A literature search of the PubMed, PsycINFO, and Embase databases was conducted. The search string was a combination of serious games-related terms, such as “serious games,” “game-based,” “videogames,” “computer-assisted therapy,” “virtual reality intervention,” “gamification,” “gaming simulation,” and mental health-related terms, such as “mental health,” “depression,” “anxiety,” “problem drinking,” “schizophrenia,” and “obsessive-compulsive disorder.” Duplicate items were removed from the records that were identified through the literature search. The remaining items were screened on basis of title, abstract, and keywords by two independent raters (Ho Ming Lau and Ka Wai Ma). Items were included if the following inclusion criteria were met (a) the intervention used a digital game delivered on any technical platform including personal computers (PCs), consoles, cell phones, and handheld devices, which means that non-digital games were excluded; (b) the intervention targeted mental disorders such as those mentioned above; and (c) the study conducted an RCT. For the purpose of our study, simulations, VR interventions, and interactive programs without game elements were not considered serious games and were therefore excluded. The remaining records were assessed for eligibility by two independent raters (Ho Ming Lau and Jan Smit Jr.). Differences in ratings were resolved by discussion till a consensus was reached. If no consensus was reached, the coauthors of this paper were consulted to make the final decision.

A data extraction sheet was developed and pretested on two studies. The variables that were extracted from the articles were divided into two categories: (1) participant and study characteristics and (2) game characteristics.

Participant and study characteristics included variables, such as target group, recruitment, treatment type (single- or multicomponent intervention), primary outcome measures, how and how much guidance during intervention was given, setting of intervention, study conditions, attrition, and results.

Game characteristics comprised variables, such as title of the game used in the study, serious game type, game genre, and purpose of the game.

The validity of the included studies was assessed using the Risk of Bias Assessment tool of the Cochrane Collaboration (45). We used the following six criteria: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessors, incomplete outcome data, and selective outcome reporting. Each type of risk bias was rated low, high, or unclear.

Analysis of the data was done using the program Comprehensive Meta-Analysis (CMA) (46). Cohen’s d is an effect size that can be used to indicate the standardized difference between the two means divided by the pooled standard deviation at posttest. Hedges’ g, a variation of Cohen’s d, can be used to correct for potential bias as a result of small sample sizes. In this study, Hedges’ g was applied due to a number of studies with small sample sizes. An effect size of 0.2 indicates a small effect, 0.5 indicates a moderate effect, and 0.8 to infinity indicates a large effect (47). For studies that did not provide means or standard deviations, statistics such as F value or t value were used to calculate the effect sizes according to formulae of CMA. It should be noted that one study conducted two experiments (48). Both experiments were included as separate comparisons in the meta-analysis. The random effects model was used to calculate the mean effect sizes, as we expected heterogeneity among the included studies. It assumes that the calculated effect sizes differ because of true variation in effect size from one study to the next and not only of the random error within studies. The Q-statistic was calculated to assess the presence versus absence of heterogeneity, but only significance was reported. The I²-statistic was also calculated to test the homogeneity of effect sizes. A value of 0% indicates no observed heterogeneity, while larger values indicate increasing heterogeneity, with 25% as low, 50% as moderate, and 75% as high (49). Furthermore, we estimated 95% confidence intervals (CIs) around I² (50) using the non-central chi square-based approach within the Deducer GUI (51) of R (52).

The numbers needed to treat (NNTs) were calculated, using the formulae provided by Kraemer and Kupfer (53). By calculating the NNT, we obtained an estimation of the number of patients who need to be treated in order to have one who would benefit. The lower the NNT, the more effective the treatment is.

In studies where multiple primary outcomes were used, the effect sizes were averaged to produce a single summary effect size for use in the meta-analysis (54). Data were extracted by two independent reviewers (Ho Ming Lau and Jan Smit Jr.). Differences in extractions were resolved by consulting the coauthors.

Furthermore, two subgroup analyses were conducted according to the random effects model comparing (1) younger (≤18) versus older (>18) participants and (2) clinical versus non-clinical participants. We conducted additional subgroup analyses for the studies on the disorders, such as autism spectrum disorder (ASD), cognitive functioning, and post-traumatic stress disorder (PTSD) separately.

We examined the funnel plots visually in order to detect possible publication bias. A funnel plot is a simple scatterplot of the intervention effect against a measure of each study’s size. A symmetrical inverted appearance of the funnel plot indicates low publication bias, whereas an asymmetrical funnel plot indicates potential publication bias which may lead to an overestimation of the intervention effect (55). To verify an unbiased estimate of the pooled effect size, the Duval and Tweedie trim-and-fill analysis was performed (56). Moreover, Egger’s linear regression method was applied on the intercept to quantify the possible publication bias captured by the funnel plot and its significance.

The PubMed, PsycINFO, and Embase search returned 4,130 items. After removal of duplicates, 3,693 records remained. Initial screening of title and abstract excluded a further 3,612 articles, leaving 81 items. After applying the exclusion criteria, 10 studies were included in the review and 9 in the meta-analysis. See Figure 1 for a flow chart of the study selection.

We start by reviewing some characteristics of the serious games at stake, including their game design features per disorder (Table 1). We then present the results of our meta-analysis on the impact of these games on psychiatric-related disorders.

Figure 2 shows the results of the methodological assessment of the included studies in the meta-analysis. Five studies reported adequate random sequence generation (34, 59, 60, 62, 63). Two studies reported allocation to be concealed (59, 60). Two studies reported that both participants and personnel were blinded (59, 60). Three studies reported that research assistants were blinded for the outcome assessment (35, 59, 60). Four studies had a high likelihood of incomplete outcome data (35, 58, 60, 63). Several domains were rated unclear—the reason for this was incomplete reporting.

Visual inspection of the funnel plot indicated possible publication bias. Performance of the Duval and Tweedie trim-and-fill procedure also indicated possible publication bias. After adjustment for missing studies, the effect size for serious games interventions changed from g = 0.63 to 0.48 (95% CI 0.17–0.78; trimmed studies = 2). The Egger’s test did not indicate an asymmetrical funnel plot (P > 0.10).

The overall outcome of the nine studies (n = 10 comparisons) showed a moderately significant effect size of g = 0.55 (95% CI 0.28–0.83, P < 0.001) for improvement on mental disorder symptoms at posttest (see Figure 3). Heterogeneity was substantial and significant (I² = 58.53, 95% CI 1.31–34.15, P < 0.05) (Table 3). One study (60) had a control group that received TAU in contrast to the other studies (no treatment or waitlist). After excluding this study, a significant moderate almost similar effect size remained, g = 0.63 [95% CI 0.36–0.90, P < 0.001, NNT = 2.91] (Figure 4). Excluding the study with the highest (34) and lowest (63) effect size showed similar results (see Table 3).

We then conducted meta-analyses on psychiatric disorder-related symptoms for those where two or more studies were available (Table 3). This was the case for ASD (35, 63), cognitive functioning (58, 61), and PTSD (48, 64). The group that aimed at ASD-related symptoms such as lowered ability of recognition skills (35, 63) was shown to have a moderate non-significant (P > 0.05) effect size [g = 0.46 (95% CI 0.10–0.81, NNT = 3.91)]. The group targeting cognitive functioning-related symptoms such as lower attentional ability (58, 61) showed a large significant effect [g = 0.79 (95% CI 0.36–1.21, P < 0.05, NNT = 2.36)]. The group targeting PTSD-related symptoms such as flashbacks (48, 64) showed a moderate significant effect size [g = 0.59 (95% CI 0.20–0.99, P < 0.05, NNT = 3.09)].

We also conducted two subgroup analyses, grouping the studies by age and participant type. The group targeting youth (≤18) (34, 35, 59–61, 63) was shown to have a moderate effect size [g = 0.70 (95% CI 0.32–1.07, NNT = 2.63)] differing non-significantly from the adult group (18+) (48, 58, 64) [g = 0.53 (95% CI 0.07–0.99, NNT = 3.42)]. The group targeting clinical participants (with diagnosed mental disorder) (35, 59, 61, 63) showed a moderate effect size of [g = 0.59 (95% CI 0.18–0.99, NNT = 3.09)] differing non-significantly from the non-clinical group (34, 48, 58, 64) [g = 0.68 (95% CI 0.26–1.09, NNT = 2.70)].

This study aimed to give an overview of serious games for mental health-related symptoms that were evaluated with RCTs.

There were eight different games in our study. One of the games, SPARX, is both a goal-oriented and problem-solving game. The games Guardian Angel and JDTP are goal-oriented games. Five games (Tetris, Let’s Face It!, Braingame Brian, Lumosity, and “path-finding”) were categorized as cognition training games. In order to ascertain which game genre works best in targeting specific mental disorder symptoms more exploration is needed and studies need to be replicated.

All of the serious games were PC applications that required no internet connection to play. However, trends show that PC sales are declining, tablet sales are increasing, more time is spent on the smartphone and that most time spent on the smartphone is in playing games (in the US). It seems that the development and/or validation of serious games for mental health on this technical platform is lagging behind. As Fleming et al. (83) remarked, many mental health apps are already available. It is clear that considerable opportunities lie in this area. Accessibility and feasibility can be improved if this area is utilized. Software can be conveniently downloaded or distributed in app stores already available on smartphones.

We included 10 studies in the review that comprised 8 different serious games. The studies targeted depression-related symptoms, ASD, PTSD, ADHD, cognitive functioning, and AUD. The results of the meta-analysis showed a mean moderate effect size of these serious games for reducing psychiatric disorder-related symptoms. This finding is similar to that of a recently conducted meta-analysis of game-based interventions by Li et al. (36) for depression only. The effect sizes found in this study correspond to NNTs of approximately 3, indicating that three patients have to be treated in order to have one who would benefit, slightly lower than Li and colleagues. However, Li and colleagues did not confine their meta-analysis to serious games—they included simulation studies as well. The clinical impact for the youth group (≤18, g = 0.70, n = 5) is comparable to the adult group (g = 0.53), as in the Li et al.’s study (36).

These findings indicate that serious games for mental health-related symptoms have potential for various age groups. Our results compare favorably to those found in the review of serious games for depression only by Fleming et al. (83). This review was not confined to RCTs. Comparing our results with internet interventions that are not serious game based is more complex due to the differences in setting, study designs, and outcomes applied. A very generic comparison, with similar results could be made with the meta-analysis of Ebert et al. (27) on youth (2015) who likewise found evidence for the moderate efficacy of internet and computer-based CBT in the treatment of depression and anxiety (g = 0.76). Regarding adults, another meta-analysis on the impact of internet-based depression interventions (18 RCTs) of Cowpertwait and Clarke (84) likewise showed that web-based interventions targeting adult depression reduced symptoms moderately significantly compared to controls (g = 0.43). This study also showed that reminders and guidance are important moderators for treatment outcome.

The majority of studies focused on youth and young adults especially for ASD, ADHD, and depression (34, 35, 59–61, 63). Games for children with ASD-related symptoms showed comparable moderate effect as found by Grynszpan et al. (85) for technology-based (non-serious gaming) interventions for children with autism spectrum-related symptoms (d = 0.47). Studies using serious games targeting cognitive functioning were available for both adults and youth.

No serious games primarily targeting anxiety were found in this study. This finding may not be surprising, because this can be expected of a relatively new field of research. Another explanation could be that game elements do not add much more to VRET, which is found to have potential to combat anxiety (86–88).

The strength of this study is that it is the first to provide insight into the potential effectiveness of serious games on psychiatric disorder-related symptoms based on RCTs only and on a strict definition of what is regarded to be a “serious game.” This is important as RCTs are considered the gold standard of research to measure effectiveness of interventions and to make clear what a serious game is or is not. This study has some limitations as well. First, the number of studies that were eligible for inclusion in this review was small and the number of participants included in those studies likewise. The risk of bias in these studies was unclear in many cases due to incomplete reporting and may have been a problem (selection, performance, and detection bias) indicating that the methodological quality of studies could be improved and effect sizes could be overestimated. Furthermore, given the small number of studies, several mental health disorders and clinical outcomes were collated as psychiatric-related disorders instead of focusing on single ones. This means the findings can only indicate improvement instead of clear symptom reduction (89). Such an approach is, however, not uncommon in an evolving new domain of academic endeavor [see, e.g., Ref. (89)]. As a consequence, we decided not to conduct subgroup analyses of distinct mental disorders as the number of studies and power for doing so was too low.

Using serious games as interventions for reducing mental health problems appears feasible. Due to the limited number of RCTs that we have been able to include in this analysis, this review can only give an idea of the potential of serious games for treating mental disorders in the future. More RCTs are needed to determine the effectiveness of serious games. Future studies should not lose the technological development out of sight. Smartphone-based serious games for mental health need more exploration. Further, the effect and use of serious games for mental health that let players connect with other players using an internet connection need to be investigated.