To formulate and refine the eligibility criteria and search strategy, a scoping review was performed. Existing literature was reviewed to identify the population, intervention, comparisons, outcome, and study design (PICOS) principles used in this systematic review.

To be eligible for inclusion, articles must have been peer-reviewed and published in English, without any limitation on publication year. Cognition and brain function assessments must have been conducted pre/post intervention. Cognitive measures were deemed valid based on whether an appropriate citation for the assessment tool was provided within the reported study (as assessed in the methodological quality checklist described in Section 2.5). Measures of brain function specifically related to neuroimaging and physiological measurement modalities that quantify the functional activity of the brain including fMRI, EEG, MEG, PET, SPECT, fNIRS, TMS, MRS, TCD, and ASL. Studies only reporting structural brain measures (e.g., volumetric MRI, diffusion weighted imaging) were excluded. If a study involved multiple groups of the same exercise type (e.g., resistance training once/week in group one and resistance training twice/week in group two), the protocol that most closely adhered to current physical activity guidelines (Bull et al., 2020) was assessed further. Studies were excluded if: a full text could not be obtained or they were published in a format other than a journal article (e.g., review, book chapter, opinion article, thesis or editorial). Studies that did not measure the isolated effects of exercise were excluded (e.g., multimodal studies and exergaming), as were studies which included people with comorbid neurological (e.g., Parkinson's disease) or psychiatric conditions (e.g., depression).

We conducted a search for articles published in EBSCOHost (indexing CINAHL Plus, SportDiscus, PsycINFO, APA PsycArticles, Psychology and Behavioral Sciences Collection, and Ageline), Web of Science (indexing Web of Science Core Collection, Current Contents Collection, Chinese Science Citation Database, KCI-Korean Journal Database, MEDLINE, SciELO Citation Index) from inception to 10 March 2023. Search terms included words related to our PICOS. Table 1 shows the advanced search strategy used in EBSCOHost, which was adapted for the other database.

All retrieved study titles and abstracts were imported into COVIDENCE. Following the removal of duplicates, titles and abstracts were independently screened by two researchers (RC and NJM) and full texts of potentially relevant studies were obtained. The full texts of relevant studies were reviewed independently by three researchers (AS, GZS, and DK); this included the review of reference lists for potentially relevant articles. Study characteristics were extracted from the included articles by two researchers (VU and AB). Any disagreement about the selection of studies, or extraction of data, was resolved by discussion with one researcher (DK). Authors were contacted for additional information where needed. The following data were extracted from the studies: first author, publication year, location, study population, study design, number of participants (intervention and controls), cognitive decline or impairment classification criteria (if any), age (mean and SD), percentage of females included, and intervention characteristics (type, frequency, intensity [if reported], session duration, intervention duration), pre/post-intervention cognitive outcome data (mean, SD), and summary of brain function findings (between- and within-group effects, and p values). Two researchers (DK and GZS) categorized domain-specific cognitive performance according to the Cattell-Horn-Carroll and Miyake framework (Webb et al., 2018), and consensus was reached with a third researcher (IHKL).

Cognitive outcome data are presented as effect sizes with 95% confidence intervals (CI). Where data from ≥3 studies were available for a given cognitive domain, random-effects meta-analyses were conducted using Comprehensive Meta-Analysis Version 3 software (Biostat Inc., Englewood, NJ, USA). Significance was set at p < 0.05. Effect sizes were calculated from pre- to post-intervention scores between two groups for change in cognition outcomes such as global cognition, executive function, processing speed, and general short-term memory and expressed as Hedge's g with 95% CIs around the estimated effect size. Absolute between-study heterogeneity was calculated using tau (τ). If post intervention scores were not available, the absolute or relative mean change scores were used to calculate effect sizes. Where there were two comparisons from the same study for a particular outcome (e.g., an aerobic exercise group and dancing group), the sample size of the control group was halved.

Statistical heterogeneity between studies was quantified using Cochran's Q and I² statistics, both of which provide estimates of the degree of heterogeneity resulting from between-study variance, rather than by chance. Cochran's Q with p < 0.05 was classified as significant heterogeneity, and I² of more than 75% was considered to indicate high level heterogeneity, I² of 25%−75% as indicative of substantial heterogeneity, and an I² < 25% as low heterogeneity. Publication bias was tested using the Begg and Mazumdar test, with a p < 0.05 suggesting the presence of bias (Sterne et al., 2019). Where significant bias was detected, a Duval and Tweedie trim-and-fill analysis (Duval and Tweedie, 2000) was conducted to re-calculate the pooled effect size after removing any studies which may introduce publication bias (i.e., small studies with large effect sizes from the positive side of the funnel plot). Brain function effects were narratively analyzed due to heterogenous outcomes and insufficient data available for meta-analysis.

Study quality was assessed independently by two researchers (TT and AB) using a modified Downs and Black checklist (Downs and Black, 1998). Reporting quality is assessed with 10 items spanning the aims, participant characteristics, and findings; external validity is checked with 3 items concerning representativeness; internal validity for bias and confounding/selection bias with 4 and 6 items respectively. The scale was modified to include criteria on exercise supervision, whereby if a study was supervised, a “yes” was given. If an item was unable to be determined a “no” was given. Scores were compared and disagreements resolved by a third reviewer (AS).

Studies were assessed for bias independently by two reviewers (TT and RLC) using the Cochrane Risk of Bias 2 tool (Sterne et al., 2019). Included studies were evaluated as having ‘low,' ‘high,' or ‘unclear' risk of bias using the following domains: randomization, allocation concealment, blinding of intervention instructors, blinding of participants, blinding of outcome assessors, handling of incomplete data, selective reporting, and other risk of bias pertaining to exercise adherence. Discrepancies in ratings were discussed and resolved with one additional reviewer (AS). Studies were not excluded based on their bias assessment.

Figure 1 represents the study selection process for this review. A total of 5,204 records were identified with 4,477 remaining after duplicates were removed. Fifteen articles met eligibility criteria for this systematic review, representing 12 unique interventions.

The characteristics of the fifteen articles are shown in Table 2. There were three multi-paper interventions: two articles reported on different brain function outcomes from the one intervention (Zhu et al., 2018; Qi et al., 2019), two papers reported on the same intervention outcomes at different time-points (Suo et al., 2016 for intervention effects; Broadhouse et al., 2020 for follow-up), and two publications reported on different aspects of the same brain function measure from the one intervention (Tao et al., 2019; Lin et al., 2023). As we were interested in intervention effects here, the follow-up findings from Broadhouse et al. were excluded from further analysis, but are reported in Table 2 for interested readers. Further, as Lin et al. and Tao et al. referred to the same sample between their studies, only Tao et al.'s data was used in the meta-analysis. All articles were published between 2012 and 2023, with five studies located in China, two each in Canada and Australia, and one each in Israel, Pakistan, Taiwan, United States of America, and South Korea. Nine of the twelve unique interventions employed a two-group parallel arm design and the remaining four trials involved three groups where two exercise-related interventions were compared to a control.

A total of 646 participants with cognitive frailty, MCI, or VCI were included and their baseline characteristics are summarized in Table 2. For multi-paper interventions, the number of participants from the main study (i.e., Suo et al., 2016; Zhu et al., 2018; Tao et al., 2019) was used in the calculation for total participants included in this review. There were no studies involving people classified with SCD.

Population classification criteria varied between the studies with published guidelines being used in all but one trial (Zhu et al., 2018), and a MoCA or MMSE cut-off score being used in all but two studies (Xia et al., 2019; Yogev-Seligmann et al., 2021). MoCA cut-off scores were consistently applied at ≤ 26 while the MMSE cut-off was inconsistently used: <25 (Amjad et al., 2019), >20 (Hsu et al., 2018), >24 (Suo et al., 2016; Zhu et al., 2018; Qi et al., 2019; Tsai et al., 2019; Tomoto et al., 2021). Mean age ranged from 58 to 78 years, and female representation in samples ranged from 38.4 to 100%.

The last two columns of Table 2 summarize the exercise intervention characteristics and outcome measures relevant to this review. Of the nine two-arm trials, six employed aerobic exercise at a moderate intensity (Hsu et al., 2018; Zhu et al., 2018; Amjad et al., 2019; Qi et al., 2019), two used a resistance training protocol (Suo et al., 2016; Hong et al., 2018), and one implemented the mind-body exercise of Baduajin (Lin et al., 2023). In the three-arm trials, aerobic and resistance training protocols were used in two studies (Nagamatsu et al., 2012; Tsai et al., 2019), while aerobic and mind-body exercise regimes were employed in the other two studies (Tao et al., 2019; Xia et al., 2019). Most interventions were compared to usual care or lifestyle, while three employed sham exercise (Nagamatsu et al., 2012; Suo et al., 2016; Amjad et al., 2019), and a further four involved a concurrent treatment of health education across all groups (Hsu et al., 2018; Tao et al., 2019; Xia et al., 2019; Lin et al., 2023). Exercise frequency ranged from 2 to 3 sessions per week and intervention duration ranged from 6 to 24 weeks, with the most common being 24 weeks.

Supervision was employed in all aerobic exercise trials, and this was conducted through walking (Nagamatsu et al., 2012; Hsu et al., 2018; Xia et al., 2019), dancing (Zhu et al., 2018; Qi et al., 2019), or on a treadmill or stationary bicycle (Amjad et al., 2019; Tao et al., 2019; Yogev-Seligmann et al., 2021) for a duration of 20–60 minutes. Intensity varied between 60–90% of max HR (Zhu et al., 2018; Amjad et al., 2019; Qi et al., 2019; Tomoto et al., 2021) or 50–80% of HR reserve (Nagamatsu et al., 2012; Hsu et al., 2018; Tao et al., 2019; Tsai et al., 2019; Xia et al., 2019; Yogev-Seligmann et al., 2021). Resistance training protocols prescribed exercises at an intensity >65%1RM and adhered to the principles of progressive overload using an elastic band for 40 min (Hong et al., 2018) or a machine and free weights for 40–90 min (Nagamatsu et al., 2012; Suo et al., 2016; Tsai et al., 2019). Three interventions employed the mind-body exercise of Baduanjin for 60 min, three times per week (Tao et al., 2019; Xia et al., 2019; Lin et al., 2023). Seven trials reported warm-up and cool-down periods lasting 5–10 min (Hong et al., 2018; Hsu et al., 2018; Zhu et al., 2018; Amjad et al., 2019; Qi et al., 2019; Tsai et al., 2019; Tomoto et al., 2021; Yogev-Seligmann et al., 2021).

As shown in Table 3, four cognitive domains were assessed across the included studies: global cognition (as measured by the MoCA, MMSE, and ADAS-Cog), executive function (most commonly assessed using Trail Making Test B), processing speed (most commonly measured with the Trail Making Test A), and general short-term memory (assessed with Digit Span measures). The pooled effects of exercise vs control groups for these cognitive outcomes are summarized in Figure 2.

The results of the methodological quality assessment are presented in Table 4. The scores ranged from 13–20, with an average of 16.9. Seven studies were classified as being of moderate quality (Hong et al., 2018; Zhu et al., 2018; Amjad et al., 2019; Tao et al., 2019; Tsai et al., 2019; Tomoto et al., 2021; Yogev-Seligmann et al., 2021), and one was considered to be of high quality (Hsu et al., 2018). All studies did not blind their intervention groups due to the nature of the exercise trials. Most articles (n = 11) adequately reported their aims, outcomes, patient characteristics, interventions, p-values, and variabilities.

Six trials reported the adherence to exercise sessions, with an average adherence rate of ≥54% (Nagamatsu et al., 2012; Suo et al., 2016; Hsu et al., 2018; Tao et al., 2019; Tsai et al., 2019; Tomoto et al., 2021; Yogev-Seligmann et al., 2021). Furthermore, all included trials provided supervised exercise sessions. Three studies reported adverse effects, which included acute episodes of shortness of breath and non-injurious falls in one study (Nagamatsu et al., 2012), and no adverse effects in the other studies (Zhu et al., 2018; Yogev-Seligmann et al., 2021). Seven studies reported attempted to blind those measuring cognitive outcomes of the exercise interventions (Suo et al., 2016; Hsu et al., 2018; Zhu et al., 2018; Amjad et al., 2019; Qi et al., 2019; Tomoto et al., 2021; Lin et al., 2023).

The results of the risk of bias assessment are summarized in Figure 3. Five studies scored an unclear or high risk of bias on six or more domains (Nagamatsu et al., 2012; Hong et al., 2018; Amjad et al., 2019; Tsai et al., 2019; Xia et al., 2019). Two studies scored an unclear or high risk of bias on five domains (Qi et al., 2019; Yogev-Seligmann et al., 2021), while six studies scored an unclear or high risk of bias on four or less domains (Suo et al., 2016; Hsu et al., 2018; Zhu et al., 2018; Tao et al., 2019; Tomoto et al., 2021; Lin et al., 2023).

According to our analyses, exercise leads to moderate and significant improvements in global cognitive and executive functions, and these are driven by interventions of aerobic nature. The cognitive benefits of aerobic exercise (high- and low-frequency) have been reported in previous meta-analyses and reviews extending to those who have AD and non-AD dementias, MCI, and SCD (Baker et al., 2010; Groot et al., 2016; Zheng et al., 2016; Cammisuli et al., 2017). Studies utilizing strength-based exercise for cognition in older adults, however, are lacking. Only four studies included in this meta-analysis employed resistance strength-based training. Further, none of the studies explored the effects of resistance training on processing speed. Reduced muscular strength has been associated with advancing age (Lauretani et al., 2003), greater brain atrophy, white matter disease, and poorer cognitive function (Clark and Taylor, 2011; Kilgour et al., 2014; Herold et al., 2019).

Growing evidence shows that resistance exercise evokes meaningful benefits in functional brain changes (particularly the frontal lobe), reductions in white matter lesions and decreased atrophy across age-groups (Herold et al., 2019). For example, in older sedentary males, a 6-month study presented that resistance training improved memory performance (Cassilhas et al., 2007). In community-dwelling older women, progressive resistance exercise significantly improved executive function by up to 12.6% (Liu-Ambrose et al., 2010). Further, at 2-year follow-up within that same cohort, both bi-weekly and weekly resistance exercise increased peak muscle power, maintained executive functions, promoted memory, and reduced white matter atrophy (Best et al., 2015).

In an MCI population that compared progressive resistance training (PRT) with computerized cognitive training (CCT), the authors found that the PRT but not CCT significantly increased gray matter in the posterior cingulate, reduced white matter hyperintensity, and improved global cognition at 6-months follow-up (Suo et al., 2016). From our analyses, pooled effects for short-term memory were moderate, but non-significant. However, these marginal improvements in short-term memory seem to be driven by resistance training, rather than aerobic exercises, which has not been reported elsewhere. Together, this suggests that resistance exercise has cognitive benefits independent of aerobic exercise. Previous research has demonstrated plausible biological mechanisms involving homocysteine and insulin-like growth factor 1 (Liu-Ambrose and Donaldson, 2009). As an interesting aside, it has been claimed by an older meta-analysis that the greatest benefits on cognition in older adults could occur when aerobic exercise was combined with resistance training (Colcombe and Kramer, 2003). Regardless, more clinical trials examining the role of resistance strength-based exercise independently or combined with aerobic exercises are required to explore the maximal benefits on cognitive functions and associated physiological changes.

Regional differences throughout the brain were reported across studies utilizing fMRI to assess central mechanisms of exercise, with few commonalities (in study design and subsequently results), making synthesis of results challenging. Task-based fMRI assessments revealed exercise-related changes in distinct brain areas, with two of the three studies respectively correlating right lingual gyrus and superior temporal gyrus changes with improved cognitive performance (Nagamatsu et al., 2012; Hsu et al., 2018). The most pronounced resting state changes observed were in the temporal lobes, with 12–26 weeks of aerobic, resistance, and mind-body exercises altering the hippocampal region (Suo et al., 2016; Qi et al., 2019; Tao et al., 2019). While most studies found non-significant correlations between the regional changes in brain activity and cognitive performance (Suo et al., 2016; Qi et al., 2019; Xia et al., 2019), improved MoCA scores were associated with Baduanjin-related changes in resting state anterior cingulate activity (Tao et al., 2019; Liu et al., 2021). These findings are likely due to increased regional metabolic activity (Shimada et al., 2017) and may reflect exercise-induced neuroplastic changes as observed in other populations (Woodward et al., 2020), potentially providing a buffer against pathophysiological and age-related atrophy (Bugg and Head, 2011) in these vulnerable regions.

In AD, the resting state EEG signature is dominated by slow wave activity (i.e., increased delta and theta magnitude) and reduced fast wave activity (alpha and beta magnitude) (Babiloni et al., 2021), and a similar pattern has been reported in MCI (Meghdadi et al., 2021) (albeit less consistently), particularly those who have an increased risk of AD (Jelic et al., 2000). However, in healthy aging, the opposite is observed, that is, reduced delta and theta spectral power (Vlahou et al., 2014; Barry and De Blasio, 2017). Across included studies, 12–16 weeks of exercise (both aerobic and resistance) led to reduced delta and theta power in MCI (Hong et al., 2018; Amjad et al., 2019), potentially indicating a normalization of the EEG spectra and a shift away from the AD trajectory. Exercise differentially modulated alpha, with aerobic training reducing alpha-2 power (11–14 Hz) (Amjad et al., 2019) and resistance training increasing broad-band alpha (8–12 Hz) power (Hong et al., 2018). Hong's finding may have more therapeutic relevance here given that reduced alpha power (and peak frequency) is reported in MCI/AD, and the wider alpha range may have detected this. Further, aerobic exercise appears to promote neural adaptability that is measurable at a global scale. Aerobic exercise increased EEG complexity (Amjad et al., 2019), a signal proposed to reflect the underlying integrity of neuronal circuitry and its capacity to rapidly adapt to environmental changes (Bosl et al., 2011). This is further reflected in the ERP effects of aerobic and resistance exercise whereby reduced P300 latency and increased P300 amplitude indicated faster stimulus evaluation processing and enhanced neurocognitive efficiency (Steiner et al., 2013; Zhu et al., 2018; Tsai et al., 2019).

Increased CBF has been suggested as a potential mechanism underpinning the positive effects of exercise on cognitive function (Renke et al., 2022). The exact reason for this is less clear but these effects may arise from the association between neuronal metabolism and CBF and perfusion (i.e., neurovascular coupling), changes in cerebral blood vessels including elasticity and density (i.e., angiogenesis), and enhanced activity of pericytes (Bolduc et al., 2013; Barnes et al., 2021). From the two CBF studies identified in this review, aerobic exercise was found to increase blood flow velocity in the internal carotid artery (Tomoto et al., 2021), right middle cerebral and basilar arteries (Lin et al., 2023), and end diastolic velocity in the middle cerebral artery (Tomoto et al., 2021) and basilar artery (Lin et al., 2023). In the middle cerebral artery, reductions were also seen in pulsatility (Tomoto et al., 2021) and peak systolic velocity (Lin et al., 2023). No associations were found with cognitive function were found in the one study that assessed for them (Tomoto et al., 2021) so we are currently unable to infer whether these changes in arterial CBF are associated with cognitive improvements related to exercise.

A small number of studies met eligibility criteria pertaining to intervention characteristics and outcome measures, limiting the implications of this review. Most notably, there were no SCD studies. This is unsurprising given that classification criteria were operationalized in 2017 (Molinuevo et al., 2017), just as brain function measures emerged as outcomes of interest in exercise trials. This limited the scope of our review, its findings, and implications to objective classifications of cognitive impairment, which are respectively considered to represent the prodromal stages of AD and vascular dementia. These findings were also hindered by the small sample sizes, lack of reported adherence to exercise sessions (and where this was reported in two studies, the average adherence rate was 75%), moderate methodological quality, and high risk of bias in most included studies. We also did not exclude studies based on intervention from any pooled analysis. The pooled analyses were undertaken to determine the effect of exercise per se (which can include aerobic exercise or resistance training), to maximize the sample size. Further studies can explore the relative effects of individual exercise prescription variables such as modality, intensity, and volume. These review outcomes speak to the paucity and novelty of brain function research regarding exercise for dementia prevention. Any neuronal mechanistic inferences regarding exercise and cognition are thus speculative.

As reviewed here, exercise can significantly improve global cognitive and executive functions in people with cognitive impairment and increased dementia risk. Regardless of modality (aerobic, resistance, or mind-body), at least 12 weeks of exercise can reverse brain activity signatures of cognitive decline in people with cognitive impairment. We speculate that the cognitive improvements associated with exercise are likely driven by increased metabolic activity, cerebrovascular mechanisms (such as neurovascular coupling), and neuroplasticity throughout the brain, but particularly in highly sensitive and plastic regions in the frontal and temporal lobes. These hypotheses require further investigation with higher quality randomized controlled trials. Figure 4 integrates these central impacts of exercise alongside other functional changes together with molecular and physiological mechanisms established in prior reviews (mentioned in the introduction), highlighting the multitargeting effects of exercise for dementia risk reduction. As the field of dementia prevention grows, we encourage researchers to include more clinical trials investigating the role of resistance strength-based exercise independently or combined with aerobic exercises in this at-risk population. This is to identify the maximal benefits of different exercise regimes on cognition and physiological changes. Furthermore, as highlighted here, future research should integrate both cognitive and brain function measures to help elucidate the mechanisms underpinning the positive effects of exercise.