Following PRISMA guidelines [25], two investigators (Shengyu Dai [SD] and Shunling Yuan [SY]) independently conducted systematic searches in four international databases (PubMed, Web of Science, Embase, and Cochrane Library) and two Chinese databases (China National Knowledge Infrastructure [CNKI, https://www.cnki.net/] and VIP Database [VIP, https://qikan.cqvip.com/]). The search was limited to randomized controlled trials (RCTs), and each database was searched from inception to March 31, 2025. The systematic literature search employed Medical Subject Headings (MeSH) to identify core search terms. MeSH terms including “Exercise”, “Alzheimer’s Disease”, and “Cognitive Function” were selected and a comprehensive search strategy was implemented that combined both MeSH subject terms and free words for term retrieval (see Supplementary Table 1). To ensure comprehensive coverage of the literature, we also supplemented this search strategy by manually screening reference lists of identified meta-analyses and published reports to identify additional studies relevant to the topic. The review protocol was registered with the International Prospective Register of Systematic Reviews (PROSPERO), registration number CRD420250631766.

Inclusion criteria: (1) Population: patients with a confirmed diagnosis of AD meeting standardized diagnostic criteria; (2) Intervention: the intervention group received PA interventions (e.g., aerobic exercise, cycling, or walking) without restriction on the type of activity; (3) Control: the control group received conventional rehabilitation, or no intervention; (4) Outcomes: cognitive impairment was assessed using MMSE scores; and (5) Study design: only RCTs were included.

Exclusion criteria: (1) Patients without AD; (2) The intervention group received multimodal therapeutic interventions (e.g., cognitive-behavioral interventions); (3) Non-RCTs; (4) Review articles or meta-analyses; (5) Low-quality studies or duplicate publications; and (6) Studies with unavailable full-text or unextractable data.

Two researchers independently screened article titles and abstracts, with data extraction conducted using standardized forms. In the event of a discrepancy, the two researchers discussed or consulted with a third investigator to facilitate resolution. The following data were extracted: title, first author, publication date, country (region) of the study, age of participants, number of participants, diagnostic criteria, duration of disease, specific intervention measures, and outcome measures.

The risk of bias assessment method recommended in the Cochrane Handbook 5.1.0 [26] was strictly adhered to by both independent researchers when evaluating each included study. Assessment items included random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, handling of incomplete outcome data, selective reporting, and other potential biases. Each item was classified as low-risk, high-risk, or unclear-risk. In case of disagreement between the two researchers, a third investigator was consulted to arbitrate and reach a consensus. This independent dual-review with third-party arbitration ensured the objectivity of assessment.

All analyses were performed using Stata software (Version 15.0, StataCorp LLC, College Station, TX, USA; Official website: https://www.stata.com/). Since the scoring scales used in the combined trials (such as MMSE) were consistent, we employed weighted mean difference (WMD) and 95% confidence intervals (CI) to analyze continuous variable data. A fixed-effects model was applied when an I² value of 50% indicated low heterogeneity among studies. If the I² value was $\ge$50%, a random-effects model was used. Subgroup analyses were subsequently conducted to identify potential factors contributing to high heterogeneity. These were stratified by the following variables in the order reported: (1) country (China vs. other countries); (2) intervention intensity (moderate vs. non-moderate); (3) single session duration (30 min, 40 min, 60 min); (4) intervention frequency (2 times/week, 3 times/week, $>$3 times/week); (5) total weekly intervention time (120 min/week, 120 min/week, $>$120 min/week); and (6) total intervention duration ($\le$12 weeks vs. $>$12 weeks). For the country-based subgroup analysis, studies were categorized into “China” and “non-China”. This categorization was based on the distribution of included studies (7 from China and 6 from other countries) and aligns with research by Jia et al. [27]. Meta-regression analysis was performed to evaluate the specific influence of these covariates on the outcomes. Egger’s test and Begg’s test were also conducted to ensure the robustness of the findings, with p 0.05 indicating the presence of bias.

We systematically searched six databases, retrieving 5691 studies. After initial screening and removal of duplicates, 3727 articles remained. Of these, 3583 were excluded following assessment of their relevance based on the title and abstract. Next, the screening process strictly followed the predefined inclusion and exclusion criteria, with 131 studies undergoing full-text review and evaluation. This resulted in 118 articles being excluded for the following reasons: inconsistency in experimental design (n = 36), no available data (n = 39), non-RCT (n = 25), and systematic review (n = 18). Eventually, 13 RCTs were included in the final analysis (Fig. 1).

Fig. 1.

Flowchart of the literature screening process. CNKI, China National Knowledge Infrastructure; VIP, VIP Database.

Following the rigorous screening process, 13 RCTs [28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40] published between 2010 and 2024 and including a total of 813 AD patients were identified (Table 1). Among these, seven were conducted in China, three in Brazil, and one each in Germany, Denmark, and France. All studies reported baseline characteristics including age, sample size, and diagnostic criteria. However, six studies failed to document disease duration. The intervention duration ranged from 4 to 48 weeks across studies. Primary outcomes were assessed using the MMSE.

Two researchers independently evaluated the methodological quality of the included studies using the Cochrane Handbook 5.1.0 criteria [26], with any discrepancies resolved through third-party arbitration. The majority of the 13 RCTs were of moderate methodological quality (Fig. 2). However, the proportion of “unclear” ratings for allocation concealment and participant/researcher blinding was relatively high. These ratings stemmed from inadequate reporting in the original studies. Attempts were made to contact the authors of studies with missing key information, but no responses were received. To avoid speculative judgments, these items remained categorized as “unclear”.

Fig. 2.

Schematic diagram of risk bias. “?”, unclear-risk; “+”, low-risk; “-”, high-risk.

Through systematic screening of the literature, we identified 13 RCTs that assessed cognitive function using the MMSE. Analysis of the results from these studies revealed a significant improvement in MMSE scores among the PA intervention group, indicating that it can effectively enhance cognitive function in AD patients. Subgroup and meta-regression analyses revealed that sources of heterogeneity were the country in which the study was performed, and the intervention frequency. Specifically, the subgroup analysis showed that MMSE scores were significantly improved by interventions with moderate intensity, interventions with a single session duration of 30 min, an intervention frequency of $>$3 times per week, a total weekly intervention time of $>$120 min, and a total intervention duration of $>$12 weeks. Meta-regression analysis also revealed significant interactions among intervention intensity, intervention frequency, single session duration, and total intervention duration. Finally, the multivariate model found that intervention frequency and total intervention duration were significantly correlated with improved cognitive function, while the intervention intensity and single session duration showed a negative correlation.

This study specifically examined the effects of PA interventions on cognitive function in patients with a clinical diagnosis of AD. Our findings showed that PA significantly improves cognitive function in AD patients, in line with the results of previous meta-analyses. Du et al. [41] conducted a meta-analysis of cognitive function in 869 patients diagnosed with AD. Their results also showed a statistically significant improvement in cognitive function after PA interventions. However, five of the included studies in their meta-analysis did not show beneficial effects. Jia et al. [27] conducted a meta-analysis of 13 RCTs involving 673 subjects diagnosed with AD. As in the present study, these authors used subgroup analysis to explore a possible dose-response relationship between PA interventions and improvement in cognitive function. Based on the MMSE scores, they found a statistically significant cognitive enhancement in the intervention group, with different doses of PA producing different effects. A systematic meta-analysis by Yang et al. [42] demonstrated that aerobic exercise significantly improved cognitive function in AD patients. Subgroup analyses revealed that patients showed significant improvements in both MMSE and Alzheimer’s Disease Assessment Scale-Cognitive Subscale (ADAS-cog) scores when the total intervention duration exceeded 16 weeks.

These studies suggest that PA has a positive effect on cognitive function. However, most previous studies did not apply internationally recognized diagnostic criteria (e.g., National Institute of Neurological and Communicative Disorders and Stroke-Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA), Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV), or International Classification of Diseases, Tenth Revision (ICD-10)) when enrolling AD patients. This inconsistency in diagnostic standard can increase the heterogeneity of study outcomes, potentially compromising accurate assessment of the intervention effect. Moreover, previous meta-analyses typically relied only on subgroup analysis to explore potential influencing factors when heterogeneity was detected, without further investigating the specific effect of these factors on effect sizes. In contrast, the current study only enrolled patients who met established clinical diagnostic criteria for AD, thereby ensuring homogeneity of the study population. In addition to subgroup analysis, we also employed meta-regression to systematically evaluate the influence of intervention parameters on effect sizes. As a result, our study effectively controlled for bias arising from inconsistent diagnostic criteria, providing more targeted evidence for PA interventions in AD patients.

Using subgroup analysis, we systematically evaluated the effects of the following factors on cognitive function in AD patients: country of the study, intervention intensity, single session duration, intervention frequency, total weekly intervention time, and total intervention duration. This analysis revealed subtle differences in treatment effects between studies conducted in China and those in other countries, in line with previous findings by Jia et al. [27] and Yang et al. [42]. However, unlike their findings, the results of our meta-regression analysis revealed that differences in the country of study had a significant impact on intervention effects. In the five sets of bivariate meta-regression analyses, the country of study variable consistently maintained significant correlations, indicating that national differences are an important factor influencing intervention outcomes. In particular, differences between countries were most significant after adjusting for the intensity of intervention, suggesting this may be the most prominent regulatory factor responsible for the observed differences in treatment effects. Although we adjusted for several key variables in the intervention protocol, most studies did not explicitly report differences in environmental factors (e.g., level of medical infrastructure, cultural acceptance) and demographic characteristics (e.g., dietary habits, ethnicity/race), and consequently our findings may still be affected by residual confounding factors. Future research should aim to systematically collect relevant background data to further explore the association between potential environmental/demographic factors and the intervention effects.

With regard to the intervention intensity and based on the American College of Sports Medicine (ACSM) guidelines, moderate intensity was defined as 64–76% of Maximum Heart Rate (HRmax) [43]. The current results showed the intervention effect for the moderate intensity group was better and more consistent (i.e., without heterogeneity) compared to the other intensities. In contrast to previous meta-analyses by Yang et al. [42] and Zhang et al. [44], our study is the first to incorporate classification by intervention intensity. Although the results were statistically significant, certain differences were apparent in the definition of intervention intensity among the different studies. Some studies used 60–80% of HRmax as the standard [28, 35, 39], while others did not clearly report the intensity [31, 32, 36, 40]. This inconsistent methodology may be the main reason for the high heterogeneity observed in the non-moderate intensity group in our analysis. Differences in the classification of moderate intensity were also noted. Most studies used percentage of HRmax as the evaluation criterion [30, 33, 34, 38], while others employed VO₂max [29], and some described the intensity as “moderate” without precise quantification [37]. Therefore, a standardized intensity assessment system that utilizes objective indicators is necessary to determine the optimal intensity threshold for improving cognitive function in AD patients.

Our study found that interventions with a single session duration of 30 min and a frequency of $>$3 times per week were associated with significantly improved cognitive function. Previous studies also consistently found that 30 min interventions could improve MMSE scores [27, 44]. However, these studies found that interventions with a frequency of 3 times per week or fewer were more effective at improving cognitive function than $>$3 times per week. A possible reason for this discordance is that all 3 studies with an intervention frequency of $>$3 times per week were conducted in China, which may have biased the results of the present analysis towards high-frequency intervention protocols. A previous study reported that PA interventions with a frequency of 3–4 times per week can improve working memory in older adults, but that 5 or more PA interventions per week resulted in no significant improvement [45]. Studies by Venturelli et al. [46] and Wei et al. [40] support the effectiveness of high-frequency interventions by showing that aerobic exercise 5 times per week or daily Baduanjin practice, respectively, improved cognitive function in AD patients. According to joint guidelines from the ACSM and American Heart Association (AHA), for optimal health benefits, older adults should aim for either five 30 min (or longer) sessions of moderate aerobic exercise per week, or at least 20 min of vigorous aerobic exercise 3 times per week [47]. However, our study found that $>$3 PA interventions per week may have additional benefits in terms of cognitive improvement. This difference may be related to the pathological characteristics of AD patients, with high-frequency interventions ($>$3 times per week) promoting neuroplasticity and cerebral blood flow regulation more effectively through cumulative effects [48]. Moreover, a 30 min session duration can effectively stimulate the regulation of bodily functions and changes in neural plasticity, while avoiding increased cognitive load and inhibition of the brain’s adaptive capacity caused by excessive physical fatigue [45]. A shorter duration of PA can more effectively increase brain-derived neurotrophic factor (BDNF) expression and neuroplasticity, thus aligning better with the physiological constraints of AD patients. Kim et al. [49] showed that a single session duration of 60 min can lead to fatigue accumulation in AD patients due to attention deficits and physical limitations, whereas interventions with a single session duration of 30 min align better with their tolerance levels. An animal study conducted by Zhang et al. [50] provided further evidence that short-duration, high-frequency PA can more effectively promote hippocampal neurogenesis. In view of the limited attention-sustaining ability of AD patients, the application of single sessions with a duration of 30 min and for $>$3 times per week may be more in line with the physiological tolerance level of such patients.

Our subgroup analysis revealed that a total weekly intervention time of $>$120 min conferred the largest effect size for improvement of MMSE scores. This finding differs from the meta-analysis conducted by Jia et al. [27], who suggested that interventions of 120 min or less per week had a greater effect. The World Health Organization recommends that cognitively healthy adults aged 65 and older should engage in at least 150 min of moderate-intensity aerobic exercise weekly, or 75 min of vigorous-intensity aerobic exercise weekly [51]. However, our results were obtained with intervention times slightly below these guideline-recommended exercise standards, possibly due to AD patients having neurodegenerative pathology and a lower tolerance for prolonged exercise.

Our study also found that a total intervention duration of $>$12 weeks significantly improved MMSE scores compared to durations of 12 weeks or less. This result agrees with that of de Sá Leitão et al. [52], who similarly found that 12 weeks of PA can enhance the cognitive level of AD patients, but differs slightly from the findings reported by Jia et al. [27] and Yang et al. [42]. The latter authors found more significant improvements in MMSE scores when the total intervention duration exceeded 16 weeks, suggesting there may be a critical threshold for an intervention duration of approximately 12–16 weeks in total. Only when this threshold is reached or exceeded can the intervention effect be significantly improved or stabilized.

Of note, most covariates in our study were found to be non-significant in univariate models, but significant in the multivariate model. This discrepancy is likely to stem from the interconnectedness of intervention parameters. The multivariate model revealed that intervention frequency and total intervention duration were independent predictors of improved cognitive function, while intensity and single session duration exert independent negative effects. These associations were masked in univariate analyses, probably due to their interaction with other parameters, highlighting that PA interventions in AD patients require comprehensive consideration of multiple interacting parameters.

A dose-response relationship exists between PA and health benefits. An appropriate exercise load can not only mitigate the degradation of physiological function caused by aging, but also reduce the risk of various chronic diseases [53]. Another study found an inverted U-shaped relationship between PA and cognitive function [54]. PA is a crucial stabilizing factor that can slow the decline of cognitive functions [55]. A large number of studies have shown that PA can enhance cognitive function in elderly patients with AD [28, 29, 56]. A study by Machado et al. [55] confirmed that regular PA can enhance neural function in the brain. Tyndall et al. [57] found that long-term PA significantly reduces the risk of developing AD and exerts positive regulatory effects on related blood biomarkers. A systematic analysis concluded that PA had moderate positive effects on neurocognitive functions, such as improving attention, accelerating the speed of information processing, enhancing executive function, and improving memory [58]. Finally, a systematic review by Li et al. [59] also supports the efficacy of PA.

A study in transgenic AD mice demonstrated that PA can reduce extracellular A$\beta$ deposition in neurons of the prefrontal cortex and hippocampal formation, thereby improving their cognitive function, learning, and memory abilities [60]. Another study in a rat model of AD also showed that PA can mitigate the apoptosis of posterior hippocampal neurons, promote the growth and survival of neuronal processes, and improve cognitive function [61]. Nerve growth factor has a significant impact on nerve regeneration, synaptic regeneration and neurotransmitter production. Aerobic exercise can activate the nerves of adult rats, enhance neural plasticity, alleviate the degenerative changes, and reduce oxidative stress levels in the hippocampus, ultimately improving learning and memory impairment in a rat model of AD, as well as cognitive function [62].

Several mechanisms might explain the positive effect of PA on cognitive function in patients with AD: (1) PA can stimulate the release of calcium-metabolizing hormones, thereby increasing the calcium ion level in the blood. This subsequently promotes dopamine synthesis in the brain, while increasing the density of acetylcholine receptor. Ultimately, elevated dopamine levels can regulate the function of the cholinergic system, thus contributing to the improvement of cognitive function [63, 64]. (2) Cardiovascular-related risk factors are closely associated with cognitive decline. PA can improve cardiovascular function and alleviate inflammatory response and oxidative stress, thus exerting a protective effect on neurons. Concurrently, PA can enhance and redistribute cerebral blood flow, accelerate the degradation of $\beta$-amyloid protein, and increase angiogenesis and neurotransmitter metabolism. These mechanisms collectively optimize the brain environment, thereby improving cognitive function in AD patients [65, 66, 67]. (3) The hippocampus, a key brain region for memory and learning, is the most severely affected area in AD patients. PA can increase BDNF, stimulate neurogenesis in the dentate gyrus, and enhance synaptic plasticity [68]. In addition, PA reduces neuronal apoptosis, increases synaptic density, and enhances the brain’s ability to cope with pathological damage, thereby improving cognitive reserve and delaying the progression of AD symptoms [69].

This research has several strengths. First, it only included RCTs, thereby improving the quality of the study and the reliability of the results. Second, the study population was limited to AD patients who met standardized diagnostic criteria, ensuring consistency in patient disease characteristics. Third, a multi-subgroup design was employed to explore the reasons for the high heterogeneity and for the effects between PA and cognitive function in different AD patient subgroups. Finally, our study used multivariate regression analysis to quantify significant interactions between intervention parameters (frequency, intensity, session duration, and total intervention duration). However, this study also has several limitations. Firstly, it was limited to the English and Chinese language literature. Second, the included studies did not use blinding of participants during the intervention process, and hence there may have been some bias in the study design. Third, the lack of consistency in the intervention protocols and the use of different criteria across studies may have influenced the final results and increased the heterogeneity of studies. Fourth, the possible effects of potential confounding factors could not be controlled. For example, some studies did not report both disease severity and disease duration. Fifth, included studies used varying diagnostic criteria (DSM-IV, NINCDS-ADRDA, ICD-10). Because only one study used ICD-10 and the others used overlapping criteria, it was not feasible to quantify this effect using robust subgroup or sensitivity analyses. Future research should adopt a unified, standardized criterion to reduce such heterogeneity. Sixth, some subgroup analyses included only three studies. This small sample size reduces the statistical power, thereby affecting the reliability of effect size estimates. Finally, we have noted that meta-regression is an ecological analysis, and individual-level confounding cannot be excluded. Additionally, adherence to interventions and differences in implementation across studies may introduce heterogeneity that is difficult to control for in meta-analysis.