Association of Sodium–Glucose Cotransporter 2 Inhibitors With Time to Dementia: A Population-Based Cohort Study

Type 2 diabetes (T2D) is an important risk factor for dementia (1), a major contributor to overall disease burden, and a threat to functional independence among older people. However, there is limited clinical guidance on preventative strategies for dementia in older adults with T2D (2,3). Glucose-lowering medications may mitigate some of this risk, but not all classes are alike, including second-line therapies that are often substituted for, or added to, first-line agents to improve glycemic control.

Sodium–glucose cotransporter 2 (SGLT2) inhibitors, a newer class of oral glucose-lowering medications, are positioned as second-line therapies. Preclinical data indicate that SGLT2 inhibitors may mitigate various dementia pathologies (4), but existing real-world evidence for the association of SGLT2 inhibitors with dementia risk has been limited to three case-control studies (5–7) and two cohort studies (8,9). Previous cohort studies were limited by potential reverse causality (protopathic bias) (10) because the prescriptions may be influenced by early signs or symptoms of the outcome. For example, people with cognitive symptoms may have been considered at risk for dehydration and therefore less likely to have received an SGLT2 inhibitor. It remains to be determined whether the new use of SGLT2 inhibitors was associated with a lower dementia risk. Using an active-comparator, new-user cohort design with application of lag time into the start of at-risk time in large population-based administrative data (11,12), this study aimed to determine whether SGLT2 inhibitor use was associated with a lower risk of dementia.

Dipeptidyl peptidase 4 (DPP-4) inhibitors are also common second-line oral glucose-lowering medications. In large, randomized controlled trials of dementia-free people with T2D, linagliptin was shown to have no effects on cognition (13,14). Therefore, the new use of SGLT2 inhibitors was compared with the new use of DPP-4 inhibitors in this study to reduce confounding by indication and to infer whether SGLT2 inhibitors may offer more cognitive benefit.

This population-based retrospective cohort study used and linked multiple prospectively collected databases housed at ICES (https://www.ices.on.ca/DAS). Variables were ascertained from the following sources: Ontario Drug Benefit (ODB) database, Ontario Diabetes Database, Continuing Care Reporting System, Registered Persons Database, Ontario Health Insurance Plan (OHIP) Physician Billing database, National Ambulatory Care Reporting System, Discharge Abstract Database, and ICES Physician Database. All Ontario residents are eligible for OHIP coverage at any age, but only Ontario residents covered by OHIP and aged >65 years are eligible for the ODB program. Medication data for individuals >65 years old were identified from the ODB database. More information on each database is summarized in Supplementary Table 1. The study was approved by the research ethics board at Sunnybrook Health Sciences Centre (PIN: 4839).

Ontario residents covered by OHIP and aged >66 years were included to ensure new users could be defined based on a 1-year lookback window. People who were dispensed an SGLT2 inhibitor or a DPP-4 inhibitor from 1 July 2016 to 31 March 2021 entered the study. The study period started 1 year after the formulary listing of SGLT2 inhibitors to accurately identify new users. The cohort entry date was the earliest date that an SGLT2 inhibitor or a DPP-4 inhibitor was dispensed to an individual in the data.

Individuals newly dispensed an SGLT2 inhibitor or a DPP-4 inhibitor without prescriptions of either drug class in the past year were considered as new users and included. Those on both SGLT2 inhibitors and DPP-4 inhibitors, multiple SGLT2 inhibitors, or multiple DPP-4 inhibitors on the cohort entry date were excluded.

A 5-year lookback window was used for medical history. People with a history of dementia or people residing in long-term care at baseline were excluded. Long-term care residence status was ascertained from the Continuing Care Reporting System. People discharged from hospital in the 2 days before the cohort entry date were excluded to mitigate misclassification due to in-hospital exposures (15). The 1-year lag time was implemented to mitigate reverse causality and to address the issue of disease latency, and people with censoring or events <1 year from cohort entry were excluded.

Use of an SGLT2 inhibitor (canagliflozin, dapagliflozin, and empagliflozin) or a DPP-4 inhibitor (linagliptin, saxagliptin, and sitagliptin) was identified as drugs covered under ODB. Since dementia is irreversible and chronic, the primary analysis used an “intention-to-treat” approach, which did not censor data beyond initiation of the other drug (i.e., crossover) or discontinuation of baseline treatments.

The outcome was time to incident dementia. The date of incident dementia was defined using a validated algorithm for Alzheimer’s disease and related dementias: one hospitalization with a dementia record from Discharge Abstract Database, or three physician claims (earliest claim) for dementia in the OHIP database at least 30 days apart in a 2-year period, or a cholinesterase inhibitor prescription (16). When individuals met multiple criteria, the earliest identification date was picked. Previously, the algorithm showed a sensitivity of 79.3%, a specificity of 99.1%, a positive predictive value of 80.4%, and a negative predictive value of 99.0% (16).

Censoring was defined as death date, end of OHIP eligibility, or end of the study period. All-cause mortality was ascertained from the Registered Persons Database. To reduce the concern of disease latency and mitigate reverse causality that results from potential prescription preference caused by early symptoms of undiagnosed dementia (10), 1-year lag time from cohort entry was applied to the observation window (i.e., at-risk time starting 1 year after treatment initiation) (12).

All analyses were conducted in SAS 9.4. Confounders were addressed using inverse probability treatment weighting for average treatment effects among the treated (17), based on propensity scores derived from a logistic regression model that incorporated covariates listed in Supplementary Table 2. Covariates with a between-group standardized mean difference <0.1 were considered balanced (17). A 5-year lookback window before cohort entry was applied to baseline comorbidities, and a 120-day lookback window before cohort entry was applied to baseline exposure to other glucose-lowering agents and medications for other indication. Covariate coding definitions are described in Supplementary Table 3 and were adapted from a previous study from ICES (15).

Adjusted hazard ratios (aHRs) with 95% CIs between groups were obtained using Cox proportional hazards models weighted by inverse probability of treatment. Weighted Kaplan-Meier curves were used to visualize cumulative incidence proportions of each exposure group over time, and to examine whether there was a substantial time-dependent trend in relative hazards that might contribute to violation of the proportional hazard assumption.

To examine associations across different SGLT2 inhibitors, the SGLT2 inhibitor group in the primary analysis (with an intention-to-treat approach) was further stratified, in which each SGLT2 inhibitor was compared individually to the DPP-4 inhibitor group, and individual HRs were obtained in a pairwise manner.

Subgroup analyses by >75 years versus <75 years of age and by sex were conducted to explore whether the associations were consistent across subgroups. Interaction was tested using a likelihood ratio test.

To explore the association with continuous use of either drug class, a secondary analysis using an “as-treated” approach (i.e., censoring data beyond baseline treatment discontinuation or crossover) was conducted. Continuous use was defined as the next prescription dispensed during days of supplies of a previous prescription plus a 30-day grace period. In this as-treated approach, the start of at-risk time was still 1 year after cohort entry. However, as discontinuation or crossover may be related to early symptoms of undiagnosed dementia, we applied an additional observation window up to 1 year after discontinuation or crossover.

To investigate potential informative censoring due to all-cause mortality, a Fine-Gray model with all-cause mortality as competing risk was used to obtain an adjusted sub-distribution HR. Incident cataract surgery was selected as a negative control outcome (18), because DPP-4 or SGLT2 inhibitor use has no theoretical treatment effect on it, but it may share potential mechanisms of residual confounding with incident dementia, including health care literacy and access. Cataract surgery was identified from OHIP billing codes (E984, E950, E214, E138, E140–141, and E143–E146). An E value was calculated to examine robustness to an unmeasured confounder (19).

Of 106,903 new users, 36,513 initiated an SGLT2 inhibitor and 70,390 initiated a DPP-4 inhibitor (Fig. 1). Highlighted characteristics by treatment groups are summarized in Table 1. Full characteristics are summarized in Supplementary Table 2, and all covariates had a standardized mean difference <0.1 after weighting. Breakdowns of molecules in each drug class are summarized in Supplementary Table 4.

Over a mean follow-up of 2.80 years (SD 1.07) from cohort entry (i.e., treatment initiation) and a total of 192,453 person-years at-risk time, 2,731 incident dementia cases were observed. Use of an SGLT2 inhibitor was associated with lower dementia risk (aHR 0.80 [95% CI 0.71–0.89]) (Fig. 2A and Table 2) compared with a DPP-4 inhibitor.

Across different SGLT2 inhibitors, dapagliflozin showed the lowest dementia risk (aHR 0.67 [95% CI 0.53–0.84]), followed by empagliflozin (aHR 0.78 [95% CI 0.69–0.89]), when compared with DPP-4 inhibitors. However, no association was found when canagliflozin was compared with DPP-4 inhibitors (aHR 0.96 [95% CI 0.80–1.16]).

There was a trend that female subjects (aHR 0.72 [95% CI 0.61–0.86]) versus Male subjects (aHR 0.85 [95% CI 0.74–0.99]) showed greater associations in dementia risk between users of SGLT2 versus DPP-4 inhibitors, but the interaction between sex and drug class was not significant (χ² = 2.07; df = 1; P = 0.150). The effect was comparable between people >75 years of age (aHR 0.78 [95% CI 0.67–0.91]) and those <75 years of age (aHR 0.84 [95% CI 0.72–0.996]), with no interaction detected (χ² = 0.46; df = 1; P = 0.497). The subgroup analyses are summarized in Table 2.

In the secondary analysis using an as-treated approach, there was a mean follow-up of 2.09 years from cohort entry (SD 0.96) and a total of 117,024 person-years at-risk time and 1,696 incident dementia cases. Ongoing SGLT2 inhibitor use without co-exposure to DPP-4 inhibitors was associated with lower dementia risk (aHR 0.66 [95% CI 0.57–0.76]) (Fig. 2B and Table 2) compared with ongoing DPP-4 inhibitor use without co-exposure to SGLT2 inhibitors.

A breakdown of the reasons for censoring in the intention-to-treat analysis and the as-treated analysis is summarized in Supplementary Table 5.

In the negative control outcome analysis, there was no difference in time to cataract surgery between SGLT2 and DPP-4 inhibitor users (aHR 1.02 [95% CI 0.97–1.07]).

For the intention-to-treat analysis, the E value was 1.82 (lower limit 1.49), which means an unmeasured confounder would require a risk ratio of 1.82 associated with both the exposure and outcome to explain away the observed risk, and a confounder with a risk ratio <1.49 could not.

The effect size was comparable to the primary result when a competing risk model was used (aHR 0.81 [95% CI 0.72–0.90]).

This cohort study demonstrated that the new use of SGLT2 inhibitors compared with new use of DPP-4 inhibitors was associated with lower dementia risk in people with T2D aged >66 years. In the intention-to-treat analysis, a 20% reduction in dementia risk was observed. The result was robust in a sensitivity analysis incorporating all-cause mortality as competing risk. In the as-treated analysis, there was a 34% reduction in dementia risk. Discontinuation happened more frequently among DPP-4 inhibitor users, and some DPP-4 inhibitor or SGLT2 inhibitor users initiated the other treatment after cohort entry. The larger estimate in the as-treated analysis might reflect that continuous SGLT2 inhibitor use, switching to an SGLT2 inhibitor, or adding an SGLT2 inhibitor to a DPP-4 inhibitor had greater cognitive benefit.

Using multiple strategies to mitigate bias, the present cohort study strengthened previous findings that SGLT2 inhibitor use was associated with better cognitive performance over time or a lower dementia risk (8,9,20). A previous study compared current SGLT2 inhibitor users with nonusers (20), whereas the current study used an active-comparator new-user design comparing two second-line medications to reduce susceptibility to confounding by indication and to avoid bias caused by adjustment of postexposure covariates (11). In Siao et al. (9), the new use of SGLT2 inhibitors compared with other glucose-lowering agents was associated with a 11% lower dementia risk. The smaller estimate may have been due to potential selection bias introduced by differential study group selection, as well as exclusion of events during the use of other agents prior to the first SGLT2 inhibitor prescription. To minimize the potential for such selection bias, we defined cohort entry as the first prescription of an SGLT2 inhibitor or a DPP-4 inhibitor to an individual in the data, and then identified new users based on a washout period. Another methodological strength of the current study was the implementation of lag time from drug initiation to mitigate reverse causality and issues related to disease latency (12), which were not considered in prior cohort studies (8,9). We speculated that the larger estimate reported in Mui et al. (8) (59% lower risk for SGLT2 inhibitors vs. DPP-4 inhibitors) might partly reflect reverse causality. Although we observed a smaller reduction in dementia risk (20–34%), the level of this more robust estimate remains of a magnitude that would be clinically important.

According to current guidelines for diabetes management in older people, there is no clear evidence to recommend interventions to prevent dementia in T2D (2,3). Most observational studies selected non-users as the reference group; consequently, it is difficult to infer which glucose-lowering agent may be more neuroprotective, because nonuser groups were highly heterogeneous across studies (21). Among existing active-comparator, new-user cohort studies, DPP-4 inhibitor use was associated with a lower dementia risk compared with sulfonylurea use (22), and use of metformin versus a sulfonylurea exhibited a modest association with lower risk (23,24). In two large randomized controlled trials, linagliptin compared with placebo or glimepiride showed no effects on cognitive performance in dementia-free people with T2D (13,14). Those studies in toto suggest that metformin, sulfonylureas, and linagliptin are unlikely to be preferred as interventions to prevent dementia in T2D. In that context, it is important that the new use of SGLT2 inhibitors, especially dapagliflozin or empagliflozin, compared with DPP-4 inhibitors, was associated with a 20–34% lower risk of dementia. The finding may have important implications in optimizing pharmacotherapy to slow cognitive decline in older people with T2D, and it may inform opportunities to examine dapagliflozin or empagliflozin as interventions to prevent dementia in randomized controlled trials.

The current study does not speak to mechanisms of neuroprotection, but the result may be biologically plausible. Different SGLT proteins are expressed in the central nervous system, and they have been proposed to play a role in neuronal survival (4). In preclinical studies, SGLT2 inhibitors reduced Alzheimer’s disease pathology, such as amyloidosis and vascular brain damage (4). In several large new-user cohort studies, SGLT2 inhibitors compared with DPP-4 inhibitors were associated with a modest 15% lower stroke risk under an as-treated approach (25–28). Apart from the possibility of direct effects on the brain or cerebral vessels, SGLT2 inhibitors versus DPP-4 inhibitors have been associated with lower risk of heart failure (25–28) and with renal protection (15,29), which could indirectly benefit the brain. Future clinical studies might examine vascular risk factors, Alzheimer’s disease biomarkers, and cerebral microangiopathy as possible mediators of neuroprotection and cause-specific dementia outcomes.

The current study provides a novel message: that cognitive benefit may be specific to certain SGLT2 inhibitors. Unlike dapagliflozin and empagliflozin, canagliflozin exhibited a HR toward the null. The reason is unclear, but canagliflozin is known to be less selective at SGLT2 versus SGLT1 (30). Although any underlying mechanisms would be speculative, it is possible that selective inhibition of SGLT2 may be more neuroprotective, requiring further investigation. In addition, this study suggests that biological sex could be an effect modifier, but the population size was insufficient to reach a conclusion, suggesting the need to examine this explicitly in larger studies, and that future studies might consider stratifying for sex.

It should be noted that the comparison between SGLT2 inhibitors and glucagon-like peptide 1 receptor agonists (GLP1-RAs) has yet to be made. In dementia-free people with T2D, GLP1-RAs versus placebo have been suggested to slow cognitive decline (31,32), liraglutide versus dapagliflozin showed better cognitive performance in a small, randomized trial (33), and a cohort study concluded that long-term GLP1-RA use was associated with a lower dementia risk (32). However, GLP1-RAs were not covered by ODB in Ontario until 2019, and users of GLP1-RAs who self-paid or paid by private insurance could not be identified, so GLP1-RAs could not be investigated in the current study. Furthermore, the present results cannot speak to the potential of any agent to slow cognitive decline in dementia since the population did not include individuals diagnosed with dementia at baseline. A prior study showed the associations of glucose-lowering drugs with cognitive decline might differ by the presence of Alzheimer disease in people with diabetes (34). Future studies examining cognitive outcomes in SGLT2 inhibitor users with cognitive impairment would be informative, and two studies demonstrated that empagliflozin was associated with improved global cognition over 1 month in frail older adults with diabetes who were cognitively impaired (35,36).

While a large number of observed events was a strength, a relatively short mean follow-up time was an important limitation of the current study because development of dementia is generally a chronic process. We implemented a 1-year lag time to address potential reverse causality and disease latency; however, when data over longer follow-up periods are available, future studies might also consider longer lag periods. As another limitation, unmeasured confounding by baseline HbA_1c, BMI, estimated glomerular filtration rate, frailty, and blood pressure measures is possible, because these factors, especially kidney function, can be considerations in decisions between SGLT2 and DPP-4 inhibitors (2,37); however, the study attempted to minimize confounding by indication by an active-comparator new-user design, and to mitigate such unmeasured confounding by balancing proxy variables, such as diabetes duration, insulin use, diabetes complications, chronic kidney disease, hypertension, Charlson Comorbidity Index, and obesity based on administrative records. A null association with the negative control outcome reflected that the estimates between SGLT2 versus DPP-4 inhibitors were likely robust to unmeasured confounding via mechanisms such as differential health care access, and the E value indicated that an unmeasured confounder would need to be large to explain away the observed associations. Furthermore, because SGLT2 inhibitors have been associated with lower all-cause mortality compared with DPP-4 inhibitors (25–28), detection bias resulting from lower mortality among SGLT2 inhibitors might have underestimated the association with dementia; nonetheless, a sensitivity analysis with a Fine-Gray model suggested that informative censoring related to mortality might not have been a major concern in this study. Lastly, the OHIP claim data did not differentiate dementia subtypes, and therefore, time to different types of dementia could not be ascertained.

In conclusion, this large population-based cohort study demonstrated a clinically meaningful association of SGLT2 inhibitor use with lower dementia risk over a mean follow-up of 2.8 years. The associations varied across different SGLT2 inhibitors, such that dapagliflozin and empagliflozin were associated with lower dementia risk, while no association was seen for canagliflozin. Although multiple approaches were adopted to qualify and minimize potential bias, they do not preclude the possibility of residual confounding, and well-designed randomized controlled trials are needed.

Acknowledgments. This study contracted ICES Data & Analytic Services and used data from the ICES Data Repository, which is managed by ICES with support from its funders and partners: Canada’s Strategy for Patient Oriented Research, the Ontario Strategy for Patient Oriented Research Support Unit, the Canadian Institutes of Health Research, and the Government of Ontario. The opinions, results and conclusions reported are those of the authors. No endorsement by ICES or any of its funders or partners is intended or should be inferred. The authors thank IQVIA Solutions Canada Inc. for use of their Drug Information Database. Parts of this material are based on data and information compiled and provided by the Canadian Institute for Health Information. However, the analyses, conclusions, opinions, and statements expressed herein are those of the author, and not necessarily those of Canadian Institute for Health Information. The authors also thank Stefana Jovanovska (ICES) and Chelsi Major-Orfao (Sunnybrook Research Institute) for the important contribution to project coordination and administration.

Funding. W.S. was supported by the Canadian Institutes of Health Research (grant/award PJT-159711); the Natural Sciences and Engineering Research Council of Canada (grant/award RGPIN-2017-06962); Alzheimer’s Association and Fondation Brain Canada (grant/award AARG501466); and Weston Brain Institute, Alzheimer’s Research Trust, Alzheimer’s Association, and Michael J. Fox Foundation for Parkinson’s Research (grant/award BAND3). The work was supported in part through funding from the Canada Research Chairs Program. C.-Y.W. was supported by the University of Toronto Banting & Best Diabetes Centre (Tamarack Graduate Award in Diabetes Research) and Canadian Institutes of Health Research (Doctoral Research Award, Canada Graduate Scholarships, 202111FBD-476273-75801). R.H.S. receives salary support for research from a Heart & Stroke Foundation Clinician-Scientist Phase II Award, the Dr. Sandra Black Centre for Brain Resilience and Recovery, the Sunnybrook Department of Medicine, and salary support for administrative leadership from the Ontario Neurodegenerative Disease Research Initiative and as Medical Director of the North & East GTA Regional Stroke Network.

The funders had no input into the conduct of the study and the interpretation of results.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. C.-Y.W. and W.S. conceptualized the study and drafted the manuscript. C.I., C.W., and R.S. had significant input into data curation for the analysis. C.-Y.W., C.I., C.W., B.R.S., J.D.E., M.K.K., H.C.-M., R.S., and W.S. had significant contribution to the study methodology. C.-Y.W., C.I., B.R.S., J.D.E., M.K.K., M.M., R.H.S., H.C.-M., R.S., and W.S. had significant input into the investigation process. C.-Y.W. performed the formal analysis and visualized the results, and L.Y.X. validated the analysis. All authors made significant contributions to the interpretation of the results and the review or editing of the manuscript. R.S. and W.S. were responsible for project administration. W.S. acquired funding for this study. This study contracted ICES Data & Analytic Services and used data from the ICES Data Repository. R.S. as an ICES staff scientist had full access to all of the data in the study. W.S. is the guarantor of this work and, as such, had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Preliminary work of this study was presented in abstract form at the Alzheimer’s Association International Conference 2022, San Diego, CA, 31 July–4 August 2022.