Life expectancy has increased worldwide as people have better access to health care services and an improved standard of living. As a result, people are living longer [1-3]. According to the United Nations World Population Aging 2020 report [4], the number of people aged ≥65 years has increased up to 727 million worldwide. The older population group is expected to increase to 16% by 2050 compared with 9.3% in 2020 [4]. The older population group is more likely to develop cognitive impairment [5,6], which is a decline in cognitive abilities and functions such as memory, attention, concentration, learning, and language [7,8]. According to the Alzheimer’s Association, approximately 12% to 18% of people aged ≥60 years have mild cognitive impairment (MCI) [9].

MCI refers to a decline in the ability to learn new information or recall stored information and occurs along a continuum that ranges from normal to severely impaired cognition [10]. Although inconsistencies exist in screening for MCIs, it is certain that they occur because of brain changes owing to multiple factors, including older age, injuries to the brain, diabetes, hypertension, stroke, depression, and physical inactivity [11]. Memory is one of the main cognitive functions that decline with age. Memory is known as the ability of the brain to hold information and recall it as needed. There are different types of memory: verbal, nonverbal, and working memory. Verbal memory refers to a person’s ability to remember what they read or hear of information that was already learned [12]. On the other hand, nonverbal memory refers to storing, retrieving, and remembering nonverbal information, content, or experiences, such as images, feelings, tastes, sounds, shapes, and smells [13]. Furthermore, memory is divided into 3 types according to the period for which the memorized information is retained: short-term, long-term, and working memory. Short-term memory temporarily holds a limited amount of information [14], whereas long-term memory refers to the relatively permanent storage and recall of information [15]. Working memory refers to the temporary storage of a limited amount of information to be used in the execution of cognitive activities such as learning, reasoning, and comprehension [16].

Several nonpharmacological interventions can be used to improve memory, such as physical exercise, cognitive behavioral therapy, psychosocial therapy, good nutrition, and serious games [17]. Serious games are defined as electronic games that are played for purposes beyond leisure to promote the users’ mental, physical, and social well-being [18,19]. Recent evidence suggests that exergames are effective in improving physical and cognitive function in people with MCIs [20] as well as their compliance and adherence to medical interventions embedded in serious games [21,22]. Previous systematic reviews have shown that serious games have the potential to prevent or alleviate mental disorders such as depression [23], anxiety [24], and cognitive impairment [25]. Several types of serious games have been used to improve cognitive abilities, namely (1) cognitive training games (which deliver cognitive activities to maintain or improve cognitive functions) and (2) exergames (which entail physical exercises as part of the intended gameplay [25]). Compared with conventional exercise and cognitive training, serious games can positively affect mood, social functioning, mental health well-being, and cognitive flexibility in older adults [26-29].

The effectiveness of serious games in improving memory has been assessed by many studies. To draw definitive conclusions about the effectiveness of serious games, the findings of these studies need to be pooled and aggregated. Several systematic reviews have summarized the evidence from these studies; however, they had a different aim and scope from this review. Specifically, these reviews (1) focused on healthy older adults and not necessarily those with cognitive impairment [17,30-33] (therefore, future reviews should consider older adults with cognitive impairment), (2) included pilot randomized controlled trials (RCTs) and quasi-experiments [17,20,33,34] (thus, future reviews should include only RCTs), (3) performed an outdated search (>5 years [17,32,34]; therefore, an updated review or a new review are required), (4) did not assess the quality of evidence [17,20,30,33,34] (thus, the quality of the evidence should be assessed in future reviews), (5) only focused on a specific type of serious game such as cognitive training games [30,34] and exergames [17,20,33] (hence, future reviews should consider all types of serious games), (6) focused on a certain type of memory (working memory [34]; therefore, all types of memory should be considered in upcoming reviews), or (7) did not compare the effect of serious games with a specific comparator (eg, no intervention, conventional exercises, or conventional cognitive activities [17,20,30,33,34]; thus, further reviews are needed to compare the effect of serious games with a specific comparator). To address the aforementioned gaps, this study aimed to assess the effectiveness of serious games in improving memory among older adults with cognitive impairment. This review focused only on memory as other cognitive domains—for example, global cognition [25], executive functions [35], and processing speed [36]—were targeted by previous reviews.

This review included only RCTs that looked at the effectiveness of serious games in improving memory in older adults with cognitive impairment. The target intervention in this review was serious games supplied on any digital platform, such as computers (PCs), consoles (Xbox and PlayStation), mobile phones, handheld devices, Nintendo, or any other computerized device. Furthermore, components of gaming had to be used as an important and major technique for reaching the intervention’s goal. Serious games had to be used solely for the purpose of therapy. Studies combining serious games with other interventions were eligible if the control group received the same adjacent intervention. Nondigital games and those used for other purposes, such as monitoring, screening, diagnosis, and research, were excluded.

The study focused on the older adult population (aged ≥60 years) who had any type of cognitive impairment or condition (eg, MCI, Alzheimer disease, or dementia). Their diagnosis had to be confirmed by checking the inclusion criteria or baseline scores against standardized diagnostic criteria (eg, Mini-Mental State Examination and Montreal Cognitive Assessment). This review did not focus on healthy older adults, health care providers, or caregivers. No restrictions were applied regarding sex and ethnicity.

The main outcome of interest in this review was memory regardless of the type (verbal, nonverbal, or working memory) and regardless of the tool used for measuring the outcome. Studies were excluded if they assessed only other cognitive outcomes (eg, language and processing speed), cost-effectiveness, acceptance, feasibility, or satisfaction. This review focused on outcome data that were measured immediately after the intervention rather than on follow-up data.

Only RCTs conducted in English and from 2010 onward were considered. Pilot or feasibility RCTs, quasi-experiments, observational studies, and reviews were omitted. Studies published as journal articles, conference proceedings, or dissertations were included. Reviews, conference abstracts, proposals, editorials, and commentaries were all excluded. Finally, no restrictions related to the country of publication, comparator, or study setting were applied.

The studies that were relevant to this review were found by searching 7 bibliographic databases: MEDLINE (via Ovid), PsycINFO (via Ovid), EMBASE (via Ovid), CINAHL (via EBSCO), IEEE Xplore, ACM Digital Library, and Scopus. Furthermore, we searched the search engine Google Scholar. Owing to the high number of papers obtained through Google Scholar, only the first 10 pages (ie, 100 records) were taken into account as they were automatically ordered based on their relevance [37]. The first author (AA) conducted the search on August 6, 2021. An automatic alert was set up to retrieve studies that were added to the databases after that date; this continued for 16 weeks (ending on December 5, 2021). Forward reference list checking (ie, screening studies that cited the included studies) and backward reference list checking (ie, screening the reference lists of the included studies and relevant reviews) were carried out to retrieve further studies.

To develop the search query for this review, the authors consulted 2 experts in digital mental health and checked the search queries used in other systematic reviews within this field. The chosen search terms were related to the target population (eg, cognitive impairment), target intervention (eg, serious games and exergames), and target study design (eg, RCTs). Multimedia Appendix 2 summarizes the search query that was used for searching each of the 8 databases.

Relevant studies were identified taking the following steps. First, the obtained studies were imported into EndNote X8 (Clarivate Analytics) to identify and delete duplicate items. Second, the titles and abstracts of all the retrieved studies were evaluated in the second phase by 2 reviewers (AA and MH) working independently. Finally, the 2 reviewers independently evaluated the entire texts of the studies included in the previous step. Any disagreements in the 2 previous steps were resolved via discussion. The interrater agreement (Cohen κ) in steps 2 and 3 was 0.94 and 0.96, respectively, indicating a near-perfect level of interrater agreement [38].

In total, 2 independent reviewers (AA and MH) used Microsoft Excel to extract data from all the included studies. The data extraction form used to extract data from the included studies was pilot-tested using 2 of the included studies (Multimedia Appendix 3). The reviewers’ disagreements were resolved through discussion. An interrater agreement of 0.85 was observed, indicating a near-perfect degree of agreement. If data such as the mean, SD, and sample size were unavailable from the published studies, contact was made with the first and corresponding authors in an attempt to retrieve them.

The Cochrane Collaboration recommends assessing the risk of bias via 2 independent reviewers (AA and MH) using the Risk of Bias 2 (RoB 2) tool [39]; as such, these guidelines were followed for this review. The RoB 2 tool assesses the risk of bias in 5 domains of RCTs: randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result [39]. The risk of bias judgments in these domains were used to determine the overall risk of bias of each included study. Any inconsistencies in decisions between the reviewers were resolved by consulting a third reviewer. Interrater agreement between the reviewers was near perfect (Cohen κ=0.93) [38].

A narrative and statistical approach was used to synthesize the information acquired. In our narrative synthesis, we used texts and tables to describe the characteristics of the included studies (demographic, intervention, comparator, and outcome variables). The results of the experiments were categorized and pooled based on measured outcome (ie, verbal, nonverbal, and working memory) and the comparator (ie, control, conventional exercises, conventional cognitive training, and other serious games). A meta-analysis was conducted when at least two studies with the same measured outcome and comparator submitted enough data (ie, mean, SD, and number of participants in each intervention group). Owing to the type of data for the outcome of interest (memory) being continuous and the methods used to measure the outcome being variable throughout the included studies, the standardized mean difference (SMD; Cohen d) was used to analyze the overall effect of each study. The random effects model was used for the analysis because of the high clinical heterogeneity among the meta-analyzed studies in terms of serious game characteristics (eg, type, duration, frequency, and period), population characteristics (eg, sample size, mean age, and health condition), and outcome measures (ie, tools and follow-up period). As several studies used more than one outcome measure to assess memory, the dependency on effect sizes within or across studies will be introduced in the meta-analysis. As a result, a multilevel meta-analysis considering the dependency on effect sizes and sampling covariance between the effect sizes was used [40-42]. Namely, the multilevel meta-analysis should be applied when effect sizes within the same study are very likely to be more similar to each other than the effect sizes across studies [42]. The R (version 4.3.1; R Foundation for Statistical Computing) statistical package was used to perform the analysis. We used the function rma.mv in the library metafor, which is a library in R, to perform the multilevel meta-analysis [43].

If we observed a statistically significant difference between the groups in a meta-analysis, we further sought to examine if it was clinically significant. The phrase “minimal clinically important difference” (MCID) refers to the smallest change in a measured outcome that a patient would consider worthwhile and significant enough to warrant a change in treatment. The MCID boundaries were calculated as 0.5 times the SMD of the meta-analyzed studies.

We calculated I² and a chi-square P value to investigate the degree and statistical significance of the heterogeneity in the meta-analyzed studies, respectively. A chi-square P value of ≤.05 suggests heterogeneous meta-analyzed studies [44]. When I² ranged from 0% to 40%, 30% to 60%, 50% to 90%, and 75% to 100%, the degree of heterogeneity was judged to be insignificant, moderate, substantial, or considerable, respectively [44].

To appraise the overall quality of evidence resulting from the meta-analyses, we applied the Grading of Recommendations Assessment, Development, and Evaluation approach [45], which assesses the quality of evidence based on 5 domains: risk of bias, inconsistency (ie, heterogeneity), indirectness, imprecision, and publication bias [45]. In total, 2 reviewers independently rated the overall quality of the meta-analyzed evidence, and disagreements were resolved through discussion. The interrater agreement of the reviewers was considered near perfect (Cohen κ=0.87) [38].

The total number of records retrieved by searching the predefined databases was 618 (Figure 1). Of these 618 records, 161 (26.1%) duplicates were removed using the EndNote software. Checking titles and abstracts of the remaining records led to the exclusion of 52.3% (323/618). After reading the full texts of the remaining 134 publications, 116 (86.6%) were excluded, mainly because of the population (n=67, 57.8%). The list of studies that were excluded after screening the full texts is provided in Multimedia Appendix 4. No additional studies were found through backward and forward reference list checking. In total, 18 RCTs were included in this review [46-63]. Of these 18 studies, 15 (83%) were included in 10 meta-analyses [47-49,51,52,54-63]. A total of 17% (3/18) of the studies were excluded from the meta-analyses because 33% (1/3) [46] did not report the data required for the meta-analysis (eg, mean and SD) and 67% (2/3) [53,61] compared serious games with other serious games that had different characteristics; therefore, including them in a meta-analysis would not make sense.

The included studies were published between 2012 and 2021 (Table 1). The year in which the largest number of included studies was published was 2015 (4/18, 22%). The included studies were carried out in 13 different countries, and there was a general equal distribution of studies in these countries. All the included studies were peer-reviewed journal articles except for a book chapter included (1/18, 6%). The trial type was parallel RCT in most of the included studies (17/18, 94%).

The sample size of the included studies varied from 20 to 209, with an average of 81. The mean age of the participants in the included studies ranged from 66 to 83.1 years, with an average of 74.5 years. The percentage of men in the included studies ranged from 21.5% to 71%, with an average of 46.5%. The participants in most of the included studies had MCI (14/18, 78%). Participants were recruited from clinical settings in 67% (12/18) of the studies, from the community in 28% (5/18) of the studies, and from both clinical settings and the community in 6% (1/18) of the studies.

Serious games alone were used as interventions in 89% (16/18) of the included studies, whereas the remaining 11% (2/18) of the studies used serious games combined with conventional exercises [48] or sham exercises [49] (Table 2). The included studies used 16 different serious games. On the basis of the therapeutic modality that they delivered, the serious games used in the included studies were grouped into 2 types: cognitive training games (16/18, 89%) and exergames (2/18, 11%). Games were designed with a “serious” purpose from the beginning (designed serious games) in all studies except for 6% (1/18) that used a purpose-shifted game (which was not designed as a serious game from the start but rather was used for a serious purpose). The most common platform used for playing the games were computers (14/18, 78%). In 67% (12/18) of the studies, serious games were played under the supervision of health care providers or caregivers. The duration of the games in the included studies ranged from 7 to 90 minutes, and the most common duration was 60 minutes (7/18, 39%). The frequency of playing the games varied between 2 and 7 times per week, but it was 2 times per week in half of the studies (9/18, 50%). The period of intervention ranged from 2 to 25 weeks, but it was ≤12 weeks in 72% (13/18) of the studies.

The comparison groups received only passive interventions in 39% (7/18) of the studies, whereas they received only active interventions in 44% (8/18) of the studies (eg, conventional exercises and conventional cognitive activities; Table 3). In total, 17% (3/18) of the studies delivered both active and passive interventions as comparators. The duration of the active comparators ranged from 7 to 100 minutes. The frequency of the active comparators varied between 2 and 7 times per week. The period of the active comparators varied between 2 and 25 weeks. Most of the included studies (16/18, 89%) measured more than one outcome. The measured outcomes were verbal memory in 78% (14/18) of the studies, nonverbal memory in 61% (11/18) of the studies, and working memory in 67% (12/18) of the studies. The studies used 32 different tools to measure these outcomes, but the most common tool used was the Wechsler Memory Scale Third Edition (7/18, 39%). The outcomes were measured immediately after the intervention in all the included studies (18/18, 100%). The follow-up period ranged from 4 to 264 weeks. Participant attrition was reported in 89% (16/18) of the studies, and it ranged from 0 to 23.

An appropriate random allocation sequence for the randomization process was used in 44% (8/18) of the studies. Researchers in 39% (7/18) of the studies concealed the allocation sequence until participants were assigned to the interventions. The groups were comparable at baseline in all studies (18/18, 100%). Thus, the risk of bias owing to the randomization process was rated as low in only 33% (6/18) of the studies (Figure 2).

Participants and those who delivered the interventions were aware of the assigned interventions during the trial in 67% (12/18) and 83% (15/18) of the studies, respectively. None of the studies reported a deviation from the intended intervention because of experimental contexts; however, 11% (2/18) of the studies provided insufficient information to verify if protocol deviations had occurred. Appropriate analysis methods (eg, intention-to-treat analysis) were used in 89% (16/18) of the studies to estimate the effect of the intervention. According to these judgments, the risk of bias because of deviations from the intended interventions was low in 78% (14/18) of the studies (Figure 2).

Missing outcome data were <5% in 44% (8/18) of the studies. There was evidence that the findings were not biased by missing outcome data in only 6% (1/18) of the studies. The missing outcome data resulted from reasons that were documented and not related to the outcome in 28% (5/18) of the studies. Therefore, there was a low risk of bias because of missing outcome data in 78% (14/18) of the studies (Figure 2).

In all the included studies (18/18, 100%), the outcomes of interest were evaluated using appropriate measures, and the measurement methods were comparable across the intervention groups. The assessor of the outcome was aware of the assigned interventions in 39% (7/18) of the studies, but it was unlikely that the assessment of the outcome was influenced by knowledge of the intervention received in these studies. Accordingly, all studies (18/18, 100%) had a low risk of bias in the “measuring the outcome” domain (Figure 2).

In total, 28% (5/18) of the studies published their protocols in sufficient detail. In all studies (18/18, 100%), the reported outcome measurements did not differ from those specified in the analysis plan, and there was no evidence that the studies selected their results from many results produced from multiple eligible analyses of the data. On the basis of these judgments, the risk of bias because of the selection of the reported results was considered low in 28% (5/18) of the studies (Figure 2).

In the last domain, “overall bias,” the risk of bias was considered high in 22% (4/18) of the studies as they were judged as having a high risk of bias in at least one domain. A total of 61% (11/18) of the studies raised some concerns in the domain of overall bias as they had some issues in at least one of the domains and were not at high risk for any domain. The remaining 17% (3/18) of the studies were judged to be at low risk of bias for the domain of overall bias given that they were rated to be at low risk of bias for all domains. The reviewers’ judgments about each “risk of bias” domain for each included study are presented in Multimedia Appendix 5 [46-63].

This study summarized the evidence regarding the effectiveness of serious games in improving memory. Our meta-analyses showed that serious games are more effective than no or passive interventions in improving nonverbal and working memory. Surprisingly, we found that serious games are as effective as no or passive interventions in improving verbal memory, which, therefore, deems serious games ineffective. This review demonstrated that serious games are more effective than conventional exercises in improving verbal memory. However, we found that serious games are as effective as conventional exercises in improving nonverbal and working memory, indicating that serious games are comparable with conventional exercises. Evidence suggests that cognitive training and exercise work through distinct neuronal mechanisms and, therefore, if combined, they might have synergistic and more effective results compared with being used as separate interventions [65,66]. Studying this synergistic relationship will become important in future primary research and trials. With the advances in virtual reality technologies, their availability, and rising applications of the metaverse [67], more evidence is needed to assess the effectiveness of virtual reality–based exergames in improving memory [68].

The meta-analyses in this review showed that serious games are as effective as conventional cognitive training in improving verbal, nonverbal, and working memory, meaning that serious games and conventional cognitive training are comparable. Furthermore, we found that the effect of adaptive serious games is similar to that of nonadaptive serious games in improving working memory.

The findings of our review and previous reviews were consistent for some outcomes and different for others. Specifically, a systematic review conducted by Lampit et al [32] compared the effect of cognitive training games with that of passive and active interventions on verbal, nonverbal, and working memory in healthy older adults. Consistent with our findings, the review found no statistically significant difference (P>.05) in the effect of cognitive training games and of no or passive interventions on verbal memory, and there was a statistically significant difference in working memory between the groups, favoring cognitive training games over no or passive interventions [32]. In contrast to our findings, Lampit et al [32] did not find a statistically significant difference in nonverbal memory between the groups. The contrary finding may be attributed to the following reasons: (1) although the number of participants was ≥100 in 17% (3/18) of the studies in our meta-analysis, the number of participants was <100 in all studies included in the meta-analysis by Lampit et al [32]; (2) the total number of training hours was >15 in only 2 studies included in the meta-analysis by Lampit et al [32], whereas the total number of training hours was >15 in 28% (5/18) of the studies in our meta-analysis; and (3) all studies meta-analyzed in our review (15/18, 83%) recruited participants with cognitive impairment, whereas all studies meta-analyzed in the review by Lampit et al [32] recruited participants without cognitive impairment.

In contrast to our findings, Lampit et al [32] found a statistically significant difference in verbal, nonverbal, and working memory between the groups, favoring cognitive training games over active interventions [32]. This may be attributed to the following reasons: (1) our findings related to these comparisons are based on meta-analyses of 11% (2/18) of the studies, whereas the findings of Lampit et al [32] are based on meta-analyses of 6 to 15 studies; (2) Lampit et al [32] compared cognitive training games with all active interventions, but this review compared cognitive training games with specific active interventions (ie, conventional exercises and conventional cognitive training); (3) all studies meta-analyzed in our review (15/18, 83%) recruited participants with cognitive impairment, whereas all studies meta-analyzed in the review by Lampit et al [32] recruited participants without cognitive impairment; and (4) the review by Lampit et al [32] included some pilot RCTs, whereas our review included only RCTs.

Another review examined the effect of cognitive training games on verbal and working memory among healthy older adults regardless of the comparator type (ie, passive and active controls) [30]. Meta-analyses in that review showed a statistically significant difference in verbal memory (P=.03) and working memory (P<.001) between the groups, favoring cognitive training games over all types of comparators [30].

Hill et al [34] conducted a systematic review to assess the effect of cognitive training games on verbal, nonverbal, and working memory among people with MCI or dementia regardless of the comparator type. For people with MCI, the review found a statistically significant difference between the groups in verbal and working memory (P<.001), favoring all types of comparators [34]. In contrast, there was no statistically significant effect of cognitive training games on nonverbal memory when compared with all types of comparators [34]. For people with dementia, the review showed no statistically significant effect of cognitive training games on verbal, nonverbal, or working memory when compared with all types of comparators [34].

This review cannot comment on the effectiveness of serious games (1) delivered on nondigital platforms, (2) used for other purposes (eg, screening or diagnosis), (3) used for improving other cognitive abilities (eg, learning, processing speed, and executive functions), (4) among other age groups, or (5) among those without cognitive impairment. This is because such interventions, outcomes, and populations were beyond the scope of this review.

It is likely that we missed some relevant studies as this review did not search some databases (eg, PubMed and the Cochrane Library [CENTRAL]) and excluded studies that were quasi-experiments, pilot RCTs, published before 2010, or written in non-English languages. The quality of the evidence was very low in all meta-analyses except for 10% (1/10); this may decrease the internal validity of our findings. We cannot comment on the long-term effect of serious games on memory as this review focused on the short-term effect of serious games by meta-analyzing only postintervention data rather than follow-up data. This is because the follow-up period was not consistent among the studies.

The effect size for each meta-analyzed study was likely overestimated or underestimated in this review given that the authors used postintervention data for each group to assess the effect size rather than the pre-post intervention change for each group. Postintervention outcome data were used as most studies (15/18, 83%) did not report the mean and SD for pre-post intervention change in memory for each group, and there was no statistically significant difference in memory between the groups at baseline in all studies (18/18, 100%).

Serious games may have a significant role to play in improving verbal, nonverbal, and working memory in older adults with cognitive impairment. However, these findings should be treated with caution given that most meta-analyses (7/10, 70%) were based on a few studies (≤3) and judged to have a low quality of evidence for the following reasons: most of the included studies (11/18, 61%) were judged to have some concerns regarding the overall bias, the heterogeneity of the evidence was high in approximately half of the meta-analyses (4/10, 40%), and the total effect sizes were imprecise in all meta-analyses (10/10, 100%). Therefore, serious games should be offered as a supplement to existing proven and safe interventions rather than as a complete substitute until further, more robust evidence is available. Further reviews are necessary to investigate the short- and long-term effect of serious games on memory and other cognitive abilities (eg, executive function, processing speed, and learning) among people of different age groups with or without cognitive impairment.