Introduction
Type 2 diabetes mellitus (T2DM) is a chronic disease with a persistently rising incidence worldwide, often accompanied by various serious complications. According to the “Global Diabetes Atlas (11th edition)” released by the International Diabetes Federation in April 2025, the number of people with diabetes aged 20 to 79 worldwide has reached 589 million, and it is projected to increase to 853 million by 2050.1 Dementia and depression are among the most common diseases in the elderly, significantly impacting the quality of life, mortality rates, and healthcare and social costs for patients and their families. Dementia and depression have a complex bidirectional relationship: Depression is both a risk factor and a precursor symptom of Alzheimer’s disease, as well as a common comorbidity at various stages of dementia.2 In recent years, the relationship between dementia, depression, and T2DM has garnered increasing attention.
Both dementia and depression are considered significant clinical complications of T2DM.3 Individuals with T2DM have a 1.5 times higher risk of developing dementia compared to those without T2DM.4 The risk of developing severe depression is even higher, reaching 1.5 to 2.0 times.5
Novel antidiabetic medications, glucagon-like peptide-1 receptor agonists (GLP-1RA) and sodium-glucose cotransporter-2 inhibitors (SGLT2i), have become a breakthrough in the treatment of T2DM due to their weight loss and cardiovascular and renal protective effects.6 Beyond glycemic control, both classes have demonstrated pleiotropic actions on the central nervous system (CNS) through several overlapping mechanisms. First, GLP-1 receptors are widely expressed in brain regions regulating appetite, reward, and mood, including the hypothalamus, amygdala, and nucleus accumbens.7,8 GLP-1RA can cross the blood-brain barrier to exert neurotrophic, anti-inflammatory, and neuroprotective effects7,9,10 while SGLT2i contribute to metabolic regulation, anti-inflammation, mitochondrial protection, autophagy modulation, and enhanced neurotrophic signaling.11 Second, both drug classes may ameliorate brain insulin resistance via enhanced central insulin signaling and glucose metabolism,12 suppress neuroinflammation through reduction of pro-inflammatory cytokines and microglial activation,13 and modulate neuronal mitochondrial bioenergetics—SGLT2i reprogram mitochondrial responses to nutrient deprivation,14 whereas GLP-1RA reduce reactive oxygen species production and improve mitochondrial function, synaptic plasticity, and cognition.10 Observational studies also indicate cognitive benefits, and several meta-analyses show that GLP-1RA and SGLT2i are superior to traditional antidiabetic medications in reducing the risk of dementia.15–19 However, clinical studies evaluating GLP-1RA and SGLT-2i in neurodegenerative and at-risk metabolic populations have yielded heterogeneous results.9 Previous meta-analyses have shown that there is still controversy regarding the effectiveness comparison of these two drugs in preventing dementia.17,18,20–23 Regarding depressive events, recent studies have shown that both medications have beneficial effects on depression.24,25 However, the potential association between GLP-1RA and suicidal ideation has gradually garnered attention.26 Although numerous individual studies have evaluated the effects of SGLT2i and GLP-1RA on dementia and depression-related outcomes in patients with T2DM, no quantitative synthesis has yet been conducted.
This study employs a meta-analysis method to systematically compare the effects of SGLT-2i and GLP-1RA on dementia and depression-related outcomes in patients with T2DM, aiming to provide key evidence for the differences in the neuropsychiatric protective effects of the two drug classes and to guide clinical personalized treatment decisions.
Methods
Search Strategy and Study Selection
A systematic search of PubMed, Cochrane Library, Embase, and Web of Science was conducted from inception to June 1, 2026. The search strategy combined Medical Subject Headings and free-text terms for three conceptual domains: (1) Interventions: “Ertugliflozin” OR “Dapagliflozin” OR “Canagliflozin” OR “Empagliflozin” OR “Ipragliflozin” OR “Luseogliflozin” OR “Tofogliflozin” OR “Sotagliflozin” OR “Sodium-Glucose Transporter 2 Inhibitors”; “Liraglutide” OR “Albiglutide” OR “Dulaglutide” OR “Exenatide” OR “Efpeglenatide” OR “Lixisenatide” OR “Semaglutide” OR “Taspoglutide” OR “Glucagon-Like Peptide-1 Receptor Agonists”. (2) Outcomes: “Cognitive Dysfunction” OR “Dementia” OR “Depression” OR “Anxiety Disorders” OR “Suicide” OR “Self Injury” (3) Population: “Type 2 Diabetes Mellitus”.
Inclusion criteria are: (1) Study design: Observational studies (cohort, case-control); (2) Participants: Adults with T2DM exposed to both SGLT-2i or GLP-1RA; (3) Outcomes: Quantifiable data on dementia or depression-related events; (4) Publication: Original research articles in English. Reviews, conference abstracts, animal studies, case reports, letters, commentaries, studies without event counts or comparator groups and non-English publications were excluded. Two reviewers independently screened titles or abstracts of all identified records and assessed full texts of potentially eligible articles.
Any disagreements were resolved through consensus or, when necessary, by consultation with the corresponding author.
Data Extraction and Quality Assessment
To systematically assess the comparability of the included studies, we extracted detailed information on study design, age thresholds, demographic characteristics, and key eligibility criteria—with particular attention to new-user design and exclusion of baseline outcomes to ensure incident outcome assessment. We also collected data on outcome definitions for dementia (including Alzheimer’s disease and vascular dementia) and depression-related outcomes (including depression, anxiety, suicide, and self-harm), as well as the use of standardized cognitive or behavioral testing for outcome ascertainment and the incidence of each outcome. Additionally, we recorded methods for confounding adjustment, including propensity score techniques and the specific covariates adjusted for in each study.
Methodological quality of included observational studies was appraised using the Newcastle-Ottawa Scale (NOS), which evaluates three domains: (1) selection of cohorts (Maximum score 4), (2) comparability of groups (Maximum score 2), and (3) outcome assessment (Maximum score. Studies scoring ≥7 were deemed high quality.27 In addition, the overall certainty of the evidence and methodological rigor of this meta-analysis were assessed using the AMSTAR 2 (A Measurement Tool to Assess Systematic Reviews, version 2) checklist, which comprises 16 items.28 The NOS and AMSTAR 2 assessment was conducted independently by two reviewers (Yuzheng Li and Huiyun Li), with disagreements resolved by discussion.
Data Synthesis and Analysis
Two investigators (Runzhou Pan and Bo Wang) independently screened the titles and abstracts of all identified records and assessed the full texts of potentially eligible articles. Data extraction was performed independently by two investigators (Runzhou Pan and Yuzheng Li) using a standardized form. Disagreements at any stage were resolved through consensus or consultation with the corresponding author (Yongcai Zhao).
Descriptive characteristics of included studies and outcome definitions were tabulated. All data are presented as the number of events and total patients for each treatment group (GLP-1RA users and SGLT-2i users) separately. Pooled odds ratios (ORs) with 95% confidence intervals (CIs) were calculated for the SGLT-2i group and the GLP-1RA group.
All data were analyzed using a random effects model. A p-value of <0.05 was considered statistically significant. Study heterogeneity was quantified via I2 statistics (I2<25%: low; 25–75%: moderate; >75%: high) and visualized through forest plots.
Sensitivity analyses involved sequential exclusion of individual studies. Prespecified subgroup analyses examined effect modifiers including geographic region, study design, age, and follow-up duration. Publication bias was assessed using funnel plot symmetry complemented by Egger’s regression test, with P<0.05 indicating significant publication bias.
The above statistical analyses were performed using Review Manager version 5.4 and Stata version 14.0 software.
Results
Literature Search
A systematic literature search was conducted following PRISMA guidelines across four electronic databases: PubMed (n=54), Embase (n=437), Web of Science (n=105), and Cochrane Library (n=16). Initial searches yielded 612 records. After removing 17 duplicates, 595 unique records underwent title or abstract screening.
Exclusion criteria at this stage removed: Irrelevant studies (n=319), reviews (n=190), conference abstracts (n=26), basic research (n=12), case reports (n=14), letters (n=3). Of the 31 full-text articles assessed, 18 were excluded for the following reasons: baseline presence of outcomes (n = 2), non-T2DM population (n = 1), absence of GLP-1RA or SGLT-2i intervention (n = 5), unavailable data on outcome event counts (n = 2), and overlapping database (n = 8), and use of non-GLP-1RA incretin-based therapy as the comparator (n = 1) (Supplementary Table 1). 13 studies met inclusion criteria for quantitative synthesis (Figure 1).
No additional records were identified through manual searches. Figure 1PRISMA flow diagram of the study selection process.The PRISMA flow diagram outlines the study selection process for a meta-analysis. It begins with identification, where 612 records are found through database searching: PubMed (54), Embase (437), Web of Science (105) and Cochrane Library (16).
No additional records are identified from other sources. After removing 17 duplicates, 595 records remain. Screening involves assessing these records, resulting in 564 exclusions due to reviews (190), basic research (12), conference abstracts (26), case reports (14), letters (3) and irrelevant studies (319).
Eligibility assessment is conducted on 31 full-text articles, with 18 exclusions due to outcomes present at baseline (2), non-T2DM population (1), absence of GLP-1RA or SGLT-2i intervention (4), unreported outcome event counts (2), duplicate database (8) and use of non-GLP-1RA incretin-based therapy as the comparator (1). Finally, 13 studies are included in the synthesis for meta-analysis.PRISMA diagram: study selection for meta-analysis from identification to inclusion.Abbreviations: T2DM, type 2 diabetes mellitus; GLP-1RA, glucagon-like peptide-1 receptor agonists; SGLT-2i, sodium-glucose cotransporter-2 inhibitors.
Characteristics of Included Studies and Quality Assessment
A total of 13 observational studies were included in this meta-analysis, comprising 7 studies on dementia risk (total participants: n = 538,000) and 6 studies on depression-related outcomes (suicide, self-harm, depression and anxiety; total participants: n = 517,315).29–41 Details are as follows: (1) Dementia studies: seven studies investigated dementia risk,29–35 spanning five countries (USA, Korea, Israel, Denmark, and Germany), with four employing a retrospective cohort design and three using a case-control design. The sex distribution was generally balanced across cohorts, with the number of SGLT-2i or GLP-1RA users ranging from 622 to 443,766 cases. Three further analyzed dementia subtypes (Alzheimer’s disease and vascular dementia).
(2) Depression-related studies: six studies examined depression-related outcomes, including suicide, self-harm, depression, and anxiety,36–41 comprising 517,315 participants across five countries (USA, Sweden, Denmark, UK, and Spain). Five studies used a retrospective cohort design and one used a case-control design, with the number of SGLT-2i or GLP-1RA users ranging from 2603 to 298,553. The sex distribution was also generally balanced across these cohorts.
(3) Study characteristics and comparability: across all 13 studies, the core eligibility criteria—T2DM diagnosis, new-user design, and exclusion of baseline outcomes—were broadly consistent across studies, though age thresholds varied: six studies included adults aged ≥18 years; three restricted to older adults (≥60 or ≥65 years); three required ≥50 or ≥55 years; and one did not specify an age cut-off. Dementia, its subtypes, and depression-related outcomes (including suicidal ideation, self-harm, or suicide death) were primarily identified using ICD-9/ICD-10 diagnosis codes, with some studies additionally incorporating prescriptions for anti-dementia or antidepressant medications into the diagnostic criteria. In addition, one study employed the Sweet 16 cognitive screening tool to assist in dementia ascertainment,32 while another study relied on data from the FDA Adverse Event Reporting System (FAERS) for Alzheimer’s disease outcome identification.34 All included studies reported the incidence of dementia or depression-related outcomes in both the SGLT-2i and GLP-1RA groups, or reported the usage rates of these two drug classes in the context of the aforementioned outcomes.
All studies controlled for confounding factors, of which four dementia-related studies and five depression-related studies applied propensity score methods (matching or weighting), with covariates typically including demographic characteristics, comorbidities (eg, hypertension, cardiovascular disease, depression, obesity), diabetes-related complications, glycemic control indicators, and concomitant medications. Furthermore, the follow-up duration exceeded one year in all studies, and the majority had follow-up periods of more than five years (Table 1). Of the 13 included studies, 10 reported receiving financial support (Supplementary Table 2).Table 1Characteristics of the Included StudiesFirst Author, YearCountryStudy DesignPS; Adjusted VariablesPeriodFollow-up DurationKey Eligibility CriteriaMale/FemaleDrug NameNumber (n)GLP-1RA/SGLT-2iOutcome DefinitionOutcome IncidenceGLP-1RASGLT-2iHuilin Tang,292025USARCSIPTWDemographics, SES, healthcare utilization, diabetes complications, comorbidities (incl. depression, anxiety), medications, HbA1c, BMI.2014–2023Assignment to first of outcome, loss, or endInclusion: ≥50 y, T2DM, new user of GLP-1RA, SGLT2i, or other GLD.Exclusion: Baseline ADRD or antidementia drugs, T1DM/gestational diabetes, ESRD/dialysis.11,647/12,470EXE;ALB;DUA;LIR;LIX;SEM;TIRCANA; DAPA; EMPA; ERTU23,810/24,260(After IPTW)ADRD:ICD-9/ICD-10 codes;Cognitive testing:NoGLP-1RA vs other GLD:ADRD: 1.42% vs 2.15%;SGLT2i vs other GLD:ADRD: 1.39% vs 2.31%;SGLT2i vs GLP-1RA:ADRD: 0.84% vs 0.87%Mingyang Sun,302025USARCSPSM;Age, sex, race, comorbidities (hypertension, IHD, CVD, CKD), HbA1c, BMI, medications (insulin, SU, statins).2004–2024Assignment to first of outcome, loss, or endInclusion: ≥18 y, T2DM, first prescription of SGLT2i or GLP-1RA.Exclusion: Baseline dementia/cognitive impairment, prior use of study drugs, incomplete data.311,382/309,594NRNR221,883/221,883(After PSM)Dementia, VaD, AD:ICD-10-CM codes;Cognitive testing:NoSGLT2i vs GLP-1RA:Overall dementia: 2.7% vs 3.6%;VaD: 0.8% vs 1.1%;AD: 1.5% vs 1.9%Bin Hong,312024KoreaRCSPSM;Sociodemographics, comorbidities, medications, healthcare utilization, lab results (FBG, eGFR, BMI, BP).2010–2022Assignment to first of outcome, loss, or endInclusion: ≥60 y, T2DM, metformin only at baseline, new user of SGLT2i or dulaglutide.Exclusion: Baseline dementia, ESRD/dialysis, <2 y follow-up potential, prior study drug use.7870/5694DUADAPA;EMPA1038/2076(After PSM)Dementia, AD, VaD:ICD-10 codes;anti-dementia prescription;Cognitive testing:No.SGLT2i vs DUA:Overall dementia: 3.32% vs 4.14%;AD: 3.18% vs 4.05%;VaD: 0.19% vs 0.10%;Avi Cohen,322025IsraelRCSIPTW;Sociodemographics, diabetes duration, HbA1c, BMI, BP, lipids, eGFR, CVD, average daily dose.1998–20198.7 yearsInclusion: ≥65 y, T2DM, users of DPP4i/SGLT2i/GLP-1a.Exclusion: Diabetes diagnosed before 55 y, T1DM/gestational diabetes, baseline dementia, <1 y follow-up.42,301/42,497NRNR16,960/11,767(After IPTW)Dementia:ICD-9 codes;Sweet 16 ≤13;anti-dementia prescription;Cognitive testing:Sweet 16.SGLT2i vs GLP-1RA:Dementia: 2.34% vs 2.68%Ida Kim Wium-Andersen,332019DenmarkCCSNR;Age, sex, education, marital status, comorbidities/medications (IHD, CVD, hypertension), diabetes complications (acute/chronic).1995–20128.7 yearsInclusion: ≥35 y, T2DM (national registry).Exclusion: T2DM diagnosed before 35 y, gestational/secondary diabetes, baseline dementia/antidementia drugs.31,466/26,629NRNR4168/912Dementia:ICD-10 codes;anti-dementia prescriptionCognitive testing:No.Dementia vs non-Dementian=11619 vs n=46476SGLT2i: 0.38% vs 1.87%GLP-1RA: 2.12% vs 8.44%Hayato Akimoto,342020USACCSNR;Age, sex, common T2DM comorbidities (hypertension, dyslipidemia, CVD, using medication as proxy).2004–2018NRInclusion: ≥65 y, T2DM, antidiabetic monotherapy.Exclusion: Hypoglycemia, combination therapy, AD drugs started before antidiabetics.33,380/32,705ALB;DUA; EXE;LIR;LIXCANA; DAPA; EMPA7047/1574AD:FAERS reports of AD;anti-AD prescriptionCognitive testing:No.AD vs non-ADn=1250 vs n=64835SGLT2i: 0.40% vs 2.42%GLP-1RA: 2.40% vs 10.82%Jens Bohlken,352018GermanyCCSNR;Age, sex, index year, physician, mean HbA1c, diabetes duration, comorbidities (hypertension, hyperlipidemia, CHD, stroke, depression), medications.2013–2017≥1 yearsInclusion: ≥60 y, T2DM, GLD prescription, HbA1c record.Exclusion: Observation period <365 days.7250/9302NRNR307/315Dementia:ICD-10 codes;Cognitive testing:NoDementia vs non-Dementian=8276 vs n=8276SGLT2i: 1.7% vs 2.0%GLP-1RA: 1.7% vs 2.1%Huilin Tang,362024USARCSPSM;Demographics, comorbidities (depression, obesity, ADRD), medications, healthcare utilization, baseline GLD use.2017–2020Assignment to first of outcome, loss, or endInclusion: ≥66 y, T2DM, new user of GLP-1RA, SGLT2i, or DPP4i.Exclusion: T1DM, baseline suicidal ideation/behavior, ESRD, concurrent initiation of both drugs.21,722/22,090(After PSM)EXE;DUA;LIR;SEMNR21,807/21,807(After PSM)Suicidal ideation or behaviors:ICD-10/ICD-10-CM codes.Psychological assessment:No.SGLT2i vs GLP-1RA:Suicidal ideation or behaviors: 0.38% vs 0.41%GLP-1RA vs DPP4i:Suicidal ideation or behaviors: 0.46% vs 0.49%Peter Ueda,372024Sweden, DenmarkRCSPSW;Sociodemographics, calendar year, psychiatric history (recent & past), medical history, medications, healthcare utilization.2013–20215 yearsInclusion: 18–84 y, new user of GLP-1RA or SGLT2i.Exclusion: End-stage illness (incl. dementia, coma), dialysis/renal transplant, pancreatic disease, prior study drug use.182,216/116,337NRNR124,517/174,036(After PSW)Suicide/self-injury:ICD-10 codes;Depression/anxiety:ICD-10 codesAntidepressant prescriptions.Psychological assessment:No.GLP-1RA vs SGLT2i:Suicide/self-injury: 0.73% vs 0.50%;Depression/anxiety: 6.79% vs 5.26%;Samantha B Shapiro,382025UKRCSPSW;Age, sex, SES, BMI, HbA1c, diabetes duration, micro/macrovascular complications, mental health disorders (depression, anxiety, SH history), medications.2013–2020Assignment to first of outcome, loss, or endInclusion: ≥18 y, T2DM, first prescription of GLP-1RA, DPP4i, or SGLT2i.Exclusion: <1 y medical history, no T2DM diagnosis, ESRD, concurrent initiation of both drugs.65,224/63,324(After PSW)DUA;EXE;LIRCANA; DAPA;EMPA;ERTU32,336/96,212(After PSW)Suicide/self-injury:ICD-10 codes;Psychological assessment:No.GLP-1RA vs SGLT2i:Suicide/self-injury: 0.74% vs 0.47%;GLP-1RA vs DPP-4i:Suicide or self-injury: 0.83% vs 0.46%;Huilin Tang,392025USARCSPSM;Demographics, comorbidities (obesity, ADRD), medications, baseline GLD use.2014–20202 yearsInclusion: ≥66 y, T2DM, new user of GLP-1RA, SGLT2i, or DPP4i.Exclusion: T1DM, baseline depression/mood disorder, antidepressant use, ESRD.16,462/12,868(After PSM)EXE; DUA;LIR;SEMNR14,665/14,665(After PSM)Depression:ICD-10/ICD-10-CM codes.Psychological assessment:No.GLP-1RA vs SGLT2i:Depression: 6.55% vs 5.76%;GLP-1RA vs DPP-4i:Depression: 7.02% vs 7.84%;Isabel Hurtado,402024SpainRCSIPTW;Sociodemographics, lifestyle (tobacco/alcohol), comorbidities (depression, anxiety, sleep disorders), BMI.2015–2021Assignment to first of outcome, loss, or endInclusion: ≥18 y, T2DM with obesity, new user of GLP-1RA or SGLT2i, metformin at baseline.Exclusion: No VHS coverage, non-resident/temporary resident.7877/6790NRNR3040/11,627(After IPTW)Suicide/self-injury:ICD-9/ICD-10 codes.Psychological assessment:No.GLP-1RA vs SGLT2i:Suicide/self-injury: 0.33% vs 0.15%;Ida Kim Wium-Andersen,412022DenmarkCCSNR;Age, sex, education, marital status, enrollment year, comorbidities/medications, diabetes complications (acute/chronic), baseline other GLD use.2000–2012Assignment to first of outcome, loss, or endInclusion: >35 y, incident T2DM (national registry).Exclusion: Baseline schizophrenia, bipolar disorder, dementia, depression.74,806/65,190NRNR2239/364Depression:ICD-10 codes;Antidepressant prescriptions;Psychological assessment:No.Depression vs non-Depressionn=40940 vs n=32929GLP-1RA: 2.72% vs 3.42%SGLT2i: 0.33% vs 0.69%Abbreviations: AD, Alzheimer’s disease; ADRD, Alzheimer’s disease and related dementias; ALB, albiglutide; BMI, body mass index; CANA, canagliflozin; CCS, case-control study; CHD, coronary heart disease; CKD, chronic kidney disease; CVD, cerebrovascular disease; DAPA, dapagliflozin; DPP4i, dipeptidyl peptidase-4 inhibitors; DUA, dulaglutide; eGFR, estimated glomerular filtration rate; EMPA, empagliflozin; ERTU, ertugliflozin; ESRD, end-stage renal disease; EXE, exenatide; FAERS, FDA Adverse Event Reporting System; FBG, fasting blood glucose; GLD, glucose-lowering drug; GLP-1RA, glucagon-like peptide-1 receptor agonists; HbA1c, glycated hemoglobin; ICD, International Classification of Diseases; ICD-9, International Classification of Diseases, Ninth Revision; ICD-10, International Classification of Diseases, Tenth Revision; ICD-10-CM, International Classification of Diseases, Tenth Revision, Clinical Modification; IHD, ischemic heart disease; LIR, liraglutide; IPTW, inverse probability of treatment weighting; LIX, lixisenatide; NR, not reported; PS, propensity score; PSM, propensity score matching; PSW, propensity score weighting; RCS, retrospective cohort study; SEM, semaglutide; SES, socioeconomic status; SGLT2i, sodium-glucose cotransporter-2 inhibitors; SH, self-harm; SU, sulfonylureas; T1DM, type 1 diabetes mellitus; T2DM, type 2 diabetes mellitus; TIR, tirzepatide; UK, United Kingdom; USA, United States of America; VaD, vascular dementia; y, years.
According to the NOS assessment, the quality of the studies was evaluated and it was shown that all 13 included studies scored ≥7 (maximum 9), indicating high methodological quality (Supplementary Table 3). The methodological quality of this meta-analysis was assessed using the AMSTAR 2 tool. Among the 16 items, 7 were considered critical items (Items 2, 4, 7, 9, 11, 13, and 15).
Of these critical items, six were fully satisfied and one (Item 2, protocol registration) was partially satisfied. All non-critical items were fully satisfied (Supplementary Table 4).
Dementia Outcomes
A total of 7 observational studies involving 538,000 participants (SGLT-2i: n = 262,787; GLP-1RA: n = 275,213) were included in the meta-analysis of dementia risk. The sensitivity analysis confirmed robust directional stability of the primary finding (Supplementary Figure 1). The pooled analysis demonstrated a statistically significant reduction in dementia risk among SGLT-2i users compared to GLP-1RA users.
The overall OR was 0.85 (95% CI: 0.77–0.94; P = 0.001; I2=39%) under a random-effects model (Figure 2). Figure 2Forest plot of dementia risk: SGLT-2i vs GLP-1RA in T2DM.Forest plot with a study table at left and odds ratio plot at right. Columns list study or subgroup; SGLT 2i events and total; GLP 1RA events and total; weight; and odds ratio, M H, random, 95 percent confidence interval.
Seven studies are listed: Huilin Tang 1 2025, Mingyang Sun 2025, Bin Hong 2024, Avi Cohen 2025, Ida Kim Wium Andersen 2019, Hayato Akimoto 2020, Jens Bohlken 2018. The x axis is labeled odds ratio with ticks at 0.2, 0.5, 1, 2, 5. The y axis has study names and a total row.
A vertical reference line is at 1. Individual study odds ratios shown are 1.03, 0.79, 0.80, 0.87, 0.81, 0.75, 0.95 with their printed 95 percent confidence intervals. The pooled diamond is 0.85 with 95 percent confidence interval 0.77 to 0.94.
Bottom labels read favours SGLT 2i on the left and favours GLP 1RA on the right.A forest plot comparing dementia risk between SGLT 2i and GLP 1RA, with an overall odds ratio below 1.Abbreviations: CI, confidence interval; GLP-1RA, glucagon-like peptide-1 receptor agonists; M-H, Mantel-Haenszel; OR, odds ratio; SGLT-2i, sodium-glucose cotransporter-2 inhibitors; T2DM, type 2 diabetes mellitus.
The vascular dementia analysis included 2 studies with 446,880 participants (SGLT-2i: n = 223,959; GLP-1RA: n = 222,921). SGLT-2i users demonstrated a significantly lower risk compared to GLP-1RA users (OR 0.75, 95% CI: 0.71–0.80; P < 0.0001; I2=0) (Figure 3A). Three studies (n = 455,501) specifically evaluated Alzheimer’s disease risk.
SGLT-2i use was associated with a 24% lower risk compared to GLP-1RA (OR 0.76, 95% CI: 0.72–0.79; P < 0.0001) (Figure 3B). Figure 3Forest plots of (A) vascular dementia and (B) Alzheimer’s disease risk: SGLT-2i vs GLP-1RA in T2DM.Image A: Forest plot for vascular dementia comparing SGLT 2i vs GLP 1RA. X-axis: odds ratio (0.2, 0.5, 1, 2, 5).
Y-axis: Study/Subgroup. Mingyang Sun 2025: odds ratio 0.75 (95 percent CI: 0.71-0.80). Bin Hong 2024: odds ratio 2.00 (95 percent CI: 0.22-17.93).
Subtotal: odds ratio 0.75 (95 percent CI: 0.71-0.80). Image B: Forest plot for Alzheimers disease comparing SGLT 2i vs GLP 1RA. X-axis: odds ratio (0.2, 0.5, 1, 2, 5).
Y-axis: Study/Subgroup. Mingyang Sun 2025: odds ratio 0.76 (95 percent CI: 0.72-0.79). Bin Hong 2024: odds ratio 0.78 (95 percent CI: 0.52-1.15).
Hayato Akimoto 2020: odds ratio 0.75 (95 percent CI: 0.29-1.92). Subtotal: odds ratio 0.76 (95 percent CI: 0.72-0.79).Two forest plots comparing odds ratio for vascular dementia and Alzheimers disease across studies.Abbreviations: CI, confidence interval; GLP-1RA, glucagon-like peptide-1 receptor agonists; M-H, Mantel-Haenszel; OR, odds ratio; SGLT-2i, sodium-glucose cotransporter-2 inhibitors; T2DM, type 2 diabetes mellitus.
Depression-Related Outcomes
Six studies (n = 517,206) evaluated composite depression-related outcomes. The sensitivity analysis confirmed that the initial findings exhibited significant directional stability (Supplementary Figure 2). SGLT-2i demonstrated a 24% lower risk vs GLP-1RA (OR 0.76, 95% CI: 0.67–0.88; P = 0.0001).
The heterogeneity was high (I2=82%), and a random effects model was adopted (Figure 4). Figure 4Forest plot of composite depression-related outcomes: SGLT-2i vs GLP-1RA in T2DM.A meta-analysis figure lists six cohort studies comparing two glucose-lowering medication classes in type 2 diabetes, with event counts and sample sizes for each group. Each study is shown as a square with a horizontal 95 percent confidence interval on an odds-ratio axis and a vertical reference line at 1.0.
Most point estimates lie to the left of 1 and the pooled effect is displayed as a black diamond around 0.76 with a 95 percent CI of roughly 0.67 to 0.88. Heterogeneity statistics are reported (I superscript 2 about 82 percent) and the axis is labeled to indicate which side favors each treatment.Meta-analysis plot showing study odds ratios for depression outcomes with two diabetes drug classes.Abbreviations: CI, confidence interval; GLP-1RA, glucagon-like peptide-1 receptor agonists; M-H, Mantel-Haenszel; OR, odds ratio; SGLT-2i, sodium-glucose cotransporter-2 inhibitors; T2DM, type 2 diabetes mellitus.
Four studies specifically focused on suicidal or self-injury with a sample size of 485,273 individuals. Compared to GLP-1RA, SGLT-2i drugs had a lower risk of suicide or self-injury (OR: 0.69, 95% CI: 0.60–0.80; P < 0.0001; I2=51%) (Figure 5A). Three studies (n = 215,436) assessing depression or anxiety outcomes revealed a 22% risk reduction for SGLT-2i vs GLP-1RA (OR 0.78, 95% CI: 0.65–0.93; P = 0.006; I2=90%) (Figure 5B).
Figure 5Forest plots of (A) suicide or self-injury and (B) depression or anxiety: SGLT-2i vs GLP-1RA in T2DM.A two-part forest-plot figure labeled A and B summarizes pooled evidence from several cohort studies comparing two diabetes medication classes. Each panel lists studies in rows with event counts and totals for the two treatment groups, plus a study weight column. Blue squares with horizontal lines show individual study odds ratios and 95 percent confidence intervals and a black diamond at the bottom of each panel shows the combined estimate.
Panel A reports results for suicide or self-harm outcomes and shows a pooled odds ratio below 1, with moderate between-study variability. Panel B reports results for anxiety or depression and also shows a pooled odds ratio below 1, with substantial heterogeneity. The x-axis is logarithmic with a vertical reference line at 1; values left of 1 indicate lower odds for the left-hand treatment group and values right of 1 indicate lower odds for the right-hand group.Two-panel meta-analysis chart showing odds ratios for mental health outcomes in diabetes drug users.Abbreviations: CI, confidence interval; GLP-1RA, glucagon-like peptide-1 receptor agonists; M-H, Mantel-Haenszel; OR, odds ratio; SGLT-2i, sodium-glucose cotransporter-2 inhibitors; T2DM, type 2 diabetes mellitus.
Subgroup Analysis
For dementia-related outcomes, the pooled estimate for SGLT2i versus GLP-1RA was statistically significant in Asian populations (OR 0.86, 95% CI 0.75–0.99; I2 = 0%), whereas it was not statistically significant in European (OR 0.88, 95% CI 0.70–1.02; I2 = 0%) or U.S. populations (OR 0.87, 95% CI 0.69–1.10; I2 = 73%). When stratified by study design, the association was statistically significant in cohort studies (OR 0.86, 95% CI 0.75–0.98; I2 = 65%) but not in case-control studies (OR 0.87, 95% CI 0.70–1.09; I2 = 0%). Age-stratified analysis revealed that studies including younger populations (minimum age 18 years, without an upper age restriction) showed more pronounced differences (OR 0.79, 95% CI: 0.76–0.81; I2=0%), whereas the differences were relatively smaller in studies focusing on older populations (minimum age ≥50 or ≥60 years).
Stratification by follow-up duration indicated that the association was more pronounced in the subgroup with follow-up ≥5 years (OR 0.85, 95% CI: 0.76–0.95; I2=54%). For depression-related outcomes, the lower event rate associated with SGLT2i compared with GLP-1RA was more significant in European cohorts (OR 0.67, 95% CI: 0.54–0.84; I2=83%), but not evident in US cohorts (OR 0.93, 95% CI: 0.85–1.02; I2=0%). Furthermore, compared with older populations (minimum age ≥50 or ≥60 years) or those with follow-up <5 years, the subgroups of younger populations (minimum age 18 years, without an upper age restriction) and long-term users (follow-up ≥5 years) showed more pronounced differences (younger subgroup: OR 0.67, 95% CI: 0.54–0.84; I2=83%; long-term users subgroup: OR 0.71, 95% CI: 0.59–0.86; I2=79%).
These differences remained statistically significant across different study designs (cohorts: OR 0.80, 95% CI: 0.70–0.92; I2=81%; case-control: OR 0.60, 95% CI: 0.48–0.76; I2=0%) (Supplementary Table 5).
Publication Bias
The distribution in the two funnel plots was roughly symmetrical, indicating a relatively low risk of publication bias (Supplementary Figures 3 and 4). The Egger regression test results for the studies on 7 dementia outcomes and 6 depression-related outcomes also showed that the risk of publication bias was relatively low (dementia outcomes: p = 0.153, depression-related outcomes: p = 0.433).
Discussion
This meta-analysis shows that among patients with T2DM, SGLT-2i users demonstrated a 15% lower risk of dementia and 24% reduced incidence of depression-related outcomes compared to GLP-1RA users. In subtype analyses, SGLT-2i use was associated with lower proportions of vascular dementia, Alzheimer’s disease, suicide/self-harm events, and depression/anxiety diagnoses.
The effects of GLP-1RA and SGLT2i on the risks of dementia and depression have garnered increasing attention.15–17,19,42 In recent years, network meta-analyses have provided new perspectives for indirectly comparing the dementia prevention efficacy of different glucose-lowering drugs. Li et al demonstrated that SGLT2i and GLP-1RA may be optimal treatment options for reducing dementia risk in patients with T2DM, with similar SUCRA values and comparable efficacy.17 Tian et al integrated three randomized controlled trials (RCTs) and 24 observational studies in a network meta-analysis, finding that SGLT2i, GLP-1RA, thiazolidinediones, and DPP-4 inhibitors were all associated with a decreased risk of dementia, with SGLT2i ranking as the optimal intervention according to SUCRA.18 Regarding depression-related risks, multiple studies suggest that GLP-1RA possess neuropsychiatric protective potential: activation of central GLP-1 reduces anxiety- and depression-like behaviors in animal models;43 meta-analyses based on RCTs and observational studies consistently indicate that GLP-1RA improve depressive symptoms and lower depression scale scores;24,44 and population-based studies have shown that SGLT2i may be associated with mood improvements.25 In addition, both drug classes may exert positive effects on mood and cognitive function through indirect pathways, including improved renal function, enhanced cardiac ejection fraction, and reduced cardiovascular and stroke events.45–48 However, current meta-analyses have primarily focused on individual drug effects or indirect comparisons, lacking a direct head-to-head comparison between SGLT2i and GLP-1RA regarding dementia risk reduction and depression-related outcomes in patients with T2DM.
To our knowledge, this is the first meta-analysis to directly compare SGLT2i and GLP-1RA with respect to dementia and depression-related outcomes in patients with T2DM. By synthesizing head-to-head observational data, our study fills this evidence gap and demonstrates that SGLT2i users had lower observed rates of all-cause dementia, vascular dementia, Alzheimer’s disease, and depression-related outcomes (including anxiety, depression, and suicidal behaviors) compared with GLP-1RA users. These findings suggest that SGLT2i may offer more comprehensive neuropsychiatric protection than GLP-1RA for T2DM patients at elevated risk—such as those with a history of depression, cognitive decline, or cerebrovascular disease.
However, given the observational nature of the included studies, these findings should be considered hypothesis-generating and warrant confirmation through RCTs. Furthermore, contextualizing our findings alongside a recent network meta-analysis by Tseng et al (62 RCTs, n=200,068) provides valuable insights.49 Tseng et al reported that SGLT2i, particularly high-dose dapagliflozin, may mitigate delirium and depression, whereas GLP-1RA, notably dulaglutide and liraglutide, may enhance cognitive function.49 Of note, their observation that only high-dose dapagliflozin showed significant protective effects suggests that dosing may be a critical factor—an aspect not addressed in our meta-analysis due to limited data. Both studies consistently support agent-specific neuroprotective profiles; however, they diverge on dementia outcomes—Tseng et al found no treatment significantly reduced dementia incidence, whereas our meta-analysis observed lower dementia rates among SGLT2i users versus GLP-1RA users.
This discrepancy likely reflects methodological differences: Tseng et al included RCTs with limited follow-up (mean 129.2 weeks), capturing dementia events via adverse event reporting, whereas our meta-analysis included observational studies with longer follow-up (up to 8.7 years) and real-world ICD-coded diagnoses—better suited to detect late-onset outcomes, though susceptible to confounding. Collectively, these findings underscore the importance of agent-specific and dose-specific effects, highlight the need for head-to-head RCTs with dedicated neuropsychiatric endpoints, and warrant careful consideration when selecting antihyperglycemic therapies for individuals at elevated neuropsychiatric risk.
To systematically assess the comparability of the included cohorts, we compared the 13 studies in terms of methodological and outcome-related characteristics. Overall, the studies demonstrated good consistency in core design features: all studies required a confirmed diagnosis of T2DM in adults as the entry criterion, adopted a new-user design, and excluded baseline outcomes. Regarding age range, six studies included adults covering younger populations (≥18 or ≥35 years), six studies were restricted to older adults (≥60 or ≥65 years), and one study set the cutoff at ≥50 years.
In terms of outcome definitions, dementia and depression-related outcomes were primarily identified using ICD-9/ICD-10 diagnosis codes, with some studies additionally incorporating prescriptions for anti-dementia or antidepressant medications into the diagnostic criteria, and one study (Cohen et al, 2025) employing the Sweet 16 cognitive screening tool to assist in dementia ascertainment. Follow-up durations ranged from 1 to 8.7 years, with the majority exceeding 5 years. For confounding control, five of the seven dementia studies and all six depression studies applied propensity score methods (matching or weighting), consistently adjusting for demographic characteristics, comorbidities (eg, hypertension, cardiovascular disease, depression, obesity), diabetes-related complications, glycemic control indicators (HbA1c), and concomitant medications.
Moreover, multiple studies directly compared GLP-1RA and SGLT-2i in head-to-head analyses, providing direct evidence for the primary comparisons. Although methodological variations may partly explain the observed statistical heterogeneity—particularly for depression-related outcomes (I2 = 82%)—and the relatively wide confidence intervals in some pooled estimates, the consistency of findings across sensitivity and subgroup analyses supports the robustness of our conclusions. In addition, the relatively wide confidence intervals in the pooled estimates may also be attributable to the use of a random-effects model, which accounts for both within-study and between-study variance.
The potential mechanisms underlying this result may be as follows: (1) SGLT2i can enhance cerebral blood flow, improve brain ischemia, and reduce the incidence of cerebrovascular events, thereby decreasing cognitive decline.50,51 However, the cerebrovascular benefits of GLP-1RA have not been fully confirmed. (2) Mitochondrial dysfunction is one of the core pathological features of neurodegenerative diseases, including dementia. SGLT2i may intervene in this pathological process by optimizing mitochondrial energy metabolism efficiency in the brain, promoting neuronal survival, and improving cognitive function.19,52,53 However, the direct effects of GLP-1RA on mitochondrial function remain unclear.
(3) The accumulation of β-amyloid (Aβ) and the hyperphosphorylation of tau protein are hallmark features of Alzheimer’s disease. Dysfunction of AMP-activated protein kinase (AMPK) is associated with various neuropsychiatric disorders, including Alzheimer’s disease, Parkinson’s disease, stroke, and depression. SGLT2i can activate the AMPK signaling pathway, reducing Aβ production and inhibiting tau protein, thereby slowing the neurodegenerative process.51,54–57 In contrast, a randomized trial found that liraglutide did not significantly reduce Aβ accumulation after 26 weeks of treatment in patients with Alzheimer’s disease.58 (4) SGLT2i have been shown in mouse models to restore overactivated mTOR to a physiologically activated state, thereby mitigating cognitive decline and preventing the onset or progression of Alzheimer’s disease.59 (5) SGLT2i can exert an inhibitory effect on acetylcholinesterase by acting on the same targets as acetylcholinesterase inhibitors, improving memory dysfunction caused by scopolamine, and increasing levels of serotonin in the brain.
This may be a potential mechanism for alleviating depression and slowing the progression of Alzheimer’s disease.60 (6) Chronic neuroinflammation and oxidative stress constitute the core driving mechanisms of dementia pathology. SGLT2i demonstrate stronger central anti-inflammatory effects by effectively protecting neurons and delaying cognitive decline through significant reductions in pro-inflammatory factors and oxidative stress markers.19,53,61 While GLP-1RA have anti-inflammatory potential, their exact neuroprotective strength and mechanisms of action in the central nervous system require further investigation.62 (7) SGLT2i can induce the production of ketone bodies. Endogenous ketone bodies, as potential neuroprotective metabolites, may improve mood by enhancing brain bioenergetics, whereas GLP-1RA do not have this effect.25 (8) Although GLP-1RA have been shown to have potential antidepressant effects, caution is still warranted regarding their potential association with suicidal ideation: the appetite-suppressing phenomenon induced by these drugs, which acts on brain regions involved in appetite and reward regulation, such as the hypothalamus and nucleus accumbens, has been significantly associated with suicidal thoughts and self-harming behaviors in adolescent populations.26,63,64 Moreover, while the neurotrophic and anti-inflammatory properties of GLP-1RA may help prevent the degeneration of dopaminergic neurons, which can contribute to mood regulation, there is also a risk of increased impulsive behaviors, including self-harm.65 (9) Microvascular dysfunction is a core pathophysiological mechanism underlying brain lesions associated with T2DM and may constitute a shared mechanistic basis for cognitive impairment and comorbid depression in T2DM patients.3 Although GLP-1RA also possess anti-inflammatory and microvascular regulatory effects, SGLT2i demonstrate more significant improvements in microvascular function through the induction of physiological ketosis, reduction of chronic low-grade inflammation, and promotion of sodium-glucose co-excretion, leading to an energy deficit.
This benefit has extended to non-diabetic populations: SGLT2i have been shown to improve microcirculatory disorders independently of glucose-lowering effects in patients with heart failure or renal impairment.66
To explore the sources of heterogeneity and to evaluate whether the results vary under different characteristics, including region, study design, patient age, and follow-up time, we conducted a subgroup analysis. The results showed that, first, the protective effect of SGLT2i against dementia is more pronounced in the Asian population. ApolipoproteinE (APOE) epsilon4 is associated with increased production of β-amyloid, tau pathology, neuroinflammation, and disruption of the blood-brain barrier, making it the strongest genetic risk factor for Alzheimer’s disease.67 Farrer et al have confirmed that the risk amplification effect of APOE epsilon4 for Alzheimer’s disease is stronger in Asians than in Caucasians.68 SGLT2i primarily exert neuroprotective effects by reducing Aβ production, inhibiting tau proteins, suppressing neuroinflammation, and enhancing cerebral blood flow.
Whether this protective effect is mediated by APOE epsilon4 warrants further investigation. Second, depression-related outcomes improved significantly in the European cohort, while there was no difference observed in the American population. The high treatment accessibility, fewer individuals with low socioeconomic status (SES), and moderate inflammatory background in Europe may have contributed to the more pronounced improvement in depression-related outcomes.
In the American population, high comorbidity, polypharmacy, and SES gradients may have masked the independent antidepressant effects of SGLT2i.69,70 Third, compared with studies restricted to older adults, the subgroup of studies that enrolled younger adults (minimum age 18 years, without an upper age restriction) demonstrated more significant effects on both dementia outcomes and depression-related outcomes. Younger patients with T2DM typically have a shorter disease duration, and clinicians are more inclined to initiate SGLT2i or GLP-1RA early to balance weight loss and long-term cardiovascular and renal benefits. In contrast, prescriptions for older patients tend to be more cautious due to concerns about adverse reactions.
This prescribing bias may have amplified the observation that “the effects are more pronounced in the younger group”. In addition, young adults generally have lower overall levels of inflammation than the elderly, which gives SGLT2i a greater advantage in alleviating chronic low-grade inflammation.66,71 Fourth, the observed difference in risk reduction between the two medications in the subgroup with long-term follow-up and the lack of benefits in dementia risk in the case-control study subgroup indicate that the benefits related to dementia and depression are time-dependent. Only by long term, continuous treatment can the delayed protective effects of SGLT2i be fully realized.
Several limitations of this study should be considered when interpreting the results. First, the total number of included studies was limited, with pooled estimates for each outcome derived from no more than seven studies, and even fewer available for subgroup analyses. Second, although included studies employed propensity score methods and other statistical techniques to adjust for confounding, observational studies are inherently susceptible to residual confounding and clinical biases.
Therefore, the observed differences in incidence rates do not imply causal effects, and definitive conclusions await confirmation through head-to-head RCTs. Third, considerable heterogeneity was observed for some outcomes. We used a random-effects model for pooling and conducted subgroup and sensitivity analyses to explore and mitigate the impact of heterogeneity.
Fourth, due to the limited nature of the original research reports, we were unable to conduct subgroup analyses based on social demographic characteristics such as sex, socioeconomic status, or education level, nor could we explore the mechanisms related to biomarkers such as amyloid plaque load, blood-brain barrier integrity, or microglial cell function.
Conclusion
In conclusion, this meta-analysis of observational studies found that, among patients with type 2 diabetes, SGLT-2i users had lower rates of all-cause dementia, Alzheimer’s disease, vascular dementia, suicide/self-harm events, and depression/anxiety diagnoses compared with GLP-1RA users. Subgroup analyses suggested that factors such as age and geographic region may influence the magnitude of these differences. These findings suggest that SGLT-2i use may be associated with lower rates of neuropsychiatric outcomes, although causality cannot be established.
Future large-scale RCTs are warranted to validate these findings and elucidate the underlying mechanisms, and subsequent studies are encouraged to report outcomes stratified by sex, socioeconomic status, and educational level.