This systematic review conforms to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement [35] and was registered in advance in the international Prospective Register of Systematic Reviews (PROSPERO) database (registration number CRD42021250727).

The literature search was conducted on PubMed, Web of Science, EMBASE, CINAHL with Full Text, the Cochrane Central Register of Controlled Trials, CNKI, WanFang, VIP, and CBM, from inception to May 25, 2021. The search terms were chosen to be inclusive of VR (eg, “virtual reality”, “VR”, “virtual environment”) and breast cancer (eg, “breast neoplasm” OR “breast tumors”). Medical Subject Headings (MeSH) and Embase Subject Headings terms were used. See Multimedia Appendix 1 for the specific search strategy adapted for each database. Details of the search strings of the PubMed database are displayed in Table 1. Searches were limited to English and Chinese language sources.

The Population, Intervention, Comparison, Outcomes, and Study design (PICOS) model was used to establish the article inclusion criteria:

Studies were excluded if they (1) did not specify the type of cancer; (2) described the technologies only; (3) were conference papers, workshop papers, literature reviews, posters, comments, letters, study protocols, or proceedings papers.

Records from searches were imported into an EndNote library (EndNote X9.1) and duplicate studies were removed. The remaining records were transferred into an Excel spreadsheet (Microsoft). Screening was conducted by 2 independent reviewers (XB and WX) who assessed the article titles, abstracts, and full texts. Articles that did not meet the established inclusion criteria were excluded. Any disagreements between the 2 reviewers were resolved by discussion or in consultation with other investigators (QC, AC, and XL).

Data extraction was performed independently by 2 reviewers (XB and WX) using a predesigned standardized form in Word (Microsoft). Any discrepancies between the 2 reviewers were resolved by discussion with other reviewers (QC, AC, and XL), who acted as arbiters where necessary. We removed duplicate data published in different manuscripts. Additionally, the authors of the included trials were contacted to obtain any unclear or missing data. Data extraction included study characteristics (the first author, study design, and study region), participant characteristics (sample size, age), intervention details (characteristics of interventions, duration), patient-important outcomes, measuring instrument, and main results.

Two reviewers (XB and WX) independently assessed the methodological quality of all included trials. The Cochrane risk-of-bias tool was used to assess the quality of included RCTs [36]. The Cochrane risk-of-bias tool includes 6 domains of bias: selection bias, performance bias, detection bias, attrition bias, reporting bias, and other bias. After assessing the risk of bias of each study, the studies were categorized as “low risk,” “high risk,” or “unclear risk.” The Risk Of Bias In Non-randomized Studies of Interventions (ROBINS-I) was used to assess the quality of the included non-RCTs, covering 7 distinct domains: bias due to confounding, selection bias, bias in measurement classification of interventions, bias due to deviations from intended interventions, bias due to missing data, bias in the measurement of outcomes, and bias in the selection of the reported result [37]. The ROBINS-I is a new tool for evaluating the risk of bias in estimates of the comparative effectiveness (harm or benefit) of interventions from studies that do not use randomization to allocate units (individuals or clusters of individuals) to comparison groups [37]. After assessing the risk of bias of each study, the studies were categorized as “low risk of bias,” “moderate risk of bias,” “serious risk of bias,” “critical risk of bias,” or “no information.” In case of doubt, the final decision was determined through discussion or consultation with other reviewers (QC, AC, and XL).

The meta-analysis compared VR interventions with other interventions or nonintervention. The studies were divided into subgroups based on the measuring instrument that was used in the study. If more than 1 instrument was used in the same study, we included the study in more than 1 subgroup. The differences in the effect size between the groups were analyzed in terms of the standardized mean difference (SMD). Review Manager version 5.3 (Cochrane Collaboration) was used to conduct a meta-analysis. The mean difference (MD) and SDs with 95% CIs were used to calculate continuous variables. Initially, a fixed-effect model was used in the data analysis. An I² value over 0.5 was considered to represent substantial heterogeneity and a random-effect model was used [38]. Subgroup analyses were not possible due to the lack of patient-level data. All P values were 2 sided.

A total of 964 potentially relevant articles were initially identified from the 9 databases; 271 articles were removed due to duplication, and the remaining 693 studies were screened. We excluded 664 articles due to insufficient relevance based on the title and abstract. The characteristics of the excluded studies are shown in the PRISMA diagram (Figure 1). Twelve studies were included in the systematic review, 10 of which were further included in the meta-analysis.

The characteristics of the 12 studies are shown in Table 2. Three studies were from the United States, 3 from China, and 1 each from Turkey, Italy, Egypt, France, Australia, and Jordan. The included studies were published between 2003 and 2021 [14-17,23,30,39-44]. Of the 12 studies, 6 were RCTs [15,16,23,41-43], 2 were quasi-experimental design studies [39,40], 1 was an externally controlled trial [14], and 3 [17,30,44] were pre–posttest study designs with a single arm. The number of participants ranged from 6 to 80. All participants were adult patients with breast cancer or BCSs. VR-based interventions included both immersive and nonimmersive formats. The intervention duration varied from 15 minutes to 10 months. All studies examined the effects of VR-based interventions on health-related outcomes, including shoulder ROM, hand grip strength, anxiety, depression, pain reduction, cognitive function, fatigue, incidence of complications, cybersickness symptoms, and fear of movement.

The results of the assessment of risk of bias are presented in Figures 2 and 3 and Table 3. The Cochrane risk of bias tool was used to assess the quality of the included RCTs. For RCTs, allocation concealment and blinding will not seriously influence the selection of patients and the measurement of outcomes. Two of the 6 RCTs reported randomized methods in detail [16,22], while the remaining 4 trials did not provide the methods of sequence generation, nor demonstrated that the participants were recruited randomly. None of the trials provided concealment methods, except 1 trial that reported the use of anonymization by placing numbers into opaque, sealed envelopes to conceal the allocation sequence [22]. In all trials, no blind method was used on participants due to the particularity of the intervention methods. None of the trials reported employing blinding of assessors, except for 1 trial that reported the person who collected the data [16]. Only 3 studies performed power calculations and reported adequate statistical power [16,22,41], while the other studies did not perform power calculations and were without dropouts [15,42,43].

ROBINS-I was used to assess the risk of non-RCTs. Four studies [14,17,40,44] had a moderate risk of bias, with confounding, outcome measurement, and selective reporting being the primary sources. Two studies [30,39] had a critical risk of bias due to missing data. The risk of selection bias was judged to be low for all studies. A detailed list of the risk of bias assessments is provided in Table 3.

This research aimed to use qualitative and quantitative methods to evaluate the effectiveness of VR-based interventions in the rehabilitation management of patients with breast cancer. Twelve studies were included in the systematic review, 10 of which were included in the meta-analysis. A total of 604 participants were involved in different studies. In view of our results, we can conclude that VR-based interventions are more effective in improving the emotional, cognitive, and physical well-being of BCSs. For other outcomes (and comparators), the evidence was less compelling in improving learners’ skills, attitudes, satisfaction, and patient-related outcomes.

For RCTs, due to the nature of the intervention, we judged the majority of studies to be at a high risk of bias for nonblinding of participants and study personnel and nonblinding of outcome assessment. Although blinding of outcome measurement would not seriously influence the results, nonblinding of participants might bias the effect [45]. For non-RCTs, most of the studies had a moderate risk of bias in confounding, outcome measurement, and selective reporting. Some studies had a serious or critical risk of bias, most frequently due to the outcome measurement, missing data, and choice of analyses, which did not allow controlling for missing data.

Moreover, the lack of randomization and power analysis to calculate the appropriate sample size, as well as different duration, schedule, intensity, and type of interventions, and different scales of measurements all contributed to the heterogeneity of the included studies. Additionally, some outcomes could not be analyzed quantitatively due to data format (eg, incidence of complications, cybersickness symptoms) or the fact that the data were only reported in 1 study (eg, fatigue, fear of movement). Furthermore, we could not assess the risk of publication bias because funnel plot–based methods are not accurate for less than 10 included studies per outcome.

We identified 12 studies; however, most of them were limited by small sample sizes. Therefore, additional studies are needed to confirm our findings. We are encouraged that a number of larger RCTs are currently underway [46-52], which are likely to inform the field further.

Our findings must be viewed with caution owing to the limited number of trials with low quality. Moreover, as we only included a small number of trials, it was not feasible to conduct subanalyses regarding VR-based interventions or study design. A more empirical study is needed to determine the applicability of VR-based interventions in BCSs according to intense physical and psychological symptoms, function defects, and adverse effects. Additionally, further study is needed to standardize the contents of VR-based interventions, especially for upper limb recovery. Moreover, to successfully implement VR-based rehabilitation exercises into daily practice, it is better to provide detailed information on training frequency, duration of the intervention, and targeted motor skills [53].

Although we performed extensive searches of the literature, there is a possibility that we did not identify all relevant studies. Two review authors independently completed data screening, extraction, evaluation of risk of bias, and certainty of evidence rating. Any discrepancies between the reviewers were resolved by discussion or in consultation with other reviewers in the event that disagreement persisted. Even though we contacted all relevant study authors for additional information, we did not always receive a response. Low study quality, inadequate methodological details, and significant inconsistencies across trials decrease the overall quality of the evidence. Moreover, the variability in VR-based interventions, as well as the timing, regimens, and definitions of outcome measurements all have the potential to contribute to inaccuracies in the assessment of the intervention effects.

The findings of this systematic review and meta-analysis indicated that VR-based interventions have a positive effect on physical and psychological symptom management and ROM. Our findings are a valuable extension of recently published systematic reviews and meta-analyses. Previous similar studies mainly focused on the effect of immersive or nonimmersive VR-based interventions in cancer survivors or stroke survivors; however, the results of the studies were inconsistent.

Ahmad et al [54] and Chow et al [55] reported that VR-based interventions may be more effective in the management of pain and anxiety in patients with cancer, whereas a nonsignificant difference in pain and anxiety was reported by Zeng et al [56]. By contrast, Ioannou et al [57] reported that VR-based interventions demonstrated a trend toward improvement in pain. One possible explanation for this difference is that the intervention effects of immersive and nonimmersive systems differ and that these reviews focused on cancer survivors. Moreover, different cancer survivors may have various physical and psychological symptoms that lead to variability within the findings. Therefore, it is necessary to conduct controlled trials between different interventions and populations. In addition, although no significant improvement was observed in most studies, we cannot ignore the potential health-promoting effects of VR-based interventions. In agreement with our review, Aminov et al [58] found evidence of a significant effect in improving upper limb function using VR-based rehabilitation interventions and suggest VR as an adjunct for stroke rehabilitation.

Overall, previous reviews have presented similar conclusions to those of our review, suggesting that although the evidence is limited, it does exist. However, the results of this review should be interpreted with caution due to the limited number of controlled trials analyzed, the small sample sizes, and low methodological quality. The majority of previous reviews/meta-analyses indicate that more high certainty of evidence is needed before VR-based interventions can be considered as potential strategies for rehabilitation management in BCSs consistently.

The examination of VR-based interventions is recommended to ascertain whether there is a role for technology‐based exercise in improving the late and long-term side effects of breast cancer treatment. Furthermore, empirical evidence is required to provide well‐substantiated recommendations regarding the frequency, duration, and content of the VR intervention. Finally, future studies on VR-based interventions could utilize more consistent reference standards, such as standardizing the frequency, duration, and content of the VR interventions. Such standardization minimizes bias and heterogeneity between studies. Future studies could focus on (1) the late and long-term side effects of breast cancer management; (2) the mechanism of symptom management; and (3) combination of VR with artificial intelligence, physiological indexes, and electroencephalogram.

The late and long-term side effects resulting from breast cancer treatment are persistent and prominent. The findings from this review suggest that VR has the potential to facilitate immediate and longer-term improvements in symptom management and the performance of upper extremity function following a surgery for BCSs. Although the use of VR-based interventions has expanded in the rehabilitation management of BCSs, the current evidence for using VR-based interventions for both immediate and long-term improvements among BCSs remains limited. Future trials would benefit from using multicenter data, with larger sample sizes, longer follow-up periods, and high methodological quality.