Mild cognitive impairment (MCI) is a prodromal stage of Alzheimer’s disease (AD) and other types of dementia. In patients with MCI (pwMCI), circumscribed cognitive abilities are commonly below age expectation despite generally intact daily functioning (Petersen, 2004; Smith and Bondi, 2013). However, while pwMCI remain independent in primary daily activities, they may encounter difficulties performing complex functional activities (e.g., managing finances, medications, or shopping) and request increased caregiver attention (Albert et al., 2011). MCI is associated with an approximate 12% annual conversion rate to dementia while the comparable normal control group rate is only 1–2% (Petersen et al., 1997, 2001; Shah et al., 2000). In longer-term follow-up studies approximately 80% of pwMCI converted to dementia within six years (Petersen et al., 1999).

While neurodegenerative forms of dementia are irreversible, non-pharmacological interventions (i.e., behavioral interventions such as physical exercise, note taking, social engagement, and computerized cognitive training) administered at an early stage (e.g., MCI) can preserve functional independence, slow cognitive decline, and thereby delay the onset of dementia (Gauthier, 2005; Levy et al., 2022). A review by Chandler et al. (2016) revealed the benefits of behavioral interventions in improving mood (k = 26, Cohen’s d = 0.16, 95% CI = [0.03–0.28]), functional ability (k = 31, d = 0.23, 95%CI = [0.16–0.47]), and metacognition (k = 26, d = 0.30, 95%CI = [0.15–0.58]) in pwMCI (Chandler et al., 2016). Since that review, numerous additional multicomponent interventions have been reported in pwMCI or other at-risk groups. Large multicomponent behavioral interventions such as Vivifrail, which consisted of physical resistance, balance, flexibility, and gait-retraining exercises for three months, have shown significant improvements in functional capacity, cognitive function, and depression (Casas-Herrero et al., 2022). Alternative interventions including lifestyle training might also play an essential role in mood and functional improvement (Gale et al., 2019; Yu et al., 2019).

These observed benefits lead to hypotheses that repeated cross-domain training might stimulate “compensatory scaffolding” and neuroplastic reorganization (Sherman et al., 2017). In other words, the combination of several approaches in a multicomponent treatment program interventions targeting multiple domains may exhibit additive efficacy. In one systematic review only multicomponent (k = 16, Hedges’ g = 0.40, 95%CI = [0.16, 0.63]) and multidomain-focused cognitive training (k = 13, g = 0.23, 95% CI = [0.108, 0.352]) yielded statistically significant improvement in cognitive outcomes post-intervention in pwMCI when compared to MCI controls (Sherman et al., 2017). Thus, combining multiple interventions has been increasingly emphasized as a tool to facilitate functional retention. Previous systematic reviews and meta-analyses have reported benefits in combining physical exercises with cognitively challenging activities in both clinical and non-clinical older adults (Zhu et al., 2016; Gheysen et al., 2018; Gavelin et al., 2021). In one meta-analysis, combined cognitive-physical interventions showed small-to-medium positive effects (k = 10, standardized mean difference (SMD) = 0.32, 95%CI = [0.17–0.47]) on global cognitive function and moderate-to-large effects (k = 4, SMD = 0.65, 95%CI = [0.09-1.21]) on activities of daily living (ADL) in MCI or dementia patients (Karssemeijer et al., 2017). In contrast, despite the significant benefits evidenced in most studies, a recent systematic review found no difference between combined cognitive-physical training and interventions with isolated elements in executive function, processing speed, attention, mood, and cardiorespiratory fitness (Yang et al., 2020). However, the review focused primarily on cognitive outcomes, which might not reflect the overarching efficacy of multimodal interventions across domains (e.g., quality of life and independent daily functioning).

A few limitations were identified in existing systematic literature reviews and meta-analyses. First of all, while the effects of combined interventions have been extensively studied in the past decade (see Supplementary material A), research has focused predominantly on comparative effectiveness analysis (Amofa et al., 2021; Levy et al., 2022), a tool commonly used to explore the additive effect of a specific arm instead of changes an overall program has exerted. For example, Imaoka et al. (2019) used comparative effective analysis to investigate the additive effect of soy peptide as a supplement to memory exercise in pwMCI but did not study the overall efficacy of both when compared to an untreated control group. Secondly, some studies and reviews have mixed samples of pwMCI with healthy older adults or early dementia patients (Li et al., 2011; Straubmeier et al., 2017; Bruderer-Hofstetter et al., 2018; Stephen et al., 2019; Santos Lopes da Silva et al., 2023) due to the small amount of available literature (Gheysen et al., 2018, k = 9; Han et al., 2022, k = 3; Karssemeijer et al., 2017, k = 5). Nevertheless, primary preventions in cognitively healthy older adults can serve distinctive roles from interventions for those with known risk of decline (i.e., secondary preventions). Secondary preventions usually incorporate compensation training and adjustment-related treatments to slow or prevent further decline (Smith, 2016). On the other hand, tertiary preventions for those with dementia diagnoses rely heavily on participants’ capacity to grasp the ideas, which might include differential strategies and evaluation systems from interventions designed for pwMCI. Thus, an essential question regarding the effectiveness of multimodal intervention as a secondary prevention in pwMCI remains unclear. Thirdly, there is a lack of consensus on targeted outcomes. Some studies focused primarily on mobility (Kiper et al., 2022; Mai Ba and Kim, 2022) while others focused on cognition (Yan et al., 2022). Lastly, while one meta-analysis (Meng et al., 2022) has synthesized clinical trials combining cognitive intervention and physical exercise on multiple cognitive domains in pwMCI, this meta-analysis excluded behavioral interventions other than physical exercise and included single intervention comparisons to study the additive effects instead of the overall impact of multimodal interventions. Furthermore, this study also suffered from a limited number of reports (k = 8) of randomized control trials (RCTs).

In addition, the definition of “multimodal” varied across studies and was often mixed with terms including “multicomponent” or “multifaceted.” For example, a combination of different physical exercises (Lau et al., 2015; Trautwein et al., 2020; Barisch-Fritz et al., 2022) or cognitive training targeting multiple domains (Tsolaki et al., 2011; Olchik et al., 2013) were treated as multimodal in several studies. While these interventions have included multiple strategies, the target was often limited to one area of concern instead of a comprehensive approach that can target multiple interrelated areas of concern simultaneously. Studies have also used the term “multimodal” to describe treatments conducted in different settings (e.g., home vs. clinic) or through different delivery methods (e.g., computer vs. paper). To establish an operational definition and delineate the targeted treatment types for this review, multimodal interventions generally refer to combining several training approaches that target different outcome domains in a treatment program (Giusti et al., 2017).

In summary, we believe that examining truly multimodal interventions that focus on or at least partition pwMCI for separate analysis might assist future explorations of comprehensive and efficient intervention programs for persons at the highest risk for dementia. Therefore, the aims of the current systematic review and meta-analysis are (1) to perform a synthesis of existing research of multimodal interventions on cognition and mood for individuals who meet the criteria of MCI and (2) to investigate the clinical implications and limitations of these results for future treatment planning.

The eligibility criteria are consistent with the PICO criteria and the PRISMA 2020 reporting guidelines (Page et al., 2021a), and incorporate participants, interventions, comparators, and outcomes. Only RCTs were included in the review with no restrictions on cohort studies, longitudinal studies, and crossover designs.

This review only included published studies and abstracts written in or translated into English. PubMed, Embase and Cochrane Library were searched for articles published before January 1st, 2024. In addition, references from relevant publications and symposiums were examined and manually searched as an additional source of literature. Please see Supplementary material B for searching items.

Targeted variables and measures were identified and extracted by GY and APL independently to an Excel spreadsheet and compared to ensure no errors. Outcomes included changes in cognition and mood. Outcomes were identified by searching the specific terms in the report regardless of measuring tools. Authors were not contacted when information regarding the primary outcome was not available in the text.

Participant age, study attrition rate, diagnostic criteria, specific multimodal intervention strategies and characteristics (duration, frequency, and follow-up duration), comparator characteristics (no treatment vs. alternative treatment), outcome measures, effect sizes for each outcome, and results reported by the authors were extracted and documented for all eligible publications.

The revised Cochrane Collaboration software (RoB 2) (Sterne et al., 2019) assessing the risk of bias in RCTs was employed in the current review. Detailed criteria of focus in each domain can be found in the Cochrane Handbook Chapter 8.2 (Higgins et al., 2019). An overall risk-of-bias judgment was obtained for individual domains by both GY and APL. Similarly, a consensus meeting was arranged to resolve any discrepancies during the process. Results of the risk-of-bias assessment were then visualized through another web-based R package, robis (McGuinness and Higgins, 2021). Because several studies included both targeted outcomes, each outcome was assessed separately and weighted equally in the evaluation. Figure 1 depicts the results of 26 parallel design evaluations conducted for 15 clinical trials.

Comparisons of the baseline characteristics were employed to evaluate any effects raised by the randomization process. Studies that failed to report any differences between the intervention and control groups regarding demographic variables (e.g., age, gender, etc.) or targeted outcomes (e.g., cognition) raised concern about whether an appropriate analysis was used to estimate the effect of assignment (Domain 2) and whether baseline differences suggested a problem with randomization (Domain 1). “No information” on the randomization process (Domain 1) was given to a few studies, which led to a rating of “some concerns,” due to a failure to clarify the sequence allocation method. In addition, “probably no” was given to one study using consecutive recruitment with no information on the randomization strategy (Kurz et al., 2009). Studies with a larger than 5% dropout rate, according to the guidelines, were rated as “probably not” for whether the outcomes were provided for almost all the participants (Domain 3). If the reasons for attrition were provided and were irrelevant to participants’ cognitive functioning, the overall rating for the domain remained “low risk.”

The goal of a meta-analysis is to estimate the overall effect of treatments across studies. However, because studies vary in the quantity and quality of information, different weight is assigned to each study (e.g., higher weight assigned to larger studies) to calculate a combined effect. Due to the variability of sample sizes and characteristics among the reports included in the current study, we used the random effect model of meta-analysis, which assumes that each study is estimating a different effect size, to estimate the mean of a distribution of true effects for each outcome.

Between-study variance, Ƭau², was calculated through total variance (Cochrane’s Q), which denotes the squared deviations of each study from the combined mean, and the degrees of freedom (df). Due to the small sample size and heterogeneity across study populations, the random effects model with maximum likelihood (Borenstein et al., 2007) was employed to compute the heterogeneity and combined effect of the studies. For cognition, the aggregated model was used to indicate heterogeneity attributed to the variance across studies. In addition, to account for the impact of sample size on Q, we calculated the total proportion of variance owing to heterogeneity (I²) for each outcome (Higgins et al., 2013). In general, I² categorizes results into low (25%), moderate (50%), or substantial (75%) heterogeneity. The analyses were performed on Metafor (Viechtbauer, 2010).

Publication bias generally refers to the probability of bias stemming from unpublished results of studies with non-significant data (Borenstein et al., 2009). A common way of assessing publication bias is through the level of symmetry of a funnel plot, which depicts the relationship between effect sizes and standard error in each study. Because small studies are more likely to generate non-significant results and have a larger standard error, they are less likely to be published. The funnel plot inverted the y-axis (standard error) to position these smaller studies at the bottom while placing the larger ones on the top. Thus, the top of the funnel should distribute closely to the mean effect size whereas the bottom should scatter heavily on both the left and right sides (the shape of a funnel) when there is no publication bias. Aside from the graph, we also used the modified Egger’s regression test by Pustejovsky (Egger et al., 1997; Pustejovsky and Rodgers, 2019) to assess asymmetry of the funnel plots incorporating the standard error of between group SMD using the following formula: S E ∗ S M D between = n 1 + n 2 n 1 n 2 . The resulting value is equivalent to a z-score with a similar rejection range above 1.96 or below −1.96 for a significance level below 0.05. These tests were all performed through Metafor and Dmetar (Viechtbauer, 2010; Harrer et al., 2021) in R (R Core Team, 2014).

A total of 482 results were identified after a systematic search of PubMed (k = 126), Embase (k = 106), and Cochrane Library database (k = 250). Among them, 105 duplicates were removed prior to screening, which yielded 377 results for review. A preliminary abstract/title review excluded 356 articles, of which the majority were study protocols or interventions targeting combined MCI and dementia populations. In the remaining 21 reports, 10 were excluded after a full-text review. A list of excluded reports was provided in Supplementary material C. Specifically, three studies were excluded due to a lack of multimodal intervention. Two studies used comparative effectiveness analysis. In addition, four studies were excluded because the group receiving multimodal interventions was not directly compared to the double-sham control group but to other single-modal interventions, and one study lacked randomized groups. In the end, 11 clinical trials were included from the databases for review.

Manual citation searching from previous literature reviews (Chandler et al., 2016; Karssemeijer et al., 2017; Gheysen et al., 2018; Gavelin et al., 2021; Han et al., 2022; Meng et al., 2022) found 28 results that did not overlap with the primary database search. After abstract/title screening, 12 remained for full-text screening. Of those clinical trials, two multi-group studies with no direct comparison between the multimodal and control groups, one study with a wrong comparator (i.e., the control group received mixed interventions), and one report with a mixed sample of MCI and dementia patients were removed. As a result, eight studies were included in the final review. A detailed PRISMA 2020 flowchart is demonstrated in Figure 3 (Page et al., 2021b).

Overall, 19 journal articles were eligible for the final review, and 18 were included in the meta-analysis. One report (Troyer et al., 2008) was excluded due to the high risk of bias (Figure 1). Of these, 18 reports of cognition (n = 1,555, mean age = 73.54 years old) and seven reports of mood (n = 343, mean age = 72.08 years old) were identified. A few reports failed to include an effect size or a p value for nonsignificant results, for which certain outcomes were not included in data extraction.

Participants’ mean ages were obtained from baseline characteristics for most of the studies except for Griffiths et al. (2020), which only included the number of participants in two age groups (60–69 years) and (70–79 years). Mean ages ranged from 67.82 to 87.20 with a standard deviation of 4.26. The attrition rate ranged from 0 to 34.78% with three studies (Kounti et al., 2011; Rojas et al., 2013; Bae et al., 2019) reporting above 30% dropout rates at the end of the intervention. Details regarding age, attrition rate, intervention methods, sample size, country, and follow-up durations are presented in Table 1.

Some concerns were reported for most of the studies due to the lack of published protocols for a proper comparison between the actual analysis and an analysis plan before unblinded outcome data were available (Domain 5). Other common concerning criteria included whether participants were aware of their assigned intervention during the trial (Domain 2) and whether the allocation sequence was concealed from participants until enrollment (Domain 1). Maffei et al. (2017), Lam et al. (2015), Xu et al. (2020), and Montero-Odasso et al. (2023) were the only studies that explicitly stated that participants were not informed of their group assignment until the beginning of the intervention. Troyer et al. (2008)’s randomization process (Domain 1) was rated “high risk” due to missing information regarding allocation concealment and significant group differences favoring the control group on cognitive functioning at baseline. Therefore, the study was not included in the final meta-analysis. In the end, only two studies (Lam et al., 2015; Montero-Odasso et al., 2023) received an overall rating of “low risk.”

In general, heterogeneity was low for mood (Ƭau² = 0.046, Q(7) = 10.33, p = 0.17, I² = 29.7%) and minimum-low for the aggregated cognition outcomes (Ƭau² = 0.040, Q(17) = 21.71, p = 0.20, I² = 21.7%). Study characteristics such as sample size, education, frequency of intervention, and intervention modalities might serve as potential sources of heterogeneity. Heterogeneity as indicated by I² represented between-study variability regardless of the number of studies. In this case, studies involving either mood or cognition outcomes only differed by sampling error, which did not appear to impact the overall aggregated meta-analysis model.

Egger’s test with adjustment did not indicate asymmetry in the funnel plot for cognition (bias = −1.59, intercept =0.62, t(16) = −1.29, p = 0.216) or mood (bias = 0.46, intercept = 0.51, t(6) = 0.37, p = 0.727), which reflects the absence of publication bias in both outcomes (Figures 4, 5).

In the past decade, clinical trials on pharmacological interventions have not demonstrated improvement in cognition for pwMCI (Ströhle et al., 2015; Fink et al., 2018). While the FDA has recently approved Aducanumab for early stages of AD, findings did not support cognitive benefits in pwMCI (Knopman et al., 2021). Even in RCTs that showed cognitive improvement of donepezil (SMD = -0.90), the benefit was rather subtle (1 point between group difference on the 89-item ADAS-cog scale) (Doody et al., 2009). In addition, research has emphasized the frequent treatment-emergent adverse events such as diarrhea, nausea, abnormal dreams, and even increased mortality in the treatment group (Winblad et al., 2008; Doody et al., 2009). A meta-analysis of 41 RCTs has suggested small to moderate effect sizes of cholinesterase inhibitors on cognitive function (SMD = 0.10–0.46) (Cooper et al., 2013). Thus, results from this meta-analysis showed generally comparable or larger effects of multimodal nonpharmacological interventions on cognition and mood, which are consistent with previous reports (Sherman et al., 2017) and further supported the utility of these interventions to maintain functionality and facilitate adjustment to cognitive changes.

Studies failed to mention the race and ethnicity of participants, mainly due to the homogeneity of the populations. Impacts of racial/ethnic background on the effects of multimodal or single-modal interventions have not yet been studied. Another limitation of the studies pertains to the absence of control of repeated measure effects except for Kurz et al. (2009). Because most interventions were conducted within a short time frame, a repeated testing effect at the end of the intervention, especially in cognitive tasks, might have mediated the observed changes post-intervention (Roediger and Payne, 1982). Furthermore, only one report included dementia conversion rate as an outcome (Rojas et al., 2013). Conversion to dementia was seen in one trained and three non-trained patients at the 12-month follow-up, and significant declines in global cognition were seen in the non-trained group at the six-month follow-up assessment (Rojas et al., 2013). However, since the conversion rate was low in both groups and no significant improvement was observed in the intervention group immediately after the intervention, the results need further examination to determine whether long-term effects were present. Thus, a longitudinal analysis of whether these multimodal interventions have delayed dementia progression is needed.

The Yang et al. (2022) study was found to be an outlier on the Funnel plot, indicating potential heterogeneity/publication bias. Findings in the study suggested significant cognitive improvement in the intervention group but a decline in untreated controls. Despite observed deviations from other studies, further evaluation of study population, methodology, interventions, and outcomes did not demonstrate evidence of bias or poor data quality. Therefore, we speculated that the distinctive results might stem from the relatively intense schedule for a long intervention period (6 months). The study was also unique in its short and frequent follow-ups (1-, 3-, and 6-month follow-ups). However, these hypotheses might not completely explain the reason for the deviation, and the results of Yang et al. (2022) should be interpreted with caution.

One of the limitations of this review is the lack of consensus in MCI diagnostic criteria across reports. Most studies included older adults with an MCI diagnosis regardless of subtype. However, four reports included only single or multidomain aMCI (Buschert et al., 2011; Park et al., 2019; Jeong et al., 2021) and one used the term mNCD and MCI interchangeably (Griffiths et al., 2020). Additionally, this study did not investigate the effects of different modes of delivery (simultaneous vs. sequential). Sequential designs were defined as delivering intervention modalities in separate sessions during the same period (e.g., exercise followed by cognitive training). In contrast, simultaneous designs were usually delivered by asking participants to perform certain cognitive tasks while exercising at the same time or by using exergaming. Most of the interventions in the current review delivered different modalities through a sequential design whereas several dual-task trainings were administered using exergaming (Delbroek et al., 2017; Donnezan et al., 2018; Shimada et al., 2018; Park et al., 2019). In healthy and cognitively impaired older adults, simultaneous training was found to be more efficacious for cognition than sequential combinations of physical exercises and cognitive training (g = 0.32–0.38) (Zhu et al., 2016; Gheysen et al., 2018; Gavelin et al., 2021). However, whether simultaneous or sequential delivery is superior in pwMCI has yet to be studied. An analysis to compare the modes of delivery was beyond the scope of this review. Future research could focus on differences in efficacy associated with modes of delivery.

Another limitation pertains to the number of databases searched in the study. We only searched three major databases. However, research shows that using Embase combined with PubMed can cover approximately 88% of the available literature (Frandsen et al., 2021). Previous studies have also indicated high coverage rates when combining the Cochrane Library and EMBASE (88% in hypertension systematic review) (Rathbone et al., 2016) or the three search engines (97% in orthopedic research) (Slobogean et al., 2009). Additional bibliographic databases did not provide unique records when two or three of the above databases were searched due to significant overlaps across databases (Royle and Milne, 2003; Hirt et al., 2021). The Cochrane Library was also found to have the highest precision rate in literature reviews and to be sensitive in identifying RCTs (Royle and Milne, 2003). Therefore, a combination of these three databases and a manual reference search were considered sufficient to identify all the studies meeting our inclusion criteria.

Studies of multimodal cognitive and behavioral interventions on pwMCI demonstrated small to moderate positive effects on cognition and mood. A few directions for future research are postulated: (1) including long-term follow-ups to evaluate adherence and efficacy in delaying dementia conversion, (2) comparing effects of similar interventions in patients from diverse racial/ethnic backgrounds to inform adjustment in designs, and (3) considering simultaneous vs. sequential modes of delivery.