The eligibility criteria were formulated by using a modified patient population, intervention/indicator, comparator, outcome, and study design (PICOS) framework (Santos et al., 2007). The population comprised older adults of 50+ years of age, and the intervention/indicator was the assessment or training of balance using HMD, which was compared with traditional assessment or other training programs. By using exergames (e.g., Nintendo Wii or Xbox Kinect), participants can still receive contextual information from their surrounding and adjust their balance accordingly. Therefore, articles that used other VR technologies, such as exergames, for presenting the VR scenarios were excluded. The outcomes that we considered were balance control or the ability to maintain balance. We considered full-text articles with different designs that investigated balance assessment or training. While we considered conference papers (if full article was available) for inclusion, those that provided only the abstract (conference abstract) were excluded. We also excluded review articles.

The EBSCOhost, Scopus, Web of Science, and PubMed electronic databases were searched for articles and conference abstracts that focused on the use of HMD to assess and train balance and were published between 1990 until 1 September 2020. The following search terms and Boolean operators were used: (HMD OR head mount^* display OR virtual reality OR artificial environment OR simulated 3D environment OR simulated three-dimensional environment) AND (balance OR posture OR fall) AND (vestibular OR visual OR somatosensory OR context^*). Appendix 1 details the search strategy used. The search was restricted to English peer-reviewed articles. Each author independently screened the titles and abstracts of the identified records for relevance. The full text of the articles that appeared relevant and those without obvious relevance (after inspecting the title and abstract) were retrieved. The full-text records were read, and their eligibility was screened with the criteria mentioned above. The reference lists of all eligible studies were hand-searched to identify additional relevant studies. Any disagreement was resolved by mutual discussion.

One author (PS) extracted all data from each article into a piloted form. The data extracted comprised the type, number, age, and sex of the participants, HMD type, VR scenario, protocol, main outcomes, and the actual HMD effect. Due to the heterogeneity on the population characteristics, HMD interventions, and outcome assessment tools, a meta-analysis was not pursued.

The methodological quality of each article was assessed using a custom quality assessment questionnaire adapted from Downs and Black (1998) and Campos et al. (2011). These questionnaires provide scores for study quality, and internal and external validity, in randomized and non-randomized studies. The results from these questionnaires provide insights on the completeness of the reporting which helps future studies on commonly overlooked methodological steps, and how future studies should comprehensively report their methodology. Each article was evaluated on 13 questions that considered the reporting parameters, internal and external validity, and study power (Table 1). Each question was answered by “Yes,” “No,” or “Unable to determine.” As this review represents a qualitative summary of using HMD in assessing and training balance, no total score was computed for quality assessment. The risk of bias was appraised using the Risk of Bias Assessment tool for Non-randomized Studies (RoBANS; Kim et al., 2013). The RoBANS tool can be applied to different types of studies and eliminates the need to use several tools for each type of study. We opted to use solely the RoBANS tool to display all the risk of bias judgements in a single tool. The RoBANS tool includes the following domains: selection of the participants (selection bias), confounding variables (selection bias), measurement of exposure (performance bias), blinding of outcome assessments (detection bias), incomplete data outcome (attrition bias) and selective outcome reporting (reporting bias). For intervention studies, we also considered two additional domains to account for bias arising from the interventions, which included planning and implementation of interventions (performance bias) and deviation from intended interventions (performance bias). Each domain was judged as unclear, low risk, or high risk of bias. Any disagreements were discussed until consensus.

A total of 1,694 articles were found in the initial search. After duplicated and non-English removal, 385 articles were screened for titles and abstracts. After full text review, a total of 19 articles evaluating the use of VR HMD in balance assessment and training in older adults were included (Figure 1).

The results of the quality assessment are mentioned in Figure 2. Most of the studies properly described their aims, main outcomes, participants' characteristics, inclusion/exclusion criteria, and main findings (Q1 to Q5, respectively). For reporting the information (Q6 and Q7), most of the studies mentioned that their data were either normally distributed or provided standard error, standard deviation, confidence interval, and actual probability values. External validity was measured using Q8 and Q9. Since most of the studies were only using older adults without comparing them with a younger control group, the distribution of the main confounding factors was not the same in the study sample. Finally, internal validity was evaluated using Q10 to Q13. Very few studies were based on a predefined hypothesis or provided justification for their sample size, power description, variance, and effect estimates.

The judgement of each domain for each study, as well as the summary of each domain can be seen in Figure 3. Note that in the summary of domain risk of bias there is a high percentage (~20%) of “not applicable,” which is related to the non-intervention studies; this high percentage is due to the higher sample size of these studies that translated into greater weight in the summary plot. All studies showed high risk of detection bias because none of the studies blinded the outcome assessor. While selection bias due to uncontrolled confounding variables was a major concern (74% of studies), selection bias due to selection of participants had low risk of bias all but one study. Attrition and reporting bias were also not common, with only one study judged as high risk of bias attrition bias and none for selective reporting. Performance bias due to measurement of exposure was judged as high risk in 16% of the studies. From the 15 studies that included HMD as an intervention, only two studies were judged to having high risk of performance bias due to planning and implementation of interventions or deviations from intended interventions.

The included studies presented a wide variety of health conditions and age ranges (Table 2). Eleven studies involved healthy older adults (58%; n = 230), four studies included individuals with vestibular disorder (21%; n = 193), four studies included individuals with Parkinson's disease (PD; 21%; n = 67), one study included participants with mild cognitive impairment (MCI; 5%; n = 12), two studies involved participants with stroke (10%; n = 8), and five studies included balance impaired and at risk of fall individuals (26%; n = 127).

Most of the included studies used the Oculus Rift (DK 1 and 2) as HMD tool (33%). Other devices included HTC Vive (19%), Balance Rehabilitation Unit (BRU; 33%), and other systems (eMagin Z800 3D Visior, Sony Glasstron, Virtual Research VR8, Solo Myvu, Samsung Gear VR, and Revelation 3D; 43%). Twenty-nine percent of the studies did not report the name of the HMD. The VR scenarios were heterogenous across studies. Seven studies (37%) used VEs with adjustable visual perturbances while four studies (21%) used stationary visual scenes. Two studies (11%) used simple avatar VE and one study (5%) used (semi)realistic VEs. In 17 studies (89%), participants had to interact with VR to complete the tasks.

Sixteen studies (84%) used validated physical test batteries that utilized force plates, plantar pressure, posturography, physiological profile assessment, and human machine interaction. Two studies (11%) used specific perturbation-based protocols either by pulling and pushing, or by including visual perturbations inside the VR scenario. Six studies (32%) used questionnaires including Motion Sickness Susceptibility Questionnaire (MSSQ), Simulator Sickness Questionnaire (SSQ), Falls Efficacy Scale International (FES-I), and Activity-specific Balance Confidence (ABC). Four studies (19%) used surveys including gaming habits, Dynamic Gait Index, Berg Balance Scale (BBS), Tinetti Scale, Subjective Visual Vertical (SVV) test, Romberg, and Motion Sickness Rating. Actual HMD effect varied between studies. Studies that included healthy older adults, measured changes in gait patterns, postural stability, balance, fear of falling, body sway, simulator sickness, and balance strategies. Studies comprising patients with pathological conditions, evaluated both improvement and deterioration of balance and its related metrics.

The studies were heterogenous in terms of HMD features, VR scenarios, and comparison outcomes. Therefore, a direct comparison was not possible. Below, we report the results of HMD systems to assess balance in regard to validity, reliability, and consistency of their measurements with gold standard methodologies (Tables 3, 4).

The purpose of this systematic review was to explore the validity, reliability, safety, feasibility, and efficacy of HMD systems for assessing and training balance in older adults. The combined findings of the included studies show that VR HMD systems offer ecologically valid scenarios to assess and train functional balance. There is however a need to define standardized norms and protocols according to age, health status, and severity of disease. The studies also showed that several parameters of display type, VR scenario, and the duration of exposure can contribute to the safety and feasibility of HMD VR systems for balance training. These features could be adjusted according to participants' needs to ensure safety and efficacy of training. The use of HMD systems was effective in training balance and can be a useful tool to augment previously established interventions. Various visual scenarios can be added, removed, isolated, and manipulated to identify and treat specific balance-related impairments of different clinical conditions. The level of difficulty can also be adjusted to the patients' baseline levels to allow progress or regress when needed.

Only one study showed that HMD VR systems were valid and effective in healthy older adults. Those who trained with HMD VR systems showed improved gait parameters and lowered frequency of falls (Giotakos et al., 2007; Parijat et al., 2015; Kim et al., 2017). These improvements are probably due to the challenges that are induced when using VR scenarios. While some researchers reported that older adults have similar balance performance in traditional and VR scenarios (Alahmari et al., 2014), others have argued that participants' behaviors in VR versions of balance tests differ significantly (Muhla et al., 2020). One explanation could be that HMD systems remove the surrounding visual cues that older adults require for maintaining balance. Older adults will have to rely on VR scenarios, presented through HMD systems, which offer different visual stimulation compared to the real world. Another explanation is that HMD VR systems evoke various proactive and reactive postural adjustments (Liu et al., 2015; Parijat et al., 2015). It has been shown that participants generally walk slower, and their performance is lower in VR balance tests as compared to the traditional ones (Nyberg et al., 2006; Almajid et al., 2020). Participants also acquire more cautious behaviors when completing VR tasks in less secure VEs (Muhla et al., 2020) and take more time and steps which are adaptive strategies to ensure balance and avoid falling (Nyberg et al., 2006). Other strategies that healthy adults employed for increasing the stability in unfamiliar situations included a reduced forward displacement of whole-body CoM and lowered height of CoM (Thomas et al., 2016).

Motion sickness is believed to happen when the brain's assumption about sensory information does not match the actual received signals (Reason, 1978). This mismatch could be due to hardware and software limitations of VR HMD systems. The motion-to-photon latency, or the delay between the users' movements and their reflections on the display, can contribute to the motion sickness (Choi et al., 2018). The effects of latency are so important that even the smallest lag degrades the sense of balance compared to the naked eye with a similar FOV (Kawamura and Kijima, 2016). The interactions of gender and sickness history can also influence the risk of motion sickness and instability and should be considered when designing the HMD hardware and software (Kawamura and Kijima, 2016; Munafo et al., 2017). It has been shown that females are more susceptible to motion sickness (Munafo et al., 2017). In pathological conditions, such as PD, the sensory deficits could make the individuals less prone to the simulator sickness and could be less influential on their perceptions of balance (Kim et al., 2017). On the other hand, HMD related head movements pose sensory mismatch to the central nervous system of UVH patients, which later tries to resolve the mismatch (Clendaniel, 2010).

The duration of VR exposure should also be carefully monitored as it can provoke sickness symptoms. Previous research has associated verbal reports of cybersickness severity, as well as relatively high incidences of simulator sickness with the duration of HMD use (Treleaven et al., 2015; Dennison et al., 2016). HMD systems could also cause visual fatigue due to a disruption of the natural relationship between vergence and accommodation (Mon-Williams and Wann, 1998). Various physiological measurements have been used to estimate post-immersion cybersickness (Dennison et al., 2016). Those with greater levels of cybersickness may show less variations in postural sway (Dennison and D'Zmura, 2017). Therefore, the physiological measurements are important for differentiating between motion sickness and actual imbalance (Dennison et al., 2016).

While postural instability increases with time, it only happens when visual perturbations are present. It seems that elements of VR scenarios could also affect the way healthy older adults interact with the systems. Some researchers reported that increasing rotation speed in VR scenarios could increase cybersickness (Dennison and D'Zmura, 2017). Others have shown that speed manipulations in simple and complex VR scenarios do not cause motion sickness (De Keersmaecker et al., 2020). Adding an independent visual background to the scene could also reduce postural disturbance and could reduce the simulator sickness (Prothero et al., 1999; Duh, 2001).

As compared to younger adults, the older population is relatively more egocentric and more accurate when they are in the first-person perspective compared to the third-person perspective (Mattan et al., 2017). Research has also highlighted that a sense of actual presence in VR could be weakened if older participants view the VE from a third-person perspective (Lenggenhager et al., 2007). Therefore, perspective-taking capacity should be considered when designing virtual scenarios for older adults. Concerning the interaction with avatars, only one study included animal avatars (Cortés-Pérez et al., 2020). It seems that other contextual information around the avatars, including ground and objects, could act as visual reference points for the participants to adjust their balance. Naturalness of the movements, or the alignment between visual and proprioceptive senses, is also reduced when using HMD systems (Sander et al., 2006). Mismatch in proprioceptive and visual senses, or the inability to get information from body, as well as the low FOV and stability of the headset, could also reduce the naturalness of the movements. Additional motion tracking sensors, such as Leap Motion, can embed hands' motions directly into HMD systems, which can act as visual reference points for older adults (Scheggi et al., 2015). After a few trials, participants start to adapt to the gaming and protocol mechanics and adjust their walking to the VR and visual perturbations (Liu et al., 2015). As a result, VR might lose its ability to induce perturbations after 2–3 trials. It is thus important that HMD VR trainings use different scenarios and intensities so that they reduce the risk of adaptation to the VR environments and allow progression to optimize the results.

There is no standard protocol for defining postural instability in patients with different pathologies. Therefore, the results of different studies are conflicting regarding both assessment and training paradigms.

Fear of falling and perceived postural threat can affect gait patterns in older adults and could induce stiffing strategies in their joints. VR exposure therapy scenarios could increase self-efficacy beliefs of falling and provide a sense of control over falling (Giotakos et al., 2007). In the case of PD and in addition to imbalance issues, many patients have lower access to visual cues. They might have a fear of imbalance and decreased gait performance, preventing them from benefiting from rehabilitation. Patients with neurological conditions (PD and stroke) show improvement of balance and reduction of the risk of falls after VR training as compared to non-VR training (Caudron et al., 2014, Cortés-Pérez et al., 2020). Vestibular patients who are suffering from chronic balance impairments are likely to benefit from HMD VR training. Those who trained with HMD showed overall improvements in vestibulo-ocular reflex gain and in posturography parameters (Micarelli et al., 2017; Viziano et al., 2019). Vestibular patients without MCI have also shown improvements in otoneurological outcome measures after HMD VR training (Micarelli et al., 2019).

The HMD VR scenarios could also help in identifying different patients within certain pathology. For example, Chiarovano et al. (2018) could identify three types of vestibular patients: those with high dizziness scores who did not fail the balance test, those with low dizziness scores who failed the test, and other participants whom their dizziness scores were correlated with balance measurements. Visually dependent older adults with greater risks of fall show smaller acceleration ranges in VR TUG test compared to visually dependent older adults without any risks (Almajid et al., 2020). Patients with PD also show different CoP and balance functional reserve values in static and dynamic VR scenes (Suarez et al., 2011). The differences in the balance-related measures could be used to identify patients with higher balance impairments, to understand who will benefit the most from HMD VR training and guide users on how to adjust the training scenarios to the patients' specific needs. Nevertheless, there is a need for further studies to establish realiable cut-offs for each of these balance-related metrics and to adjust them to the associated pathological disorders. Then, the benefit of using VR is that sensory loads can be tailored and increased according to individuals' visual perception and threshold values to enhance balance performance (Proffitt et al., 2018). Notwithstanding the results of these studies showing the usability of HMDs to identify balance-impaired individuals, there are also conflicting results. Some researchers have shown that the CoP area and velocity were similar to healthy older adults during traditional and VR tests (Alahmari et al., 2014) and others have shown that CoP values were worse in vestibular patients compared to healthy older adults (Gazzola et al., 2019).

As patients get accustomed to the training, postural responses tend to adapt to the visual scenarios (Suárez et al., 2006). It has been suggested that different VR HMD systems could be more salient by improving stimulus presentation through encouraging participants to explore the VE (Menzies et al., 2016). Instructing participants to follow targets that move out of their direct FOV could also increase their presence. VR tasks should also challenge known sensorimotor deficits of individuals with special clinical and functional needs. For example, in patients with PD, more challenging walking tasks like turning, obstacle handling, or passing through doorways are impaired (Mirelman et al., 2011). Similar VR scenarios should be incorporated in the rehabilitation programs to increase their clinical relevance for their targeted population. Different types of virtual scenarios and frequencies of movement disturbance could also affect participants' stability (Jurkojc et al., 2017). It should be noted that participants get accustomed to these disturbances faster in open (seeing the horizon in 100 m distance) compared to closed sceneries (a room; Jurkojc et al., 2017). It is therefore important that patients are exposed to various and progressive training scenarios. The VR tests are easier to standardize than ordinary tests performed in gymnasiums or rehabilitation training facilities, which is a great advantage. Using VR systems, lighting conditions, room size, and other features can be set according to desired specifications. Each condition can also be manipulated individually to calculate their contribution in balance outcomes (Ferdous et al., 2018) and help to further understanding of complex mechanisms involved in balance.

Older adults may suffer from technological illiteracy. The reluctance of not using VR, stems from common beliefs about required equipment, financial costs, and unfamiliarity with the benefits of VR systems (Schwartzman et al., 2012). An insufficient knowledge of the systems both by the patients and their therapists could also hamper their acceptance. Therefore, VR training systems should have user-centric and user-friendly design. Developers should focus on improving the usability and accessibility of VR training systems, for promoting adherence and motivation. As participants' behaviors in VR differ from traditional environments, new norms should be defined to compare different participants. Different amounts of time spent in the VR rehabilitation programs, as well as motivation and other psychological parameters may also affect the outcomes and practitioners should consider them when interpreting the results.

Researchers, developers, and practitioners should also consider participants' gender and age, as well as sickness and medication history because one system and scenario may not work for everyone. Carefully planned visual scenarios with right frequencies (speed) and duration should be used to prevent visual fatigue and cybersickness. The timing and intensity of introducing new VR elements and scenarios should also be considered when developing training programs to ensure a rate of adaptation that allows continued progression and maximizes training efficacy. Future VR systems should be able to adjust these scenarios and their complexity, as well as sensory load, duration, and game algorithm for different participants. As technology is advancing rapidly, future HMD and processing systems will be faster, smaller, more powerful, and less obtrusive. Hopefully lower latency could be translated in more realistic interactions with VR HMD systems.

Almost all the studies display statistical limitations that limit external validity of their findings. Future studies should include more subjects, control confounding variables, and match for sex to enhance the generalizability of their findings. The effects of other features of VR, such as field of regard, display size, head-based rendering, stereoscopy, realism of lighting, and scene update latency, on balance should be further explored. Future research should also target the effects of biofeedback features, including audio and vibrotactile biofeedback, on postural control. It is also important to explore whether participating in VR trainings translates into functional improvements in real life. Despite the increasing amount of studies in this area, there is still much room to improve the current technologies, and the limitations identified in this review, should be considered when planning future studies.

There are a few study limitations to note. The low samples sizes and the high risk of confounding from most studies are the major concerns, and limit both the internal and external validity of findings. The aims and populations of the included studies were heterogenous, which hinder generalizing the results. Many clinical and balance-related measurements could be influenced by medication and severity of the disease. Inevitably, healthy and unhealthy older adults might have different SSQ or arousal levels because of the medication they were taking. Therefore, it is important to be mindful of such parameters when interpreting the results. Considering that some diseases are more prevalent in specific sex, sample size should be sex-matched to avoid selection bias. Some patients, such as vestibular patients, could also be less susceptible to any form of rehabilitation, possibly due the progressive physiological activity decline. Finally, some questionnaires are only validated in healthy younger adults and therefore should be validated in older adults and in individuals with specific diseases to establish the psychometric properties and thresholds before being used as reliable comparative measures. Unfortunately, there is still not enough evidence in the literature to provide a focused and critical review on the effects of HMD VR systems in a specific population or clinical condition. Still, this systematic review provides a comprehensive view on how HMD VR systems could be used to assess balance, and which are their potential benefits if used for training balance in older adults.