HbA1c trajectories after SGLT2 inhibitor initiation and the risk of dementia

Introduction

Diabetes mellitus represents a rapidly growing global health challenge, with its prevalence projected to reach more than 1.3 billion adults worldwide by 2050. ¹ Beyond its cardiovascular, renal, and metabolic complications, dementia has emerged as one of the most consequential long-term sequelae, both clinically and from a public health perspective.^2–5 Globally, more than 55 million people are currently living with dementia, a figure expected to more than double by 2050. ⁶ Substantial epidemiologic evidence indicates that individuals with diabetes face a 40–60% higher risk of dementia compared with those without diabetes, with particularly strong associations for vascular dementia.^2–5 Earlier age at diabetes onset, longer disease duration, poor glycemic control, and recurrent hypoglycemia further magnify this risk.^3,7,8 These data underscore the intertwined epidemics of diabetes and dementia in aging societies and highlight the urgent need for strategies that extend beyond glycemic targets to encompass long-term neurocognitive outcomes.

Accumulating evidence links glycemic control with dementia risk in diabetes, yet most studies have relied on a single HbA1c measurement. Higher HbA1c levels, greater variability, and longer diabetes duration have each been associated with elevated risk of cognitive decline and dementia.^2,4,7,9,10 However, a static value cannot capture treatment response or the dynamic course of glycemic control, nor does it account for hypoglycemia or glycemic instability, which are themselves risk factors.^1,11 Indeed, recent cohort and meta-analytic data suggest that stable HbA1c trajectories within the moderate range confer the lowest dementia risk, whereas persistently high or highly variable values substantially increase risk.^2,10 These findings underscore the need for more refined metrics, such as HbA1c trajectories, that reflect the quality and stability of glycemic control over time, and may more accurately predict long-term neurocognitive outcomes.^4,7,9,10

Sodium–glucose cotransporter 2 inhibitors (SGLT2is) have become central in type 2 diabetes mellitus (T2DM) management, offering not only modest glucose-lowering (HbA1c reduction of ∼0.5–1.0%) but also robust cardiovascular and renal protection.^12–16 Evidence now suggests potential neuroprotective effects, with population-based studies showing reduced risks of dementia and Parkinson's disease among SGLT2i users.^14,17–19 Yet glycemic response to SGLT2i is heterogeneous: while some patients achieve marked HbA1c improvement, others show little change, despite accruing cardio-renal benefits.^15,18 Importantly, there are no clear clinical guidelines on how to interpret post-treatment HbA1c trajectories to guide therapeutic intensification. This gap underscores the need to evaluate whether early HbA1c dynamics after SGLT2i initiation can serve as a prognostic marker for long-term neurocognitive outcomes.

Despite accumulating evidence linking glycemic control to dementia risk, nearly all prior studies have relied on single time-point HbA1c values, which cannot capture treatment response or dynamic change. Long-term high HbA1c levels (≥9%) are strongly associated with future dementia risk, whereas values maintained in the 6–8% range confer lower risk; greater variability and lower time in range are each linked to higher incidence of cognitive decline.^9–11,20 Current ADA standards caution that intensive glycemic lowering (HbA1c < 6%) does not improve cognition and may increase hypoglycemia risk. ¹ Although HbA1c trajectories have demonstrated prognostic value for cardiovascular and renal outcomes, their role as early markers of dementia risk has scarcely been investigated.^9–11,20 In particular, whether glycemic response following therapy initiation can serve as an early indicator of long-term neurocognitive outcomes remains unknown.^1,11,20–22 Against this backdrop, we designed the present study to evaluate whether one-year HbA1c trajectories after initiation of SGLT2i provide clinically meaningful information about subsequent dementia risk. Clarifying this association could establish a simple and scalable biomarker to inform timely treatment intensification and contribute to the prevention of diabetes-related dementia.

Methods

HbA1c categories and trajectories

Baseline HbA1c was categorized as Good (<7.0%), Fair (7.0–8.9%), or Poor (≥9.0%). Within each baseline category, patients were classified according to whether their glycemic status at the follow-up assessment improved, remained stable, or worsened, based on transitions between categories. For parsimony, composite contrasts were also analyzed, such as “any worsening” among patients with good baseline control and “any improvement” among those with poor baseline control. The associations of HbA1c trajectories with risks of dementia and mortality are reported in Table 1.

OPEN IN VIEWER

Table 1.

Associations of HbA1c trajectories after SGLT2 inhibitor initiation with risks of all-cause dementia, dementia subtypes, and mortality.

Trajectory	All-cause dementia			Vascular dementia			Alzheimer's disease			All-cause mortality
	HR (95% CI)	p	E-value (E-CI)	HR (95% CI)	p	E-value (E-CI)	HR (95% CI)	p	E-value (E-CI)	HR (95% CI)	p	E-value (E-CI)
Baseline Good (HbA1c < 7.0%)
Good → Good (ref.)	ref.	—	—	ref.	—	—	ref.	—	—	ref.	—	—
Good → Fair	1.07 (0.85, 1.32)	0.565	—	1.28 (0.85, 1.92)	0.237	—	0.89 (0.65, 1.20)	0.440	—	1.08 (0.90, 1.29)	0.407	—
Good → Poor	0.61 (0.26, 1.40)	0.241	—	—	—	—	0.82 (0.22, 3.06)	0.772	—	1.55 (0.97, 2.45)	0.063	—
Baseline Fair (HbA1c 7.0–8.9%)
Fair → Good	1.08 (0.94, 1.23)	0.246	—	1.22 (0.93, 1.57)	0.143	—	1.05 (0.88, 1.25)	0.605	—	1.10 (0.98, 1.23)	0.101	—
Fair → Fair (ref.)	ref.	—	—	ref.	—	—	ref.	—	—	ref.	—	—
Fair → Poor	1.39 (1.11, 1.74)	0.004	2.13 (1.46)	1.53 (1.03, 2.25)	0.032	2.43 (1.21)	1.42 (1.03, 1.94)	0.030	2.19 (1.21)	1.47 (1.21, 1.78)	<0.001	2.30 (1.71)
Baseline Poor (HbA1c ≥ 9.0%)
Poor → Good	0.64 (0.51, 0.79)	<0.001	2.50 (1.85)	0.49 (0.33, 0.73)	<0.001	3.50 (2.08)	0.80 (0.59, 1.07)	0.129	—	0.71 (0.60, 0.84)	<0.001	2.17 (1.67)
Poor → Fair	0.72 (0.62, 0.82)	<0.001	2.12 (1.74)	0.65 (0.51, 0.83)	0.001	2.45 (1.70)	0.82 (0.67, 0.99)	0.044	1.74 (1.11)	0.68 (0.61, 0.76)	<0.001	2.30 (1.96)
Poor → Poor (ref.)	ref.	—	—	ref.	—	—	ref.	—	—	ref.	—	—

HR: hazard ratio; CI: confidence interval; ref.: reference group; E-CI: E-value for the confidence interval bound closer to the null.

Notes: HRs and 95% CIs estimated from Cox proportional hazards models after propensity score matching. Reference groups were patients who remained in the same HbA1c category (Good→Good, Fair→Fair, Poor→Poor). Glycemic control: Good (HbA1c < 7.0%), Fair (7.0–8.9%), Poor (≥9.0%). Follow-up began on day 456 after SGLT2i initiation. Results for Other/Unspecified dementia and composite trajectory contrasts are provided in the Supplemental Material.

Bold values indicate p < 0.05. E-values are shown only for statistically significant results; dashes (—) indicate non-significant findings or reference groups. The E-value indicates the minimum HR that an unmeasured confounder would need to have with both exposure and outcome to fully explain away the observed association; E-CI applies this threshold to the confidence interval bound.

Results

Discussion

In this large, propensity-matched cohort of patients initiating SGLT2is, we found that one-year HbA1c trajectories strongly predicted subsequent dementia risk, fully consistent with our a priori hypothesis. Patients with Good baseline control (<7.0%) showed no excess risk when HbA1c worsened modestly, with hazard ratios for all-cause and subtype-specific dementia remaining null (Table 1; Supplemental Tables 4–6; Supplemental Figure 1A–D, O). By contrast, among patients with intermediate (Fair, 7.0–8.9%) baseline control, worsening to Poor control was consistently associated with increased incidence of dementia—including vascular and other dementia subtypes—and elevated mortality (HR for all-cause dementia, 1.39; 95% CI, 1.11–1.74) (Table 1; Supplemental Table 7; Supplemental Figure 1K, O). For patients with Poor baseline control (≥9.0%), improvement to Fair or Good categories was strongly protective, reducing risks of all-cause, vascular, and other dementias as well as mortality (HR for all-cause dementia, 0.64; 95% CI, 0.51–0.79 for Poor→Good; HR 0.72; 95% CI, 0.62–0.82 for Poor→Fair) (Table 1; Supplemental Tables 8–10; Supplemental Figure 1B–D, E–G, I, L–N, P–R). These findings provide clinically actionable evidence linking glycemic dynamics to neurodegenerative risk, particularly Alzheimer's disease and related dementias. These findings highlight a clinically actionable gradient: failure to maintain or improve control when baseline HbA1c is ≥7.0% identifies patients at heightened neurocognitive risk, whereas marked improvement from ≥9.0% confers substantial benefit. Current ADA/EASD consensus guidelines recommend timely pharmacologic intensification when glycemia remains above target despite SGLT2i therapy, including GLP-1 receptor agonists such as semaglutide or tirzepatide, additional oral agents, or insulin when indicated.^31–33 Our results provide empirical support for these recommendations, underscoring that persistent or worsening hyperglycemia after SGLT2i initiation should not be viewed as benign but instead as a call for prompt escalation to mitigate both dementia and survival risks. Our systematic E-value analysis further confirms that substantial unmeasured confounding would be required to fully explain the observed associations, with point estimates ranging from 2.12 to 3.50, and the strongest robustness observed for the protective effect of glycemic improvement on vascular dementia risk. ²⁹ Furthermore, the use of unified records and standardized mapping within TriNetX minimized biases from fragmented care across institutions.^23,24,34 Additionally, our dual statistical approach using cause-specific hazards and cumulative incidence functions ensured both robust etiologic inference and accurate absolute risk estimation in the presence of competing mortality.^27,28

Prior studies have consistently linked glycemic control with dementia risk, but most have relied on single HbA1c values rather than dynamic trajectories. Large cohort studies from the UK and Sweden have shown that chronic hyperglycemia is associated with elevated dementia risk, particularly of the vascular subtype,^4,9 and decades-long follow-up in JAMA Neurology demonstrated that higher cumulative HbA1c exposure predicted dementia incidence. ¹⁰ More recent work has emphasized the role of long-term stability and time in range, with HbA1c variability and instability associated with higher risk of cognitive decline.^20,35 However, none of these studies directly evaluated whether treatment response to a specific class of glucose-lowering therapy could itself serve as a prognostic marker. Our study is the first to show that HbA1c trajectories in the year after SGLT2i initiation carry independent prognostic value. We demonstrate that when baseline HbA1c is ≥7.0%, failure to achieve stability or improvement is not benign but instead portends substantially higher risks of dementia and mortality (Table 1; Supplemental Tables 7–10; Supplemental Figure 1). Conversely, improvement from very poor control (≥9.0%) was associated with pronounced reductions in all-cause, vascular, and other dementias as well as mortality, establishing treatment response itself as a clinically relevant marker of neurocognitive outcomes. This represents a shift from interpreting HbA1c as a static target to recognizing dynamic change, particularly early response to therapy, as central to dementia risk stratification. These findings also align with and extend existing literature. The particularly strong association between worsening control and vascular dementia observed in our data (Supplemental Tables 3–10) is consistent with prior evidence that diabetes accelerates cerebrovascular pathology, which in turn drives vascular cognitive impairment.^21,22 The novelty lies in showing that within the setting of SGLT2i therapy, early glycemic deterioration despite treatment should be interpreted as an early warning signal for adverse neurocognitive and survival outcomes. This perspective underscores that it is not prescription of the drug alone, but the patient's treatment response, that holds prognostic meaning. This perspective reinforces our clinical hypothesis that failure to improve HbA1c after SGLT2i initiation should not be regarded as benign but rather as a signal to intensify therapy.^31–33

Our findings suggest that HbA1c trajectories should be interpreted as actionable early warning signals rather than passive observational markers. For patients with baseline HbA1c between 7.0 and 8.9%, deterioration to poor control despite SGLT2i therapy consistently predicted higher risks of dementia and mortality (Table 1; Supplemental Table 7; Supplemental Figure 1K, O). In such cases, continuing SGLT2i monotherapy is unlikely to be sufficient, and treatment intensification with additional agents is warranted.^31–33 For patients starting with HbA1c ≥ 9.0%, achieving improvement within the first year markedly reduced the risks of all-cause, vascular, and other dementias, as well as mortality (Table 1; Supplemental Tables 8–10; Supplemental Figure 1B–D, E–G, I, L–N, P–R), reinforcing the value of early, aggressive control. These trajectories thus provide a practical trigger point for therapy adjustment, ensuring timely intervention and avoiding missed opportunities for neurocognitive protection.

The biological plausibility of our findings is supported by established mechanisms linking hyperglycemia to neurocognitive decline. Worsening HbA1c reflects sustained metabolic stress, which accelerates cerebrovascular injury, promotes white matter hyperintensity progression and subcortical infarcts, and disrupts blood–brain barrier function, thereby impairing amyloid-β clearance.^36–38 Chronic hyperglycemia also induces systemic and neuroinflammation, a pathway strongly implicated in vascular and mixed dementias. Conversely, improvement in HbA1c may mitigate these processes by reducing cerebrovascular burden, enhancing insulin sensitivity, and lowering the incidence of vascular events. Long-term stability of HbA1c has consistently been associated with lower dementia risk,^1,2,9–11 and our data extend this evidence by showing that patients with very poor baseline control (≥9%) derive substantial cognitive and survival benefits when HbA1c improves. In contrast, patients with baseline HbA1c < 7% did not experience excess dementia risk with modest worsening, suggesting that fluctuations within this low range may remain below the threshold required to trigger significant neuropathology. ²⁰ This gradient helps explain why trajectories carried prognostic value only when baseline HbA1c was ≥7%. Taken together, these mechanistic insights reinforce our central hypothesis: HbA1c trajectories after SGLT2i initiation capture not only glycemic control but also the underlying pathophysiologic processes that drive dementia risk.

The robustness of our findings was reinforced through multiple sensitivity strategies. First, a landmark design was applied to minimize immortal-time bias, with follow-up anchored after post-index HbA1c ascertainment; effect estimates remained consistent across alternative landmark windows (days 181–545; Supplemental Table 1). Second, HbA1c change was modeled categorically, and results were directionally stable across dementia subtypes and mortality endpoints (Table 1; Supplemental Tables 3–10), demonstrating consistency across outcomes, an important criterion for causal inference. Third, the measurement density of HbA1c testing was balanced after matching, and additional adjustment for test frequency and outpatient visits did not materially alter risk estimates (Supplemental Table 11a, b), reducing concern for detection bias or differential surveillance. Finally, dose–response plausibility was evident: trajectory slope analyses showed that greater upward drift in HbA1c was associated with progressively higher risks of dementia and mortality, satisfying the Bradford Hill criterion of biological gradient. Taken together, these analyses strengthen the internal validity of our study, fulfilling key causal criteria: temporality (trajectory change precedes dementia onset), consistency (effects reproducible across subtypes and mortality), biological gradient (risk increases with worsening HbA1c slope), and robustness to confounding (stable after adjustment for health-care utilization). While residual confounding cannot be fully excluded, the convergence of evidence across multiple analytic approaches enhances confidence that early HbA1c deterioration after SGLT2i initiation represents a true risk signal rather than an artifact of study design. Specifically, our approach of treating missingness as an unknown category while balancing measurement density ensures that trajectory differences reflect clinical outcomes rather than data granularity.^26,30

Our results suggest that among patients with baseline HbA1c ≥ 7.0%, one-year HbA1c trajectories should be used as a trigger for treatment escalation rather than passive observation. Patients whose HbA1c worsens after SGLT2i initiation should not remain on monotherapy but should receive timely intensification with complementary agents such as GLP-1 receptor agonists, DPP-4 inhibitors, or insulin when indicated, in accordance with current ADA recommendations.^39,40 Conversely, patients with HbA1c ≥ 9.0% who achieve improvement within the first year experienced substantially lower dementia and mortality risks, underscoring the prognostic value of early stabilization. From a public health perspective, HbA1c trajectory is a low-cost, scalable biomarker that can be readily implemented for risk stratification and early intervention in diabetes populations.^41,42 Because HbA1c monitoring is already embedded in standard diabetes care,^39,40 its dynamic use offers a practical way to identify high-risk individuals and prevent missed opportunities for intervention. From a health economics standpoint, leveraging HbA1c trajectories represents a low-cost, high-yield strategy for population-level dementia prevention, maximizing impact without adding new resource burdens. We therefore call for future clinical trials to incorporate trajectory-based decision rules, testing whether embedding early treatment response into algorithms improves both cognitive and survival outcomes. Such an approach would move beyond static HbA1c thresholds toward precision diabetes management and dementia prevention.

Several limitations merit consideration. First, as an observational analysis, residual confounding cannot be excluded despite robust matching and sensitivity checks. Second, dementia diagnoses were based on ICD codes, which may undercapture early or mild cases and introduce outcome misclassification. Third, our findings derive from routine care data and thus require replication across diverse populations and treatment settings to ensure generalizability. Finally, future work should integrate HbA1c trajectories with complementary biomarkers, including cognitive testing and neuroimaging, to refine risk stratification and elucidate underlying mechanisms.

In summary, this study is the first to demonstrate that dynamic HbA1c change after SGLT2i initiation provides a simple, feasible tool for predicting dementia risk, and it should be directly considered in clinical decision-making.

Supplemental Material