The Methods section discusses the approaches used to obtain specific publications as well as the constraints and eligibility requirements that were used [21]. Following that, an overview of the standards used to interpret the studies and associated variables is provided. This meta-analysis was carried out in compliance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [22]. A checklist of all PRISMA items is presented in Multimedia Appendix 1.

The meta-analysis included studies that (1) were published between 2013 and April 2021, (2) were written in English, and (3) primarily focused on AR for learning and training. Further, we excluded studies that (1) used other virtual platforms, such as VR and virtual patients, and (2) had no experimental data. However, based on the PICO (Population, Intervention, Comparison, Outcomes) framework, our final analysis incorporated various trials, including those with randomized controlled and mixed designs. Textbox 1 illustrates all of the PICO elements used for the inclusion criteria.

We conducted a web-based literature search in the Cochrane Library, PubMed, Embase, Web of Science and Google Scholar databases from February to April 2021. We also manually checked the reference lists of the eligible studies to identify and acquire additional relevant publications. We used the following set of keywords, terms, and logical operators: “augmented reality” OR “AR” OR “mixed reality” OR “patient simulation” AND “medical training” OR “medical education” OR “health professions training.” Some of the search strategies and terms used are presented in Multimedia Appendix 2. It is worth mentioning that the search strategies were slightly modified to suit each web-based database. EndNote X9 (Clarivate) software was used to manage and import the selected documents and to remove duplicates. This process was carried out by 2 authors (YB and GA) of this study.

Data were placed into an extraction spreadsheet by using Microsoft Excel 2019. From each of the studies, information was extracted based on the study characteristics, designs, and participants. These data included the names of authors, the years of the studies, the locations of the studies, sample sizes, participant types, study designs, and interventions and outcomes. EndNote X9 was also used to obtain some publication data, such as titles, publishers, URLs, digital object identifiers, page numbers, issue numbers, and volume numbers. The process of data extraction was initially performed by 2 authors (HA and A Alwadain) and was then validated by another 2 authors (WNWA and LFC).

The quality of the studies was assessed by 2 authors (AB and A Alghail). We followed the criteria of the Cochrane Handbook for Systematic Reviews of Interventions [23] in the risk of bias assessment process. These include 7 domains that correspond to a particular kind of bias. Each domain was given a score (a “high risk,” “low risk,” or “unclear risk” of bias), and any disagreements among coauthors were settled via consensus.

RevMan 5.4 (The Cochrane Collaboration) software was used to perform the risk of bias assessment; different scales were used by different authors to measure the outcomes; therefore, raw data could not be compared directly. Thus, the R 4.0.2 (R Foundation for Statistical Computing) software was used to transform raw data, so that STATA 14.0 (StataCorp LLC) could be used to run the analysis directly. The metacont package on R was used to transform the data on the sample sizes, SDs, and means of the control and AR groups. Data on the sample sizes, means, and SDs of only the AR groups were transformed by using the metamean package. Only 1 study [24] recorded the sample size and the proportion of participants on whom AR had a positive effect. The effect sizes and CIs were generated by using the two different packages. STATA 14.0 was used to generate the forest plot for each outcome, which used the transformed data (effect size and CI), via the metan command. The statistical heterogeneity among the selected studies was measured by using I² in each analysis and a 5% significance level. The random effect model was selected for the analysis carried out because the true effect sizes underlying all studies were stochastic.

From our initial search of 5 potential databases, we found a total of 1832 records. We identified a total of 1839 records, including relevant reference lists (n=7). Duplicate records (973/1839, 52.9%) were removed. Following that, a total of 866 of the 1839 (47.1%) studies were screened based on their titles, abstracts, and keywords, and 798 were excluded after applying the eligibility criteria. A total of 68 studies were downloaded and evaluated via full-text screening, of which 55 were excluded for various reasons (Figure 1). Finally, 13 studies with a total of 654 participants were eligible for the meta-analysis; 2 studies used the same set of participants for both the control and AR groups [25,26], while another 6 studies [27-32] that also measured both the control and AR groups divided the sample size.

As presented in Table 1, the trials of the selected studies were performed in the following ten countries: Germany [27,28,30], the United States [24,31], the United Kingdom [33], Canada [34], Italy [35], Sweden [26], Finland [32], Switzerland [29], Spain [36], and South Korea [25]. Furthermore, 10 of the 13 studies, adopted a randomized control trial [24,26-29,32-36], 2 studies had a mixed design [30,31], and 1 study used a cohort approach [25]. The number of participants in the studies ranged between 4 and 372. With regard to these participants, 5 trials recruited medical and nursing students [25,27,28,30,31]; 2 trials involved pediatric and psychiatry residents [26,29]; 1 trial recruited visitors to a center [32]; and the remaining trials (n=5) included multiple participants, such as physicians, registered nurses, technicians, residents, students, clinicians, clients, hosts, and paramedics [24,33-36].

In recent years, the use of AR in the medical field has attracted the attention of academics and researchers, as evidenced by the increasing number of publications devoted to AR. This is demonstrated by the fact that 8 of the 13 studies were conducted between 2019 and 2021. The characteristics of the selected studies, including the outcome measures, interventions, and types of participants, are shown in Table 1. Additional details regarding the AR and control groups are presented in Multimedia Appendix 3.

Figure 2 and Figure 3 present the risk of bias assessment, which was performed according to the Cochrane criteria. Of the 13 studies, 10 detailed randomized methods [24,26-28,30-33,35,36], while the rest of the studies (n=3) [25,29,34] reported no sequence generation methods. Only 3 trials reported concealment methods [25,32,34]. Furthermore, 3 studies reported dropouts [29,30,33] but did not go into detail about how they were handled. Due to the uniqueness of the intervention methods, no blinding methods were reported for participants in any of the trials. The blinding of assessors was used in 5 trials [26,27,29,35,36].

The most significant issue with meta-analyses is the possibility of bias and error arising from combining multiple studies, especially unrelated studies. Creating a funnel plot is one of the key methods used for testing common publication bias. The graphics in the funnel are obvious scatter plots for the effect size approximated in each selected study versus the unit of sample size reported in the studies [37]. If there is no bias, the graphic resembles an upside-down funnel. However, if a publication bias exists, the graphic should be asymmetric and distorted [38]. Figure 4 shows the funnel plot that we used to test for publication bias.

As can be seen in Figure 4, the 13 studies included in the meta-analysis are symmetrically placed on both sides of the vertical effect size line and are very close to the symmetrical and merged effect size lines. If there is no publication bias, studies should be distributed symmetrically on both sides of the vertical line depicting the merged effect size. If the studies are placed outside of the pyramid, they should be concentrated in the upper or middle parts of the plot. However, if there is a bias, most of the studies will be in the funnel plot's bottom part or in only 1 vertical line segment [39]. The effect sizes of the included studies are distributed in a symmetrical manner, with only 5 studies located outside of the pyramid. However, it is worth nothing that 4 of these studies are located in the upper part of the pyramid. This pattern shows that the studies included in this work have no publication bias.

As shown in Figure 5—the meta-regression plot—the true effect sizes between the control and AR groups are 1.8 for the 2013 study [27]; 0.9 for the 2015 study [26]; 2.2, 0.2, and −0.9 for the 2017 studies [28,29,34], respectively; 1.2 and 0.9 for the 2019 studies [24,30], respectively; 0.4, −1, 0.2, and 1.5 for the 2020 studies [31-33,35], respectively; and finally, −3.1 and 0.8 for the 2021 studies [25,36], respectively.

A meta-analysis was performed by using the data set in Multimedia Appendix 4 for each of the five outcomes (ie, knowledge [Multimedia Appendix 5], performance time [Multimedia Appendix 6], satisfaction [Multimedia Appendix 7], confidence [Multimedia Appendix 8], and skills [Multimedia Appendix 9]). The coding of the final analysis is also provided in Multimedia Appendix 10.

The usefulness of AR simulation methods in medical training was evaluated in this meta-analysis. In terms of skill outcomes, AR training did not outperform other educational methods for medical training. A study on the use of AR in medical education found that students’ skills improved after the intervention [25]. In terms of skill performance scores and success rates, it was found that AR groups did not exhibit any statistical relationships, even though each of the trials that reported skills (ie, those included in our study) used different education and training methods for their control groups.

There was a statistical difference (P=.02) between the AR and control groups in terms of confidence. Participants' confidence was improved by AR when compared to that of participants in control conditions. It was found that AR training had a greater effect on participants’ confidence.

The I² values reported throughout the Results section are very high, indicating that there was a huge level of heterogeneity between studies. In terms of knowledge outcomes, participants who underwent AR training did not outperform those who used other educational methods for medical training. However, a previous study that looked at how AR affects learning found that AR was more effective for knowledge retention [27] and that students are more likely to make connections between concepts when they are in an interactive learning environment. As a result, more research into the impact of AR on knowledge in medical training is needed in the future.

Between the control and AR groups, there was a significant difference in participants’ satisfaction (P=.006). Of the 7 studies that recorded satisfaction as an outcome, 4 [25,32-34] indicated partial satisfaction. Some participants remarked on how difficult it is for AR technology to affect user satisfaction. As a result, we believe that participants’ satisfaction with AR training may vary depending on technical factors. AR technology will be better able to satisfy users as the technology advances.

A meta-analysis of performance time was also carried out. The findings suggested that AR was more effective than other training methods at reducing performance time. However, we found that AR was more efficient than other methods in enhancing performance time. The observed heterogeneity could have been due to the various research designs and settings used in the selected studies, such as surgical projects, AR devices, and control group training methods. One study [36] on the efficiency of AR endoscopy simulation training analyzed performance time based on real-world data and found no significant difference between the AR and control groups, but the evidence quality was poor. Other studies that measured performance time as an outcome found that AR can assist operators in reducing performance time [24,25].

One of the main strengths of this work is being among the first meta-analyses to be conducted on the effectiveness of AR in medical training, which presents a significant contribution to the advancement of AR in the medical field. Furthermore, our work involved an assessment of 13 trials with a total of 654 participants from various countries.

Our work has some limitations as well. The first is that we only included articles that focused on AR as an intervention, which made it difficult to find a large number of studies. Second, some of the studies included in the review omitted information about allocation concealment, sequence generation, and blinding methods. Finally, the 13 trials we evaluated used different teaching and training methods for their control groups, which could lead to significant heterogeneity.

The effectiveness of AR training methods in the medical field was assessed in this work. The meta-analysis included 13 trials with a total of 654 participants that were completed between 2013 and 2021. In all trials, AR training was used as an intervention in AR groups, while conventional methods were used for control groups. Based on our findings, medical students' performance time, confidence, and satisfaction can be improved by using AR training and education methods. However, there was no statistical difference between the skills and knowledge that participants gained by undergoing AR training and those gained via conventional training methods. The use of AR should therefore be adopted in medical department training because of its significant effect on the performance time of participants. In general, AR should be used to expand knowledge and as a supplement to other simulation approaches to enhance clinical practice quality and safety. It is very noticeable that there are considerable gaps in the literature regarding the use of AR technology in medical training, and a limited number of studies exist, suggesting that further research efforts in this area are needed.