Association between glucagon‐like peptide‐1 receptor agonists and risk of dementia in older adults with type 2 diabetes: A target trial emulation

Type 2 diabetes (T2D) is associated with an increased risk of cognitive decline and dementia, likely through mechanisms involving vascular dysfunction, insulin resistance, and chronic inflammation, posing a significant public health challenge as the population ages.1, 2, 3 Epidemiologic studies indicate that individuals with T2D have a 1.5‐ to 2‐fold increased risk of developing dementia compared with those without diabetes.4, 5, 6 Concurrently, dementia has become a leading cause of disability and dependency among older adults worldwide, affecting an estimated 57 million people in 2021, a figure projected to increase by nearly 10 million annually.7 The global economic burden of dementia reached approximately US$1.3 trillion in 2019 and is projected to exceed $2 trillion by 2050, posing a significant challenge to healthcare systems worldwide.8 As the prevalence of both T2D and dementia continues to rise, identifying glucose‐lowering therapies that may also mitigate dementia risk has become a critical clinical and public health priority.

Glucagon‐like peptide‐1 receptor agonists (GLP‐1 RAs), initially approved for T2D treatment in 2005 and obesity management in 2014, have emerged as promising candidates for repurposing as disease‐modifying therapies in dementia.9, 10, 11 In addition to their glucose‐lowering properties, GLP‐1 RAs exhibit multiple neuroprotective effects, including attenuation of neuroinflammation, enhancement of synaptic plasticity, reduction in amyloid‐beta accumulation, and modulation of tau phosphorylation, key pathological features of dementia and Alzheimer's disease.12, 13, 14 Early‐phase clinical trials (e.g., ELAD, EVOKE, EVOKE Plus) and real‐world studies have suggested potential cognitive benefits of GLP‐1 RAs.15, 16, 17, 18 Furthermore, post hoc analyses from large cardiovascular outcome trials such as REWIND, LEADER, SUSTAIN 6, and PIONEER 6 have shown signals of cognitive benefit among GLP‐1 RA users.16, 17, 19 Despite these encouraging findings, real‐world comparative evidence on the cognitive effects of GLP‐1 RAs versus other frequently prescribed antidiabetic agents remains limited and inconclusive.

Sodium‐glucose cotransporter‐2 inhibitors (SGLT2is) and dipeptidyl peptidase‐4 inhibitors (DPP4is), two other widely used second‐line glucose‐lowering agents, have also been explored for their potential neuroprotective effects.20, 21 SGLT2is may confer cognitive benefits through mechanisms such as reduced neuroinflammation and oxidative stress,20 while DPP4is may enhance neuroendocrine signaling by inhibiting enzymatic degradation of incretin hormones.21 However, results from large‐scale clinical trials, including CARMELINA, have shown no significant reduction in cognitive outcomes with DPP4i use.22, 23 Although biologically plausible, the comparative effects of GLP‐1 RAs versus SGLT2is or DPP4is on dementia risk have not been definitively established in clinical practice, and findings from prior observational studies remain inconsistent.22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33

To address these gaps, we conducted a retrospective cohort study emulating a target trial to assess the association between initiation of GLP‐1 RAs and incident dementia among older adults with T2D. Using electronic health record (EHR) data from the Penn Medicine health system, we compared GLP‐1 RA use to DPP4is and SGLT2is, applying rigorous methods to emulate the design and analysis of a hypothetical RCT.34 We applied propensity score (PS) matching and inverse probability treatment weighting (IPTW) to control for confounding, conducted competing risk analyses, externally validated findings using the TriNetX platform,35 and performed meta‐analysis to enhance generalizability. This study aimed to generate real‐world comparative evidence on the cognitive effects of GLP‐1 RAs and to identify patient subgroups most likely to benefit from their potential neuroprotective properties. By directly comparing the dementia risk associated with GLP‐1 RAs versus DPP4is and SGLT2is in routine clinical practice, our findings seek to inform more personalized diabetes management strategies for older adults at risk of cognitive decline.

In this target trial emulation study of older adults with T2D, GLP‐1 RA use was associated with lower dementia risk versus DPP4i, but higher risk versus SGLT2i. These associations remained directionally consistent across multiple sensitivity analyses and were further supported by external validation in TriNetX, strengthening confidence in the robustness of the findings. Although no statistically significant effect modification was observed across subgroups, the observed patterns suggest clinically meaningful hypotheses that warrant investigation in future studies.

The observed protective association of GLP‐1 RAs versus DPP4is aligns with the known neurobiological actions of GLP‐1 signalling, including reduced neuroinflammation, enhanced synaptic plasticity, and improved central insulin sensitivity, key processes implicated in neurodegenerative disease pathophysiology.13 Preclinical studies also suggest that GLP‐1 RAs may cross the blood–brain barrier and directly activate GLP‐1 receptors.45 Although DPP4is prolong endogenous GLP‐1 activity by inhibiting the DPP4 enzyme,46 this indirect mechanism is less potent than direct receptor agonism.47 Our findings are consistent with several observational studies,24, 25 but differ from a recent Medicare claims‐based analysis32 that reported no overall difference in dementia risk between GLP‐1 RA and DPP4i users. These differences likely reflect variations in data sources, covariate capture (including adiposity measures), population characteristics, and secular prescribing patterns rather than true contradiction, and the studies should be viewed as complementary. Taken together, our findings suggest that GLP‐1 RAs may confer cognitive benefits relative to DPP4is; however, whether this reflects a true neuroprotective effect of GLP‐1 RAs or comparatively weaker cognitive effects of DPP4is cannot be fully distinguished in an observational design and warrants investigation in future mechanistic or randomized studies.

The finding that SGLT2i use was associated with a lower dementia risk than GLP‐1 RAs adds to emerging evidence suggesting potential cognitive benefits of SGLT2is. Although SGLT2is are primarily known for their cardiovascular benefits, experimental data suggest possible cognitive mechanisms, including reduced neuroinflammation and improved cerebral perfusion.20, 48 Given the mixed findings in prior observational and trial‐based studies.25, 26, 27, 28, 29, 30, 33 we must interpret these results cautiously. Some observational studies reported lower dementia risk with SGLT2i versus GLP‐1 RA,27, 28 while others found no significant difference26, 30; together, the evidence remains inconclusive and highlights the need for head‐to‐head randomized trials with cognitive outcomes; together, the evidence remains inconclusive and highlights the need for head‐to‐head trials with cognitive endpoints. Although the Penn Medicine estimate suggested a modestly higher dementia risk for GLP‐1 RAs versus SGLT2is and the TriNetX estimate was closer to null, the direction of association was similar across both datasets, with differences in magnitude likely reflecting underlying variations in population characteristics, prescribing patterns, and data structures rather than a fundamentally discordant effect.

Sensitivity analyses further supported the robustness of the findings. Across multiple approaches, including IPTW in the unmatched cohort, Fine‐Grey competing risk models, application of a clinically informed dementia onset definition, sequential exclusion of comorbid conditions, use of vascular and mixed/other dementia as alternative outcomes, incorporation of a stricter dementia definition requiring both a diagnosis code and a dementia‐specific medication, adjustment for anticholinergic burden, and a fixed 2‐year follow‐up, the direction of associations remained consistent with the primary analysis, although estimates were imprecise when case counts were low. Negative control outcome analyses did not indicate residual bias, and E‐values suggested that a substantial unmeasured confounder would be required to fully explain the observed treatment‐dementia associations. External validation in TriNetX, together with pooled random‐effects estimates, further reinforced the overall stability of the findings.

Specifically, excluding patients with baseline MCI or other related neurologic conditions did not materially alter the observed associations. This is particularly important given that medication effects may vary by stages of cognitive impairment.48, 49 Similar patterns were observed when excluding patients with Parkinson's disease, substance use disorder, or vitamin B12 deficiency, indicating that secondary or reversible causes of cognitive impairment did not drive the observed associations. Parkinson's disease was included as a baseline covariate, and although biologically relevant for GLP‐1 RAs, incident Parkinson's cases were too few to allow meaningful secondary analysis. Future work with larger datasets and longer follow‐up periods will be necessary to fully evaluate this outcome. Taken together, these multiple sensitivity analyses reinforce the stability of the primary findings despite potential limitations in outcome misclassification, confounding, and follow‐up variation.

All subgroup analyses were exploratory, not adjusted for multiple comparisons, and did not demonstrate statistically significant effect modification; therefore, numerical differences should be interpreted with caution. Although some variation was observed across racial/ethnic groups, cardiometabolic risk strata, and GLP‐1 RA subtypes, these patterns were not statistically significant and may reflect residual confounding, differential diagnosis, or coding practices, treatment selection, or limited statistical power rather than true biological heterogeneity. For example, the protective association of GLP‐1 RAs versus DPP4is appeared somewhat more pronounced among non‐Hispanic White individuals, while higher dementia risk for GLP‐1 RAs versus SGLT2is appeared in both non‐Hispanic White and non‐Hispanic Black individuals, but none of these differences were significant and should be considered hypothesis‐generating only. Collectively, these exploratory findings highlight potential areas for future research in larger and more diverse cohorts with longer follow‐up to clarify whether such numerical differences reflect biological variation or social and structural determinants of health. Subgroup analyses were not repeated in TriNetX, as this external dataset was used solely to validate overall effect estimates, and its multisite structure may limit the reliability of subgroup‐level comparisons.

Our study has several strengths. First, we employed a target trial emulation framework, which explicitly aligned eligibility criteria, treatment strategies, follow‐up, and analytic approaches with a hypothetical randomized clinical trial, thereby enhancing causal inference from observational data. Second, the use of rich electronic health record data enabled the comprehensive capture of clinical covariates, including laboratory measurements such as haemoglobin A1c, allowing for more detailed adjustment than is feasible in many claims‐based studies.24, 26 Rigorous PS matching produced well‐balanced baseline characteristics between treatment groups, and the inclusion of two active comparators (DPP4is and SGLT2is) minimized confounding by indication and provided a nuanced assessment of the relative cognitive effects of GLP‐1 RAs in a real‐world setting. Third, the robustness of the findings was supported by an extensive set of sensitivity analyses, including IPTW using the full cohort, competing risk models, alternative dementia definitions, exclusion of individuals with conditions that may influence cognitive trajectories, adjustment for strong anticholinergic medications, fixed 2‐year follow‐up, and examination of dementia subtypes. External validation using TriNetX further strengthened generalizability and reproducibility.24, 26 Finally, to address the possibility of reverse causation, we incorporated a clinically informed onset window and accounted for mild cognitive impairment through both adjustment and exclusion, with results remaining consistent across analyses, although some residual bias cannot be completely excluded.

Nonetheless, several limitations should be considered. Despite extensive adjustment, residual confounding from unmeasured variables, such as diabetes duration, frailty, genetic predisposition, environmental, and socioeconomic factors, may remain and could influence the observed associations.3, 50, 51, 52 Residual confounding by adiposity is also possible because BMI does not capture body composition or fat distribution, and adiposity changes relate differently to dementia risk across obesity strata.53 Because EHRs record prescription orders rather than dispensing, refills, or adherence, true medication duration could not be ascertained; exposure was therefore defined at first prescription, and discontinuation or treatment switching cannot be reliably captured, precluding per‐protocol or as‐treated analyses and leaving open the possibility of residual bias related to treatment changes.

EHR systems also lack formal enrolment structures, and although required ≥1 encounter in the prior year to improve covariate completeness, continuous observation cannot be guaranteed. The median follow‐up of 2.4–3.0 years is relatively short for dementia onset and may lead to under‐ascertainment; however, limited event counts made it infeasible to implement a 1‐year lagged follow‐up period that excludes early post‐initiation diagnoses. Dementia diagnoses in EHRs are prone to under‐detection and misclassification,54 and Alzheimer's disease is often coded as unspecified dementia, likely explaining the lower proportion of AD diagnoses in our cohort than epidemiological estimates.

Generalizability may be limited to academic and U.S.‐based healthcare settings. Differences in effect magnitude across Penn Medicine and TriNetX highlight the need for replication in diverse populations. Finally, PS matching improved internal validity but reduced the analytic sample; however, IPTW analyses retaining the full cohort yielded directionally consistent results, supporting the robustness of the findings.

In this large real‐world cohort of older adults with T2D, initiation of GLP‐1 RA use was associated with lower dementia risk versus DPP4i but higher risk versus SGLT2i. Results were consistent across extensive sensitivity analyses and showed consistent directional patterns across subgroups. Together, the findings suggest potential cognitive benefits of GLP‐1 RAs relative to DPP4is while underscoring the emerging neuroprotective profile of SGLT2is. Prospective, biomarker‐informed studies are needed to clarify underlying mechanisms and inform treatment decisions aimed at reducing dementia risk in diabetes populations.

Authorship was determined using ICMJE recommendations. Ting Zhou and Yong Chen (University of Pennsylvania) contributed to the concept and design; Yong Chen (University of Pennsylvania) provided datasets for data analysis; Ting Zhou and Huilin Tang performed the statistical analysis; Ting Zhou visualized the analysis results; Ting Zhou, Huilin Tang, Bingyu Zhang, Dazheng Zhang, Yiwen Lu, Lu Li, Jiajie Chen, Yong Chen (Pfizer, New York City, NY, USA), David A. Asch, and Yong Chen (University of Pennsylvania) interpreted the results. Ting Zhou wrote the initial draft of the manuscript, and Yong Chen (University of Pennsylvania) supervised the study. All authors revised and approved the final version of the manuscript.

This study was supported in part by the U.S. National Institutes of Health (U01TR003709, U24MH136069, RF1AG077820, R01AG073435, R56AG074604, R01LM013519, R01DK128237, R21AI167418, R21EY034179).

Dr. David Asch is a partner and part‐owner of VAL Health and serves on the advisory boards of Thrive Global and Morpheus. Dr. Yong Chen (University of Pennsylvania) reported receiving personal fees from Merck & Co., Inc. and Pfizer Inc. outside the submitted work. Dr. Yong Chen (Pfizer, New York City, NY, USA) is a Pfizer employee and owns Pfizer stock. No potential conflicts of interest relevant to this article were reported by Ting Zhou, Huilin Tang, Bingyu Zhang, Dazheng Zhang, Yiwen Lu, Lu Li, and Jiajie Chen.