Dementia risk linked to three diabetes drug types: GLP1 receptor agonists, SGLT2 inhibitors, and DPP4 inhibitors

The conduct of the present study was approved by the Clinical Practice Research Datalink (CPRD) Research Data Governance (#23_002545), the Research Ethics Board of Sunnybrook Health Sciences Centre (PIN: 6013), and the Research Ethics Board of University of Toronto (#46463).

The present study used the UK CPRD GOLD and Aurum databases, which routinely collect linkable but de-identified longitudinal data on demographics, medical records (coded using the Read codes or SNOMED-CT classification), clinical measures, and prescriptions from primary care practices [27, 28]. The data were recorded with dates of events; as such, the cohort entry could be constructed based on prescription dates, a lookback window could be used to define pre-exposure eligibility and covariates, and the date of outcomes could be ascertained.

The UK CPRD data were linked to the Hospital Episode Statistics Admitted Patient Care (HES APC) and the Office for National Statistics (ONS) databases, but the linkage was available only in a subset of individuals. The HES APC contains dates of hospital admissions, associated inpatient diagnoses (coded using the International Statistical Classification of Diseases version 10 [ICD-10]), and procedures (coded using the UK Office of Population, Census and Surveys classification [OPCS]) [29]. The ONS contains dates of registered death and associated data [30].

Using an active-comparator new-user cohort design [31, 32], we emulated three hypothetical active-controlled randomized trials comparing SGLT2 inhibitors, injectable GLP1 receptor agonists, and DPP4 inhibitors pairwise. The three glucose-lowering medications are common second-line therapies recommended by the guidelines [8, 9]. Supplementary Table 1 compares the hypothetical target trials and the emulated cohorts. An active-comparator new-user cohort design [31, 32] compares initiators of different active treatments by emulating an active-controlled randomized trial. An active-comparator new-user cohort design is less susceptible to healthy user bias and confounding by indication, because an active treatment with similar indication rather than non-users is the comparator [31, 32]. Target trial emulation intends to formulate a clear causal question, avoid biases that critically threaten internal validity (e.g. immortal time bias), and provide adequately robust observational evidence when randomized trials are not feasible or not yet available [33].

Individuals aged ≥ 60 years and prescribed injectable GLP1 receptor agonists (except weight loss formulations of liraglutide or semaglutide), SGLT2 inhibitors, or DPP4 inhibitors entered the cohort study. When GLP1 receptor agonists were compared to DPP4 inhibitors, the cohort entry started from January 1 st, 2007, to December 31 st, 2022. For the other two comparisons, the cohort entry started from January 1 st, 2013, to December 31 st, 2022. To avoid immortal time bias and selection bias [33, 34], the cohort entry date of each comparison was the first eligible prescription of either study drug within an individual.

To identify new users based on a 1-year washout period, only those with 1-year continuous registration before cohort entry were included. Individuals prescribed either treatment in the past year before cohort entry were excluded. Individuals prescribed multiple molecules within the same class or both treatments at cohort entry were excluded. When GLP1 receptor agonists were studied, individuals prescribed weight loss formations or oral semaglutide at baseline or in the past year before cohort entry were also excluded.

Only individuals with HbA1c ≥ 6.5% (or ≥ 48 mmol/mol) or a type 2 diabetes diagnosis at any time until cohort entry were included. Individuals with type 1 diabetes, thyroid cancer, multiple endocrine neoplasia, or end stage kidney disease at any time until cohort entry or individuals with dialysis or the latest eGFR < 20 mL/min/1.73m2 in the past year were excluded, because they were unlikely to receive SGLT2 inhibitors or GLP1 receptor agonists.

Individuals with dementia diagnoses, records of cognitive deficits (including unspecified or mild cognitive impairment, aphasia, apraxia, and agnosia), or cholinesterase inhibitor or memantine prescriptions at any time until cohort entry were excluded. Because a 1-year lag time was used to address disease latency and mitigate reverse causality resulting from delayed diagnosis or treatment effects [35, 36], those with records of dementia, cognitive deficits, cholinesterase inhibitors, or memantine within 1 year after cohort entry, or those with follow-up less than 1 year were excluded.

Dementia is a chronic outcome with long disease latency. To minimize the potential for informative censoring, the primary analysis used an intention-to-treat exposure definition.

Although an intention-to-treat analysis is less susceptible to informative censoring than an as-treated analysis, it is more prone to bias due to exposure misclassification. Thus, secondary as-treated analyses for continuous use were performed. Treatment discontinuation was defined as no subsequent prescription during the estimated or recorded days of supplies plus a 30-day grace period. The date of crossover was the date of prescription of the other study drug during follow-up.

The endpoint of the intention-to-treat analysis was incident dementia (or dementia subtypes in exploratory analyses), death (based on ONS or estimated dates from CPRD), end of CPRD registration, or end of study period (November 30th, 2024), whichever occurred first. In the as-treated analysis, the endpoint was the outcome of interest, 1 year after crossover or discontinuation, death, end of CPRD registration, or end of study period, whichever occurred first. The purpose of the 1-year extension from crossover or discontinuation was to reduce informative censoring caused by early symptoms or delayed diagnosis of dementia.

Incident all-cause dementia was defined as the date of first Read or SNOMED-CT codes for dementia (Supplementary Tables 6–7). Medical codes of dementia in CPRD were shown to have an 83% positive predictive value (PPV) [37].

Exploratory outcomes were AD and VD. The incident date was the date of the last record contributing to the definition of AD or VD described below.

AD was defined as: (1) an AD diagnosis with a prescription of cholinesterase inhibitors or memantine in any sequence, (2) an unspecified dementia code followed by two prescriptions of cholinesterase inhibitors or memantine, (3) two AD diagnoses on separate dates, (4) an AD diagnosis after a post-baseline record of cognitive assessment, referral to a related specialist (i.e., neurologist, geriatrician or psycho-geriatrician), or neuroimaging assessment (i.e., magnetic resonance imaging, computed tomography, or single photon emission computed tomography), OR (5) an AD diagnosis and a record of cognitive deficits in any sequence. This algorithm showed a 79% PPV for AD in a prior validation analysis [38].

VD was also defined as: (1) two VD diagnoses on separate dates, (2) a VD diagnosis or an unspecified dementia code within 2 years after a stroke diagnosis in primary care records or the primary position of a hospitalization, (3) a VD diagnosis after a post-baseline record of cognitive assessment, referral to a related specialist, or a neuroimaging assessment, OR (4) a VD diagnosis and a record of cognitive deficits in any sequence. This algorithm showed a 73% PPV for VD in a prior validation analysis [39].

Analyses were performed using R 4.4.1 [40]. Propensity scores were estimated using a logistic regression conditional on 105 pre-exposure covariates listed in Supplementary Table 7, including demographics, laboratory measures, healthcare utilization, comorbidities, procedures, and concurrent medications among eligible individuals followed for at least 1 year and without dementia during the first year. Covariates were adjusted using the overlap weights, a propensity score weighting approach that generates no extreme weights and yields exact covariate balance for an average treatment effect in clinical equipoise [41].

Weighted cause-specific hazard ratios (HRs) with death as a competing risk and the 95% confidence intervals (CIs) were estimated from Cox models. Cause-specific hazard models are more appropriate for etiological questions [42]. Weighted Kaplan–Meier curves were used to visualize cumulative incidences and visualize whether the hazards might differ over time. Weighted incidence rate differences (IRDs) and the 95% CIs per 1,000 person-years were estimated, using a generalized linear model with an identity link function and Gaussian distribution [43]. Because a 1-year lag time was used, the at-risk time started at 1 year after cohort entry. Robust variance estimators were applied to account for weighting in regression models.

We performed subgroup analyses by age ≤ 70 years versus > 70 years, by males versus females, by insulin use, by atherosclerotic cardiovascular disease (i.e., myocardial infraction, angina, coronary artery disease, coronary revascularization, peripheral artery disease, aortic atherosclerotic disease, or cerebrovascular disease), by renal insufficiency (i.e., chronic kidney disease or the latest eGFR < 60 mL/min/1.73m²), and BMI ≥ 30 versus < 30 kg/m². Age is a strong dementia risk factor. Females versus males are more likely to be diagnosed with dementia [44]. Insulin use was examined as a proxy for more advanced diabetes. Both SGLT2 inhibitors and GLP1 receptor agonists are recommended for those with atherosclerotic cardiovascular disease and chronic kidney disease, and GLP1 receptor agonists are recommended for people with obesity [8, 9]. These clinically relevant factors informed subgroup analyses.

Individual SGLT2 inhibitor molecules or GLP1 receptor agonist molecules were studied in separate cohorts. In each molecule-specific analysis, the cohort entry was restricted to the time that both treatments were on the market.

Propensity scores were re-estimated, and weights were re-calculated within subgroups and for molecule-specific analyses. Differences in IRDs or HRs between subgroups or across individual molecules were assessed by the Wald test for homogeneity with a 10% significance level.

First, we restricted the follow-up time to a maximum of 5 years from cohort entry to mitigate exposure misclassifications under an intention-to-treat analysis, which can dilute the relative hazards when the follow-up time increases. Second, a 2-year lag time was used to more conservatively address reverse causality. Third, to examine robustness to a longer washout period, one sensitivity analysis excluded individuals with prior use of study drugs in the past 2 years, and the other analysis excluded individuals with prior use at any time. Fourth, Fine-Gray models with death as a competing risk were used to explore the ratio of cumulative incidence function estimates. Fifth, multiple imputation (20 iterations of 20 imputations) was applied to propensity score estimation, to explore the robustness to missing values in smoking, laboratory measures, body mass index (BMI), and number of hospitalizations. The Rubin's rule was used to pool estimates from each iteration. Sixth, to examine robustness to a more specific outcome definition, incident dementia was defined as: a dementia record plus (1) another dementia record, (2) a record of cognitive deficits in any sequence, (3) a prescription of cholinesterase inhibitors or memantine in any sequence, OR (4) a post-baseline record of a cognitive test, referral to a related specialist, or neuroimaging assessment before a dementia record. Seventh, to examine the robustness to a more sensitive outcome definition, the date of first dementia diagnosis or initiation of cholinesterase inhibitors or memantine was defined as the incident date. Eighth, for statistically significant estimates, we performed a quantitative bias analysis using the rule-out approach to quantify the magnitude of an unmeasured confounder needed to explain away the observed association [45].

The crude hazard ratios (HRs), incidence rates, and incidence rate differences (IRDs) are reported in Supplementary Table 12. The observed study endpoints are summarized in Supplementary Table 13. When DPP4 inhibitors were a comparator, 74.8% GLP1 receptor agonist initiators discontinued the treatment without other endpoints over a median follow-up of 1.31 years on treatment; however, only 56.7% SGLT2 inhibitor initiators discontinued the treatment without other endpoints over a median follow-up of 1.28 years on treatment (Supplementary Table 12 and 13).

In the intention-to-treat analysis, SGLT2 inhibitor versus DPP4 inhibitor initiation was associated with a 14% lower risk of all-cause dementia (HR [95% CI]: 0.86 [0.79–0.94]; Fig. 2C). In the as-treated analysis, continuous use of SGLT2 inhibitors was associated with a 30% lower risk of all-cause dementia (HR [95% CI]: 0.70 [0.60–0.82]; Fig. 2D).

Incident all-cause dementia was comparable between those on GLP1 receptor agonists versus SGLT2 inhibitors in the intention-to-treat analysis (HR [95% CI]: 0.98 [0.87–1.11]; Fig. 2E) and in the as-treated analysis (HR [95% CI]: 1.07 [0.85–1.36]; Fig. 2F).

Continuous use of SGLT2 inhibitor versus DPP4 inhibitor initiation was associated with a lower risk of AD (HR [95% CI]: 0.54 [0.41–0.71]) and VD (HR [95% CI]: 0.64 [0.48–0.85]). Continuous use of GLP1 receptor agonists versus DPP4 inhibitors did not show statistically significant estimates for AD (HR [95% CI]: 0.84 [0.60–1.19]) and VD (HR [95% CI]: 0.71 [0.47–1.07]). The risks of AD (HR [95% CI]: 1.13 [0.76–1.69]) and VD (HR [95% CI]: 1.31 [0.88–1.97]) were not significantly higher with continuous use of GLP1 receptor agonists versus SGLT2 inhibitors.

Reduction in all-cause dementia risk with SGLT2 inhibitors versus DPP4 inhibitors was specific to those aged > 70 years (HR [95% CI]: 0.87 [0.77–0.97]). A lower all-cause dementia risk associated with SGLT2 inhibitors was found among those with atherosclerotic cardiovascular disease (HR [95% CI]: 0.80 [0.68–0.95]) and renal insufficiency (HR [95% CI]: 0.80 [0.65–0.99]). Initiation of an SGLT2 inhibitor versus a DPP4 inhibitor was associated with lower dementia risk among people with BMI < 30 kg/m² (HR [95% CI]: 0.82 [0.70–0.94]) but not among those with BMI ≥ 30 kg/m² (HR [95% CI]: 0.97 [0.86–1.09]; heterogeneity: p = 0.072). The estimates were consistent across different SGLT2 inhibitors (p = 0.913) when DPP4 inhibitors were a referent comparator.

In the cohort of GLP1 receptor agonists versus SGLT2 inhibitors, the HR for all-cause dementia was 0.90 (95% CI: 0.75–1.07) among females, and the HR was 1.06 (95% CI: 0.91–1.24) among males; nonetheless, heterogeneity was not statistically significant (p = 0.163). GLP1 receptor agonist versus SGLT2 inhibitor initiation showed an HR of 0.90 (95% CI: 0.79–1.04) among those with BMI ≥ 30 kg/m² and an HR of 1.07 (95% CI: 0.83–1.38) among those with BMI < 30 kg/m², but the heterogeneity was not statistically significant (p = 0.241).

The estimates from sensitivity analyses are summarized in Supplementary Figs. 5–7. The results across sensitivity analyses were consistent with those of the primary analysis.

The quantitative bias analysis suggested that moderate unmeasured confounding could potentially explain away an HR of 0.86 observed with SGLT2 inhibitors versus DPP4 inhibitors, whereas stronger unmeasured confounding is needed to explain away larger HRs observed in as-treated analyses (Supplementary Fig. 8 and 9).

First, residual confounding is possible in observational studies due to unmeasured confounders such as frailty, neuropsychiatric measures (e.g. Mini-Mental State Examination), socioeconomic status, education, and family history of dementia. These unmeasured confounders could account for the observed benefit to some extent, and the results should be cautiously interpreted. Randomized controlled trials are needed to confirm the observational evidence. Second, time-varying confounding due to informative censoring could be present in as-treated analyses, while an intention-to-treat analysis is not susceptible to treatment-related informative censoring. Third, dementia diagnoses are usually made in a secondary care setting, so dementia incidence in primary care could have been delayed or missed. The relative and absolute measures could have been underestimated by non-differential outcome misclassification. Fourth, the analyses with dementia subtypes should be considered exploratory and interpreted with caution, because the diagnoses were not necessarily biomarker-confirmed, and the referent standard in the validation studies [38, 39] was not biomarker-confirmed diagnoses. Fifth, BMI, as well as eGFR, was ascertained based on the latest available measure during the 1-year lookback window; therefore, the measures in some people might be less accurate measures due to missing data at the time of drug initiation. Inaccuracy in subgroup classification might dilute the observed magnitude of effect modification. Sixth, because the exposure was based on prescription records, it is unknown whether the prescriptions were dispensed, which might contribute to observed effect estimates smaller here than those in studies that used dispensing data.