0.015, Cohen’s d = 0.34). There were no significant differences in the post-intervention salivary alpha-amylase values between the control (M = 3.77, SE = 0.08) and virtual human conditions (p = 0.533, Cohen’s d = 0.12) and between the audiotape and control conditions (p = 0.066, Cohen’s d = 0.34). It is possible that the audiotape condition had higher alpha-amylase due to less sleep at the baseline, although statistically controlling for sleep made no difference to the results of the ANCOVA.

There were no significant differences between the conditions in the post-intervention heart rate (F(2,155) = 1.34, p = 0.266, η² = 0.02), electrodermal activity (F(2,155) = 1.60, p = 0.205, η² = 0.02), and salivary cortisol (F(2,148) = 0.1.46, p = 0.235, ηp² = 0.02) after controlling for the baseline scores.

Table 2 shows the mean scores and standard deviations for the physiological variables at each time point. These means suggest there may have been significant changes over time for all the groups. Therefore, we performed mixed ANOVAs to investigate the time effects. These showed main time effects for heart rate (F(1,145) = 82.35, p < 0.001, ηp² = 0.36) and cortisol (F(1,148) = 89.94, p < 0.001, ηp² = 0.38), which both reduced over time with large effect sizes. There were no main effects of time on electrodermal activity or salivary alpha-amylase.

3.6. Satisfaction and Engagement Ratings

Figure 7a shows significant differences in the mean rank satisfaction scores between the conditions (H(2, N = 159) = 8.26, p = 0.016, η² = 0.04), with a small effect size. The post hoc tests revealed that satisfaction was significantly higher in the audiotape condition (mean rank = 94.55) than in the virtual human condition (mean rank = 70.58, p = 0.021). No significant differences existed between the audiotape and control conditions (mean rank = 74.88, p = 0.081) or between the virtual human and control conditions (p = 0.629).

As shown in Figure 7b, there were also significant differences in the mean rank engagement scores of the interventions between the conditions (H(2, N = 159) = 7.56, p = 0.023, η² = 0.04), with a small effect size. The post hoc tests revealed that engagement was significantly higher in the audiotape condition (mean rank = 91.96) than in the control condition (mean rank = 67.47, p = 0.006). There were no significant differences between the audiotape and virtual human conditions (mean rank = 78.84, p = 0.139) nor between the control and virtual human conditions (p = 0.202).

3.7. Qualitative Data

Participants identified several strengths and improvements of Sam, the virtual human, from the two open-ended questions. Themes, subthemes, and representative quotes for the strengths and improvements are presented in

Table 3. Themes and representative quotes describing what participants liked most about Sam.

	Themes and Subthemes	Representative Quotes [Participant ID]
Strengths
	Sam’s behaviour
	Eye contact	“Clear instructions, eye contact, good background music” [p131]
	Mannerisms	“The person had a nice calm manner” [p79] “I liked how the person moved like an actual person it made it feel less awkward” [p85]
	Sam’s voice
	Tone	“The music and calm tone of the voice” [p29]
	Pace	“Was very calm and at a good pace” [p143]
Improvements
	Sam’s appearance
	Less uncanny	“Avatar was a bit creepy” [p3]
	More facial expressions	“Perhaps more expression in the face” [p21]
	Sam’s behaviour
	Mannerisms	“The animation was slightly unsettling, being that it was quite robotic.” [p153]
	Better lip-syncing	“The digital human’s mouth moved just a bit slower than the words so that was a little disconcerting.” [p29]
	Sam’s voice
	More human-like	“Needs more human intonation” [p52] “Maybe make the person a little more friendly and personable” [p85] “It felt a bit jolty at the start of her sentences” [p138]
	Conversation design
	More engaging	“Needs more engagement i.e., thinking about what’s on for the rest of the day and how you’re going to handle it” [p53]
	More personalisation	“Maybe more personalised with use of my name” [p99]

. Overall, the participants liked aspects of Sam’s behaviour and voice, including her maintaining eye contact, human-like movements, and the calm nature of her voice. However, participants felt that many improvements could have been made regarding Sam’s appearance, behaviour, voice, and conversation design. Many participants reported that Sam was creepy or unsettling, and they would have preferred a more human or natural-looking appearance and mannerisms. Others disliked the mismatch between the voice and the lip movements, noting that this became distracting and uncanny. Regarding the conversation design, participants would have liked the exercises to be more engaging, interactive, and personalised.

4. Discussion

Developing and testing innovative ways to deliver psychological interventions has become a focus of digital health research in the past decade. This RCT investigated whether relaxation exercises, when delivered by a virtual human or an audiotape, could improve wound healing, reduce stress, and increase relaxation compared with a control group.

This study found no significant differences between the relaxation conditions and the control condition regarding the wound healing rates. This finding contrasts with a prior study, which found that audiotape relaxation instructions delivered before or after tape stripping improved skin barrier recovery compared with a quiet reading control group [12]. The null healing findings in the current study could have been due to the sample’s lower stress levels at baseline, which may have caused a floor effect for stress reduction. Previous research suggests that psychological interventions may only influence immune parameters when the sample is sufficiently stressed at baseline [42,43]. To account for this, future studies could manipulate stress using an acute experimental stressor, such as the Trier Social Stress Test, which was shown to impair SBR in previous tape-stripping studies [31,32]. Alternatively, future studies could specifically recruit participants who identify as highly stressed.

There were also some methodological differences compared with Robinson and colleagues’ study [13] that may have contributed to the null findings in wound healing between the groups. The relaxation conditions in the prior study included 20 min of relaxation plus 20 min of quiet reading, whereas the relaxation conditions in this study included 20 min of relaxation alone. Effectively, this meant a shorter intervention than this study [13].

The current study built upon the previous study by including measures of physiological stress. The heart rate and salivary cortisol reduced over time in all the groups, suggesting that all the conditions experienced reduced physiological arousal over time. The salivary alpha-amylase was the only physiological variable to significantly differ post-intervention between the conditions. We hypothesised that both the relaxation groups would experience reduced salivary alpha-amylase, and it is unclear why this was reduced in the virtual human relaxation group and not the audiotape relaxation group. This result could have been due to the baseline variability in sleep, relaxation, or alpha-amylase between the groups or sample size considerations. Individual differences in the stress reactivity may have influenced the responses, despite the statistical control. Although this study was powered to detect moderate-to-large effects, the sample size of each condition may have been too small to detect smaller effects between relaxation conditions and the control. Future research with larger sample sizes could help investigate the effects of different relaxation delivery modalities.

To achieve meaningful improvements in healing and physiological stress between the control and relaxation conditions, individuals may need to engage in longer relaxation sessions or in repeated, consistent use of relaxation techniques [10,44]. A longer relaxation period may have revealed greater differences in the physiological outcomes between the relaxation interventions and the control. Future research could build on the current study and examine the effects of the repeated use of relaxation delivered by a VH.

Both relaxation conditions increased relaxation more than the control group did, which is a promising outcome that aligns with previous studies. There were also main time effects on all the psychological variables, which showed improvement in all the conditions over time. While the control condition was designed as a neutral activity, participants surprisingly reported enjoyment in sitting and doing something easy. It is likely they welcomed a break or distraction from their devices and home or work responsibilities (especially during the COVID-19 pandemic). Distraction was previously linked with improved wound healing in previous studies [45,46], and therefore, the magazines may have acted as a distraction that could have had a similar stress reduction as the relaxation intervention groups. All participants (regardless of condition) believed they were receiving a relaxation intervention, which might have increased the expectation effects across all the conditions. Studies showed mixed findings across magazine controls, suggesting variability in their appropriateness for valid intervention comparison [11,13,47].

The satisfaction and engagement ratings of the virtual human were adequate, but higher ratings would be desirable. The qualitative analysis found concerns about the virtual human’s robotic voice tone, jolts, lip-syncing, and unsettling appearance, which may suggest uncanny valley effects were present for some people. This phenomenon occurs when a robot or artificial agent resembles a human but has subtle features lacking full realism that can evoke discomfort and unease [48]. The VH introduced novel visual and auditory sensory information, while the audiotape was a human voice recording. Although previous research has shown that multimodal relaxation can enhance the relaxation response, it is possible that the mismatch between the speech and lip movements, and the synthetic voice in the current VH were disappointing, therefore reducing the participants’ overall satisfaction and engagement ratings. This could also have been one of the reasons why the virtual human had weaker effects on anxiety than the audiotape condition. However, VH technology is rapidly evolving, and more advanced models with higher behavioural realism and more natural voices are continually being developed.

Another limitation of this study was the absence of a human video condition, which could have provided a direct comparison between different visual modalities. This study focused on a pre-recorded VH video to ensure technical consistency and compared it with an audio recording because audio is a common way to deliver relaxation instructions. Other research compared human telepresence (and human video) to an interactive digital human delivering mindfulness, and more work could be performed using these methods [49].

Tape-stripping is a minor wound that is sensitive to environmental conditions and lacks real-life clinical or chronic wound characteristics. Whilst the TEWL guidelines were followed, unexpected changes in conditions due to COVID-19 (e.g., wearing a mask, social distancing, extra cleaning procedures) and temperature fluctuations over the recruitment period may have limited the number of valid TEWL measurements and underpowered the SBR analysis. Future work could consider a different type of wound with more ecological validity.

Lastly, most participants in this study had never interacted with a VH before, so there may have been some hesitancy, negative expectations, or excitement in seeing Sam. As people become more accustomed to using VHs, they will likely become more comfortable with them and results may differ. Future research could include measures of familiarity to test whether previous experience with VHs affects outcomes.

5. Conclusions

This RCT study tested the psychological and physiological effects of relaxation delivered by a VH compared with audiotapes and quiet reading. The analysis over time showed that all three conditions showed reductions in stress, relaxation, anxiety, and pain after the intervention. The between-group analyses showed no significant differences in wound healing between the groups nor stress reductions. There were between-group differences in anxiety and relaxation, where the audiotape group showed greater reductions in anxiety, and both the VH and audiotape showed greater increases in relaxation compared with the control group. Physiologically, alpha-amylase was significantly reduced only in the VH group.

The audiotape had significantly better satisfaction than the other groups and higher engagement than the control group. The qualitative findings highlighted issues with the VH’s synthetic voice quality, appearance, and interactivity, which might have been why it was rated less satisfactory.

These findings suggest that while this VH holds promise for delivering a relaxation intervention, further technological improvements are needed. Future research should explore ways to improve their voice and behavioural realism. Replication in a more stressed population or with a different wound model may clarify the clinical benefits.