Liraglutide and Exenatide may affect thinking skills in Alzheimer's and mild memory problems: A review and combined analysis

This systematic review and meta-analysis followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) statement [15] and the MDPI Pharmaceutics reporting standards for systematic reviews.

The protocol was prospectively registered in PROSPERO (CRD420251143171). The original registration specified broad population eligibility, including adults with AD or prodromal AD, and allowed inclusion of studies with or without type 2 diabetes mellitus. Although the protocol began with a broader proposition, full-text screening identified randomized controlled trials with comparable populations, interventions, and cognitive endpoints, which permitted conducting a prespecified random-effects meta-analysis, making the inferences from our results more comprehensive and cohesive. This methodological modification was made prior to data extraction and follows PRISMA 2020 recommendations on adaptive synthesis.

Eligible studies included RCTs with either placebo or no-treatment control groups. Open-label RCTs were included when random allocation and prospective outcome assessment were clearly reported. Eligible studies included adults aged ≥50 years or older with clinically or biomarker-confirmed AD or MCI due to AD, diagnosed according to established criteria such as the National Institute of Neurological and Communicative Disorders and Stroke, the Alzheimer's Disease and Related Disorders Association (NINCDS-ADRDA), or the National Institute on Aging, Alzheimer's Association (NIA–AA). Although broader inclusion criteria would have increased the number of eligible trials, the population was intentionally restricted to clinically or biomarker-confirmed AD/MCI to minimize etiological heterogeneity and avoid indirect inference from metabolically driven or mixed dementia populations. Interventions consisted of GLP-1 RAs, specifically liraglutide or exenatide, administered at approved therapeutic doses for a minimum duration of 12 weeks. The comparator was a placebo. The primary outcome was the change in global cognitive performance from baseline to study end, assessed using validated instruments including the ADAS-Cog, MMSE, or Montreal Cognitive Assessment (MoCA). Secondary outcomes included fasting plasma glucose (FDG-PET) and body weight, analyzed as changes from baseline to the longest follow-up, in order to explore potential systemic metabolic effects of GLP-1 RAs therapy in AD populations. These outcomes were reported descriptively across included trials when sufficient data were available.

Studies were excluded if they were non-randomized, open-label, uncontrolled, or enrolled participants without AD or MCI. Alternative diagnoses, including type 2 diabetes without AD, post-stroke cognitive impairment, vascular dementia, Parkinson's disease-related dementia, or other non-Alzheimer neurodegenerative conditions, were identified based on explicit diagnostic criteria reported in the original articles, such as clinical diagnosis, neuroimaging findings, or study inclusion definitions.

When cognitive impairment was attributed primarily to metabolic disease, cerebrovascular events, or mixed or uncertain etiologies without clear AD confirmation, studies were conservatively excluded to preserve diagnostic specificity and reduce indirectness.

A comprehensive search was conducted in PubMed/MEDLINE, Embase, and the Cochrane Central Register of Controlled Trials from database inception to October 2025. The search strategy combined MeSH terms and free-text expressions related to AD, MCI, and GLP-1 RAs (including liraglutide, exenatide, dulaglutide, semaglutide, and tirzepatide). No language or publication restrictions were applied. Reference lists of included studies and recent reviews were manually screened to identify additional eligible reports.

In addition to electronic database searches, the reference lists of all included articles and relevant recent reviews were manually screened to identify potentially eligible studies. Manual screening followed the same predefined inclusion and exclusion criteria applied during database searching. The search covered studies published from database inception through October 2025, with no restrictions on language or publication status. Titles and abstracts identified through citation tracking were independently assessed by two reviewers, and full texts were retrieved when eligibility could not be determined at the abstract level.

Two reviewers independently screened titles and abstracts retrieved full-text manuscripts, and assessed eligibility using predefined inclusion criteria. Any disagreements were resolved through discussion or consultation with a third investigator. For each included study, detailed data were extracted on sample size, age, sex distribution, body mass index, diagnostic criteria, intervention regimen, comparator, follow-up duration, and cognitive outcomes. Cognitive outcomes measured on different scales were standardized using the standardized mean difference (SMD). When change-from-baseline data were expressed as standard errors or confidence intervals, standard deviations were derived using standard formulas [16]. When multiple cognitive instruments were reported within a trial, the measure corresponding to the primary cognitive endpoint was selected.

Two reviewers independently evaluated each trial using the Cochrane Risk of Bias 2 (RoB 2) tool [17]. The domains assessed included randomization process, deviations from intended interventions, missing data, measurement of the outcome, and selective reporting. The overall judgment was categorized as low risk, some concerns, or high risk of bias. Graphical visualization was performed using the robvis package in R (version 4.5.2).

The certainty of the evidence was evaluated using the Grading of Recommendations, Assessment, Development and Evaluation (GRADE) approach [18]. Outcomes were rated as high, moderate, low, or very low based on risk of bias, inconsistency, indirectness, imprecision, and publication bias.

This meta-analysis adhered to the Cochrane Handbook for Systematic Reviews of Interventions [19]. Figures, tables, and Supplementary Materials were prepared according to Pharmaceutics author guidelines. The GRADE table is provided in the Supplementary Materials.

We pooled cognitive outcomes using standardized mean differences (Hedges g) based on change from baseline. For each arm, the change-score was computed as follow-up minus baseline. None of the included trials reported the SD of the change directly. Therefore, SD change was imputed from reported SDs (or from SEs when required, converted using SD = SE × √n) using the canonical variance propagation formula: S D c h a n g e = S D b a s e l i n e 2 S D f i n a l 2 + − × × × 2 r S D b a s l e i n e S D f i n a l

Because several trials reported more than one post-baseline cognitive assessment, the longest available follow-up was selected to avoid pseudo-replication and to maintain one effect estimate per study, in accordance with Cochrane recommendations. To address concerns related to non-linear cognitive trajectories, attrition, or potential floor effects at later timepoints, sensitivity analyses using intermediate follow-up assessments were prespecified when available.

We prespecified r = 0.50 as the primary working correlation and undertook sensitivity analyses with r = 0.30 and r = 0.70. Random-effects models using the restricted maximum likelihood estimator with Hartung–Knapp adjustment were applied. Between-study variability was quantified using I². Negative SMD values indicate greater cognitive decline in the intervention group relative to the control. All analyses were performed in R (version 4.4.1) using the meta and metafor packages. Supplementary Table S2 reports SD change estimates under r = 0.30, 0.50, and 0.70. Given the small number of included trials, the quantitative synthesis was prespecified as exploratory and hypothesis-generating rather than confirmatory.

The initial search identified 353 records from PubMed/MEDLINE, Embase, Scopus, and CENTRAL. After removal of duplicates and screening of titles and abstracts, 18 articles were assessed in full text. Nine randomized trials met preliminary inclusion criteria, but six were excluded because participants did not have AD or MCI due to AD, instead enrolling individuals with type 2 diabetes or post-stroke cognitive impairment. Trials enrolling individuals with cognitive impairment primarily attributed to type 2 diabetes or other metabolic conditions were excluded to avoid misclassification of Alzheimer's pathology and indirect inference regarding disease-modifying effects. Three studies fulfilled all eligibility criteria and were included in the quantitative synthesis [11,12,13]. The PRISMA flow diagram summarizing the selection process is provided in Figure 1.

The three included randomized controlled trials encompassed a total of 278 participants (mean age 68 ± 7 years; 51% women). Sample sizes ranged from 30 to 206 participants, with treatment durations between 26 weeks and 18 months. Liraglutide was evaluated in one trial [12] and Exenatide in two [11,13]. All studies compared GLP-1 RAs therapy with placebo (Table 1). Diagnostic confirmation of AD or MCI was established using NIA–AA or NINCDS-ADRDA criteria. Cognitive outcomes were assessed with validated scales, primarily the ADAS-Cog and MMSE.

Although baseline cognitive performance showed some variability within trials, participants were generally clustered within early disease stages, and none of the included studies performed stratified analyses according to predefined cognitive severity bands. Dei Cas et al. [11] enrolled a relatively homogeneous population predominantly within the MCI spectrum, as reflected by baseline MMSE scores around 26 and mean CDR values below 0.5. In Mullins et al. [13], most participants were classified as CDR 0.5 (prodromal AD/MCI), with only a small proportion meeting criteria for CDR 1 (mild AD), and no stratified analyses by severity were performed. Finally, in Gejl et al. [12], cognitive outcomes were assessed using a composite neuropsychological battery (primarily the Wechsler Memory Scale), rather than global screening tools, and baseline scores indicated balanced groups with mild impairment (27.2 ± 17.0 vs. 27.1 ± 12.8, respectively, for the placebo and liraglutide groups), but without categorical severity stratification.

All included studies were randomized and double-blind. Two trials [12,13] were judged as having some concerns due to incomplete outcome reporting, while Dei Cas et al. [11] presented some concerns due to its open-label pilot design, which introduced potential performance and measurement bias. Randomization and allocation concealment were adequate in all studies. The summary of the analysis is presented in Figure 2. The complete description and the reasons for the reduced scores are presented in the Supplementary Material (Table S1).

Meta-analysis of the three RCTs demonstrated no significant difference in global cognitive change between GLP-1 RAs and placebo. The pooled SMD was −0.21 (95% CI −0.81 to 0.38; I² = 47%; τ² = 3.77) (Figure 3). These results should be interpreted as exploratory, given the limited number of available trials. The direction of effect slightly favored placebo in short-term studies and liraglutide in the 12-month trial, but these differences were not statistically significant. Sensitivity analyses using intermediate follow-up timepoints produced effect estimates that differed by less than 0.03 SMD units from the primary analysis, indicating that the choice of the longest available follow-up did not materially affect the pooled results and that no evidence of floor effects was observed. Sensitivity analyses excluding shorter trials (<12 months) or those with a higher risk of bias yielded nearly identical results (SMD = −0.05; 95% CI −0.27 to 0.17; I² = 0%). Variation of the imputed correlation coefficient (r = 0.30–0.70) minimally influenced the pooled estimate (<5% difference). Leave-one-out analyses confirmed the robustness of the findings. Funnel plot inspection and Egger's regression test (p = 0.48) revealed no evidence of small-study bias.

Forest plot showing SMD in change from baseline for global cognitive scores comparing GLP-1 RAs versus placebo. Random-effects model using restricted maximum likelihood (REML) estimator with Hartung–Knapp adjustment. Negative values indicate greater decline with GLP-1 RAs relative to placebo. Random-effects meta-analysis using REML estimator with Hartung–Knapp adjustment; this figure represents the primary quantitative synthesis.

Exploratory analyses of metabolic parameters demonstrated no significant differences between GLP-1 RAs and placebo. The pooled SMD for fasting plasma glucose was −2.04 (95% CI −5.87 to 1.79), and the mean difference for body weight was −2.27 kg (95% CI −7.37 to 2.83) (Figure 4 and Figure 5). Only two of the three included trials reported body weight outcomes and were therefore included in the quantitative synthesis shown in Figure 5.

Both analyses demonstrated wide confidence intervals and were graded as very low certainty evidence according to the GRADE framework (Supplementary Table S2). Heterogeneity was low to moderate (I² ≈ 40%), and no systematic pattern favored either intervention or control. HbA1c data were not uniformly reported across the included studies, and BMI information was unavailable for two trials; therefore, these metabolic outcomes were excluded from the quantitative synthesis.

These neutral metabolic findings indicate that the absence of cognitive benefit was not confounded by differential changes in systemic metabolism. Collectively, the evidence suggests that peripheral glycemic modulation alone is unlikely to translate into measurable cognitive improvement in established AD.

Adherence was high across all trials, with completion rates above 85%. Summary of study-level data on adherence rates and adverse events among participants receiving GLP-1 RAs versus placebo (Table 2). Adverse events were mild to moderate and mainly gastrointestinal, consistent with the known pharmacologic profile of GLP-1 receptor agonists. No serious treatment-related adverse events were reported in any trial.

According to the GRADE assessment (), the certainty of evidence for the primary outcome (global cognitive function) was rated as low, downgraded for imprecision and small sample sizes. For secondary outcomes (fasting plasma glucose and body weight), the certainty was very low due to wide confidence intervals and limited sample sizes. No evidence of serious indirectness or publication bias was detected. Supplementary Table S3

This meta-analysis is limited by the small number of available randomized controlled trials and modest sample sizes, resulting in limited statistical power and imprecision of effect estimates. Restricting inclusion to biomarker- or clinically confirmed AD/MCI populations resulted in a limited number of eligible trials; however, broader inclusion would have compromised clinical specificity and increased indirectness, which we considered methodologically unacceptable for addressing treatment effects in established AD. The included studies also differed in disease stage, intervention duration, and cognitive instruments, which may have reduced sensitivity to detect subtle cognitive changes.

The available randomized clinical evidence in AD and MCI is currently restricted to two GLP-1 RAs, liraglutide and exenatide. Consequently, the present findings should be interpreted as drug-specific rather than representative of class-wide effects and cannot be extrapolated to newer GLP-1 RAs such as semaglutide or dulaglutide, for which no published randomized cognitive outcomes in AD/MCI are yet available. The number of eligible trials and total sample size were modest, resulting in limited statistical power and imprecision of effect estimates. Although heterogeneity was low, confidence intervals remained wide, and small treatment effects cannot be definitively excluded. Additionally, the included studies differed in disease stage (prodromal AD versus established AD), intervention duration, and primary study objectives. In some trials, cognitive outcomes were secondary or exploratory endpoints, which may limit sensitivity to detect subtle cognitive changes. Biomarkers such as amyloid burden, tau pathology, neuroinflammation, or cerebral perfusion were not systematically assessed. Although exploratory analyses of metabolic outcomes (fasting glucose and body weight) were performed, these were limited by small sample sizes and should be interpreted with caution. Despite these constraints, the low heterogeneity and consistent direction of effect across studies strengthen the internal validity of the findings.

Future research should prioritize biomarker-driven, long-duration RCTs in early or preclinical AD, ideally integrating multimodal neuroimaging and molecular endpoints to confirm central target engagement. Combination approaches addressing both metabolic and neuroinflammatory mechanisms could enhance therapeutic potential. Despite the neutral cognitive findings, GLP-1 RAs remain biologically attractive candidates for neuroprotection. The results of the EVOKE and EVOKE+ phase 3 trials will be critical to determine whether sustained receptor activation can truly modify the course of AD [14].

In this context, the design of the EVOKE and EVOKE+ phase 3 trials directly addresses several limitations observed in the current body of randomized evidence. These trials enrolled large, well-powered cohorts of individuals with early-stage, biomarker-confirmed AD, employed long treatment durations of up to 104 weeks, and used clinically meaningful progression endpoints such as the Clinical Dementia Rating—Sum of Boxes. In contrast to earlier trials, EVOKE and EVOKE+ were specifically structured to evaluate sustained disease-modifying effects rather than short-term cognitive changes, thereby providing a more robust framework to detect potential long-term benefits of GLP-1 RAs therapy.