1 Background

1.1 Rationale

1.1.1 Virtual reality as a new technology

Virtual rehabilitation using Virtual Reality (VR) technology is a promising novel approach to motor rehabilitation after a stroke (Levin et al. 2014), in multiple sclerosis (García-Muñoz et al. 2022) and Parkinson’s disease (Pazzaglia et al. 2019). It can also add some positive elements to current rehabilitation strategies. Several advantages of virtual rehabilitation can be suggested in terms of rehabilitation intensity and motivation. VR can encourage patient participation by enhancing the pleasurable and fun aspects (Mihelj et al. 2012; García-Bravo et al. 2019; Choi and Paik 2018). It also offers the option to design personalised and flexible rehabilitation programmes based on the motor disability of the patient (W.-S. Kim et al. 2020). Its use can also be considered as a remote or home-based rehabilitation tool (Holden et al. 2007). Finally, it is possible to functionally assess and digitally monitor patient progress using motion sensors combined with VR systems (W.-S. Kim et al. 2016).

1.1.2 Difference in motor skills between real-life and virtual environments

Despite the progress that has been made, significant differences remain between VR and real life. In terms of kinematics, upper limb movements in stroke patients have been shown to differ between VR and real-life environments (Viau et al. 2004). Similarly, several studies using movement-to-target tasks have also shown that Immersive Virtual Reality (IVR) movements are slower than those in a real-life environment and that spatial and temporal kinematics differ between IVR and real-life environments (Hussain et al. 2018; Levin et al. 2015). The distribution of the centre of pressure differs between real-life environments, non-immersive VR with 2D flat screens and IVR via an HMD (Lott et al. 2003). As the goal of rehabilitation is to improve independence in real life, these different movement kinematics may affect the transfer of learning between VR and real-life environments. In addition, visual feedback provides information about body position and minimises postural sway during standing tasks (Horiuchi et al. 2017). This means that just wearing a Head Mounted Display (HMD) can alter postural control during standing tasks. The 2021 study by Chander confirmed that IVR environments can induce increased postural instability (Chander et al. 2020).

Moreover, the presence of an avatar in IVR environments is a recent occurrence; however, we already know that it affects participant (Mann 2020). Having the user control the body movements of the avatar using their own can even induce a sense of ownership in which the user considers that the body parts of the avatar are a substitute for their own, a phenomenon called “virtual embodiment” (Lugrin et al. 2015). Mohler (Mohler et al. 2008) found that participants who explored nearby space while seeing an avatar made more accurate judgements of absolute egocentric distance for locations ranging from 4–6 m away from where they were standing than participants who saw no avatar. Ries (Ries et al. 2009) showed that the accuracy of distance estimation was related to the notion of presence and not to the memorisation of spatial structures in the real-life environment. In a study of obstacle negotiation with and without lower limb representation, the authors showed that the fidelity of visual information related to the lower limbs influences both the anticipation and feedback aspects of visual/motor coordination during obstacle negotiation (Kim and Lee 2019). It has also been shown that having a self-avatar, or seeing an animated VR character, improves performance during an interaction task in a virtual environment compared to having no visual representation (Seinfeld et al. 2022).

1.1.3 Telerehabilitation and IVR

Numerous studies, including those by Laver (2020), Nikolaev (2022), and Su (2023), attest to the comparable effectiveness of telerehabilitation with traditional one-on-one rehabilitation. Nevertheless, the widespread adoption of immersive virtual reality remains constrained. Although devices designed for upper limb deficits in a seated position are available, challenges related to balance and the risk of falling emerge when patients transition to a standing position. A preliminary trial underscored the critical necessity for hands-on patient assistance during rehabilitation sessions to mitigate fall risks. This risk is heightened, particularly when balance is compromised by the use of a Head-Mounted Display (HMD) (Proffitt 2018). Kim’s study in 2017 demonstrated the feasibility of walking in an HMD; however, results revealed an increase in the Center of Pressure (CoP) sway area with eyes closed and open among all participants with Parkinson’s disease (Kim 2017). Similarly, Epure et al. (2016) concluded that HMD usage leads to increased instability. The pivotal question is to precisely quantify this degradation and explore diverse avenues to alleviate it.

1.1.4 How to enhance instability through visual flows?

Equally, vertical visual information helps generate appropriate muscle activation models to control balance. The central nervous system uses feedback from the somatosensory, visual and vestibular systems (Redfern et al. 2001). Lord et al. confirmed the importance of vision, particularly contrast sensitivity and stereopsis, in controlling posture in challenging conditions (Lord and Menz 2000). Deteriorating visual function in older people has also been shown to impair gait and balance control and therefore increases the risk of falls (Black and Wood 2005). However, the mean visual field of virtual reality headsets ranges from 39.6° to 55° (Lynn et al. 2020) much lower than the 110° in normal vision. This loss of peripheral visual field causes postural instability, reduces the weighting of the visual input and increases that of the vestibular input to maintain balance (Taneda et al. 2021). Elsewhere, the results of the Simeonov study show that a simple vertical structure can be used as a visual cue to improve balance in construction workers (Simeonov et al. 2008).

Outstanding questions requiring research are: (1) a determination of reference values between real-life and virtual environments and a measurement of the difference between the two (2) followed by a determination to establish whether the balance disturbances created by virtual reality can be reduced by using a full-body avatarisation technique and/or by increasing the amount of visual information required in the environment.

2 Methods

2.1 Design

This was a prospective, quasi-experimental, one-group study involving two sessions.

  • The purpose of the first session, which was carried out in multiple centres, was to collect reference values on balance performance in real life, firstly, and in virtual reality, secondly.

  • The purpose of the second session, carried out at a single centre, was to investigate the performance of balance tests in three virtual conditions: with a full-body avatar and/or reinforced vertical visual cues in the virtual environment. A measurement was taken in real-life conditions after the three virtual conditions. The goals of this session were to determine if the impaired balance could be corrected.

The method did not change significantly after the start of the study.

2.2 Participants

In both sessions, the eligibility criteria for participants were to be aged over 18 years. Exclusion criteria were immobilisation or incapacity of the upper limbs, pain or joint blockage of the lower limbs, major visual impairment. To avoid causing discomfort to participants or making them feel queasy due to motion sickness, participants were excluded if they had a Motion Sickness Sensitivity Questionnaire MSSQ score > 25 (Golding 2006). Since under certain conditions women may be more susceptible to motion sickness (Zhang et al. 2015), gender parity was sought.

The Saint-Hélier Fondation in Rennes (France) is a rehabilitation centre and an appeal was made to the staff employed by the facility.

2.3 Interventions

2.3.1 Balance tests

These consisted of four 30 s tests. The first was a static single-leg balance test, on the left foot (30 s) then on the right foot (30 s). The second was a ‘dynamic’ test. In a single-leg balance for 30 s, each participant mobilised both upper limbs in four sequences (Fig. 1). The first sequence started with the arms by the sides of the body. The arms were then raised above the head, in an upward stretch. The second sequence consisted in lowering both arms to the side and then horizontally. The third sequence consisted of moving horizontally both upper limbs from the supporting leg to the opposite leg. The final sequence required participants to bring the arms back to the side of their body. These four movements explore the sagittal, frontal, horizontal and then frontal again planes to complete the cycle. The two tests are performed on the left foot and then the right foot (Fig. 1). The centre of pressure was recorded on a force plate (TYMO® from Tyromotion GmbH).

Fig. 1
figure 1

Sequence of movements for the dynamic imbalance test, starting with the arms along the body in unipodal balance; 1 = raising the arms upwards; 2 = inclination of the two arms on the side of the supporting leg until they are horizontal; 3 = horizontal movement of the arms on the opposite side; 4 = return to the starting position

Full size image

2.3.2 Virtual environment

The virtual environment was created based on a room in the rehabilitation center (Fig. 2A, B). The room was scanned to replicate the volumes and configuration. The furniture was reconstructed in 3D to match the room’s furniture, including the bed, a nightstand, and a red armchair. Evacuation instruction posters on the wall were replicated. The window is placed in the same location, and the room’s brightness and lighting represent the average conditions of early afternoon. The force platform was also scanned and integrated into the virtual room. All balance tests, whether in real or virtual settings, take place in the same location facing the bed, one meter from the foot of the bed. No interaction is possible with elements in the virtual environment. Participants are brought to the force platform in a testing position and are instructed to put on the Head-Mounted Display (HMD) at that moment. The environment then appears identical to reality for them.

Fig. 2
figure 2

Images of four experimental conditions. A photograph of the real room; B reproduction of the room in virtual reality; C room in virtual reality with enhanced visual information; D projection of the avatar in the virtual environment. The avatar is represented in the first person for the participants

Full size image

External control allows operators to modify the wall covering, making it appear as a vintage decor with vertical stripes (Fig. 2C) or to make the avatar appear (Fig. 2D).

The four experimental conditions, three in virtual reality plus the real condition, are available to operators.

2.3.3 Avatarisation

The movement of participants was captured using an RGB-D camera (colour camera with depth card), of a Kinect Azure type. The IRT b-com (Technological Research Institute) in Rennes has developed 3D body movement tracking algorithms. Artefacto was responsible for rendering the avatar in IVR and for its recalibration to ensure its consistency with the perspective of the immersed user.

2.3.4 Conduct of the intervention

  • Session 1 Participants performed static and dynamic balance tests in real-life and immersive situations. In the immersive situation, conditions were identical to reality with no extra information or avatarisation. In both scenarios, the view of the real-life room and the artificially-created room were identical (i.e. one metre from the bed and pointing towards the back of the room). Each participant performed three tests for each condition. The sequence order between real and virtual environments was randomised. Three tests were conducted for each of the two conditions. The aim was to measure the difference between reality (reality1) and virtual environments, and to investigate the influence of the sequence order and determine whether a learning effect occurred over the course of the three tests. The goal was also to determine reference values for session 2.

  • Session 2 A second series of tests was conducted involving three different conditions by taking the results of the first measurements into account. A single test was performed for each condition. In virtual reality, the three conditions were: (1) the presence of a full-body avatar, (2) replacement of white walls with wallpaper of contrasting vertical stripes, and (3) the combination of the two previous conditions. The sequence of the three conditions was randomised into four arms. A test in real-life conditions (real2) was always repeated at the end of the session.

2.4 Outcome measures

The primary outcome measure was the distance travelled by the centre of pressure (COP) (Chen et al. 2021). Data was recorded for a standard duration of 30 s for each exercise using the Tymo plate (Tyromotion V 5.2).

The secondary outcome measure was motion sickness, measured by the SSQ.

No changes were made to the outcome measure after the start of the study.

2.5 Sample size

Cohen’s d for calculating the detectable internal equivalence was performed using the.

TOSTER package version 0.3.4 under r statistic, “powerTOSTpaired” function.

The equivalence bounds to achieve 80% power for a small effect [− 0.35; 0.35] correspond to a sample of 70 matched participants.

2.6 Randomisation

Two randomisation sequences (R studio, Random V 4.1.3 function) were edited:

  • Session 1: the order of sequences between real-life and virtual environments was randomised.

  • Session 2: the four arms were randomised, thus a different order of sequencing of the conditions, i.e. with avatar, with visual reinforcement or finally with the combination of both conditions.

A real-life test was carried out at the end of all virtual tests.

Allocation was based on the order of enrolment of the volunteers.

2.7 Blinding

This was an open-label study for all participants. Four evaluators were trained to administer the tests and instructions to participants. A Tymo therapy plate was used for data collection.

2.8 Statistical methods

The data from each condition are examined to ascertain normality and the types of tests to be employed, including parametric tests such as paired t-tests for pairwise comparisons and ANOVA for the four datasets, or non-parametric tests such as Wilcoxon tests for pairwise comparisons or Kruskal–Wallis tests for the four datasets.

The effect sizes and their confidence intervals are calculated using Cohen’s d (standardized mean difference, SMD) (Faraone 2008). The types of tests used will be specified in the tables.

Scores from session 1 were analysed by comparing the virtual and real conditions. A mean and standard deviation of the three tests was calculated as a reference value. The difference between the two data sets was measured. These calculations were consistent with objectives of measuring the difference between real-life and virtual conditions and setting reference values for both conditions.

Scores from session 2 were analysed in relation to the reference values obtained in session 1. The differences and effect sizes were calculated. The calculations aimed to measure the correction made by each condition.

The difference between the two real-life measurements of the two sessions (reality1 and reality2) was calculated and utilized to address the question of whether a learning effect is transferred from the virtual environment to real life.

The calculations were performed on r statistic: R version 4.0.5 (2021-03-31).

3 Results

3.1 Recruitment and flow of participants

For session 1, between 15 October 2020 and 21 January 2021, 70 subjects were recruited at Pôle St Helier (40 subjects) and at La Musse Hospital (30 subjects).

One participant stopped the tests due to knee pain prior to the testing; and a technical recording issue meant that no data was collected for one participant. Sixty-eight (68) data were analysed.

For session 2, 74 people were recruited at Pôle Saint Hélier between October and November 2022; a technical recording issue meant that no data was collected for five participants. Sixty-nine (69) data were analysed.

3.2 Initial data (Table 1)

Table 1 Baseline description
Full size table

In line with the objective, gender equity was observed in both sessions. No significant differences were observed between the different randomisation arms in either session 1 or session 2. Participants had a low level of sensitivity to motion sickness. MSSQ overall, session 1 mean (sd) = 5.25(4.98), session 2 mean (sd) = 5.71(5.09).

3.3 Primary outcome measures and estimations (Table 2)

Table 2 Comparison of center of pressure distance (Dist: cm); mean (sd) of 3 trials, between real condition and VR condition
Full size table

A statistical analysis of data normality using the Shapiro test reveals that no data can be accepted as following a normal distribution. Consequently, no-parametric tests (Wilcoxon signed-rank and Kruskal–Wallis) were utilised for p-value calculations.

3.3.1 Session 1

Three tests were conducted for each condition (virtual and real) to establish reference values. Regarding COP distance, the comparison of distance scores (Table 2) revealed a significant difference between left static, right static, right dynamic (SMD = − 0.40 [− 0.73, − 0.06], p = 0.02; SMD = − 0.40 [− 0.73, − 0.06], p = 0.05; SMD = − 0.61 [− 0.95, − 0.27], p < 0.001) and a non-significant score in left dynamic (SMD = -0.22 [− 0.56, 0.11], p = 0.19). There was no statistically significant difference among the three tests. The dynamic tests induced instability, resulting in higher scores compared to the static balance scores (p-value = 9.917e−07). The data indicated that static balance was compromised in virtual reality and even more so in dynamic balance. The standardized effect sizes ranged from a low of − 0.22 [− 0.56, 0.11] to a high of − 0.61 [− 0.95, − 0.27]. Expressed as a percentage, virtual reality caused a mean impairment of 13.12% (SD = 16.45) in static balance and 22.99% (SD = 45.10) in dynamic balance.

3.3.2 Session 2

All three conditions (full-body avatar, visual information or a combination of both) have a positive influence on balance by decreasing the distance of travel of the centre of pressure (Table 3). The effect sizes show that these corrections allow the recovery of balance postures that are almost identical to the real-life situation, with the exception of the dynamic balance avatar with right single-leg support.

Table 3 Comparison of COP distance covered data between 3 VR conditions (full body avatar, visual information, both) with real data: effect size
Full size table

On average, impairment is reduced by the avatar to 7% in static and dynamic, by the visual information to 6% in static and − 2% dynamic, and by the combination of the first two in 6% static and 4% dynamic (Fig. 3).

Fig. 3
figure 3

Comparison of conditions left and right combined. A static balance, B dynamic balance; virtual = virtual condition only, avatar = fulbody avatar implemented in virtual environnement, info = wallpaper with vertical cue in virtual environment, both = avatar + info, reel = balance in reality

Full size image

The comparison of the two real-life situations (Table 4) shows that when it is preceded by three virtual reality training sessions, performance tends to improve. The size of the learning effect ranges from the lowest in left static balance 0.09 [− 0.24, 0.43] to the highest in right static balance 0.59 [0.25, 0.93].

Table 4 Comparison of Center of Pressure (COP) Displacement Performance in Real-Life Between Session 1 and Session 2
Full size table

3.4 Secondary analyses

3.4.1 Influence of physical characteristics on the results

We fitted a linear model (estimated using OLS) to predict result with physical parameters.

The effect of age was statistically significant and positive for all results (p < 0.05). The greater the age, the greater the increase in instability.

The Welch Two Sample t-test testing the difference of Taille by Sexe (mean in group 1 = 177.95, mean in group 2 = 165.41) suggests that the effect is positive, statistically significant, and large (difference = 12.54, 95% CI [9.08, 16.00] p < 0.001).

No statistically significant effect on the MSSQ score in terms of static or dynamic balance performances were observed.

3.4.2 Comparison of male and female results in real-life conditions

For dynamic balance, the Welch Two Sample t-test testing the difference by Sexe suggests that the effect is for distance: positive, statistically significant, and very small (difference = 20.96, 95% CI [0.80, 41.13], p = 0.042).

There is no statistically difference for reel statique condition by sexe for distance of cop: small (difference = 7.21, 95% CI [− 4.77, 19.19], p = 0.234).

The difference between male and female scores can be explained by the average height of the two populations which affects the distance travelled by the centre of pressure. No difference was observed in the height to distance ratio between men and women in either static (p value = 0.36) or dynamic (p value = 0.41) conditions.

3.5 Risks and adverse events

No adverse events were recorded. The mean motion-sickness score from the Sickness Simulator Questionnaire (SSQ, scale from 0 to 48) after the virtual environment experience was low: mean (SD) = 2.21 (2.63). However, for six participants, the calibration of the avatar was a lengthy process (more than 5 calibrations to produce a re-adjusted avatar).

4 Discussion

4.1 Limitations

The technology used for the real-time immersive avatarisation had two weaknesses at the time of the experiment. Firstly, a lag with respect to real movement still persisted and we did not have an integrated mirror in the virtual environment which limited the strength of the embodiment process. It is probable that these results do not measure the true effects of the full-body avatar on postural responses.

4.2 General comments

Data from the first session provided sound reference values due to the three tests performed, i.e. 287 data sets for each test for a population aged from 19 to 61 years. Our study observed an increase in instability with age, suggesting that it cannot be applied to individuals over the age of 61 years. Virtual Reality (VR) scores were also enhanced by initially performing a real-life sequence.

4.3 Interpretation

4.3.1 Interpretation of CoP data

The CoP variable is a reliable measure of postural stability in healthy individuals (Corriveau et al. 2000). However, there is evidence that this variable can be influenced by various factors. Evidence has been reported of a systematic influence of the time of day on postural balance (Jørgensen 2014). Literature also reports the influence of age. In general, older adults have lower muscle strength and asymmetric torque resistance compared to younger subjects, which affects their ability to recover postural control (Pizzigalli et al. 2015). A logical influence of height on the measured scores is observed but we did not see any influence of weight unlike some authors (Chiari et al. 2002). However, it is not possible to identify postural strategies through these results such as hip strategies, which are more predominant in older people (Manchester et al. 1989). Likewise, we were unable to observe whether virtual reality involved a change in postural strategy.

Regarding this same criterion, in the scientific literature, an impairment in balance has been calculated at 9% in children after 60 min (Tychsen and Foeller 2018). Our data showed a much higher mean impairment (13%, static, 22% dynamic), which was more in line with the results of Lee (2019), who also reported that impairment increased with time of use up to 26% (Lee et al. 2019).

4.3.2 Impact of the avatar on instability

The correction provided by the avatar was below our expectations, which can be explained by several reasons.

Firstly, in Virtual Reality (VR), embodiment is the ability to visually replace a person’s real body with a life-sized virtual body, which can then be seen from that person’s perspective. In other words, when we put on a VR headset, a virtual body temporarily replaces our real body (Caserman et al. 2019). Several studies have shown that first-person embodiment enhances the sense of presence and leads to deeper immersion in the virtual environment compared to third-person embodiment (Neyret et al. 2020). The sense of embodiment does not differ between healthy volunteers and patients with conditions such as stroke (Borrego et al. 2019). The embodiment process can be triggered by a combination of visual-tactile information, as demonstrated by the rubber hand illusion experiment (Crucianelli 2023). Maintaining these tactile cues appears to contribute to the quality of immersion (Bovet et al. 2018) when stimuli are synchronous (Droit-Volet et al. 2020). The other option is to provide a combination of visual/motor information. Embodiment is based on the association between the visual perception of body movement and the internal sensation of that movement. The avatar’s body part moves synchronously with its real counterpart (Neyret et al. 2020). This is truly achievable only with the presence of a mirror in the virtual environment. We relied solely on visual-tactile information without a mirror, which seems insufficient to achieve a strong sense of immersion.

The second issue with the embodiment procedure is the duration of visual-motor stimulations needed for the brain to fully integrate the virtual body. By relying solely on visual-tactile association and not visual-motor, the embodiment time is undoubtedly longer. Several participants reported difficulties moving with their virtual body, which some perceived as discomfort.

Finally, our avatar is generic and lacks details on the face or finger movements. Focused on instability, it seemed sufficient for the experimentation addressing balance and overall upper limb movements. However, the presence of details is likely necessary to further enhance the sense of presence.

Therefore, it is likely that a full-body avatar with better quality and a visual-motor embodiment procedure may be able to correct instability associated with virtual reality. Further experiments are needed to confirm this point.

4.3.3 Learning in virtual reality

In session 1, the difference between real and virtual scores depended on the sequencing order. In all cases, scores tended to improve when the real or virtual condition was placed in second position; however, the difference increased significantly when the virtual condition preceded the real-life condition, particularly in the dynamic tests. The learning effect appears to be more powerful from real-life to virtual than from virtual to real-life.

In session 2, this learning effect was observed in the comparisons of the two real1 and real2 scores (Table 4). Real2 was always preceded by virtual reality tests. Performances were better in real2 compared to real1. The data are consistent with the literature which has shown learning effects on complex movements, particularly upper-limb movements (Bovet et al. 2018). Similarly, an effective transfer of skills occurred from real to virtual conditions but was more limited from virtual to real-life (Weber et al. 2022).

5 Conclusion

Immersive Virtual Reality (IVR) adversely affects both static and dynamic postural balance. However, this impairment can be mitigated, and in some cases completely rectified, through the incorporation of vertical visual cues within the virtual environment or by the introduction of a full-body avatar. Further studies are needed to investigate the effects of avatar on motor control given the rapid development of this technology. Developers can now offer options for scenery in the virtual environment that can help maintain balance. The use of a cursor could modulate the presence of these cues, meaning that they could be used for initial security but steps then taken to gradually remove these cues.