Ассоциация сахарного диабета с когнитивной функцией у взрослых американцев: перекрестный анализ мультицентральной группы АИ-РЕАДИ.

STRENGTHS AND LIMITATIONS OF THIS STUDY

This study uses data from a large, multicentre cohort with standardised cognitive testing protocols.

Diabetes status was defined using a structured, study-wide classification algorithm integrating laboratory values and medication history. Cognitive function was assessed using the electronically administered Montreal Cognitive Assessment, which provides standardised scoring and time-based performance metrics.

The cross-sectional design limits the ability to infer temporality or causality between glycaemic status and cognitive function.

Introduction

Type 2 diabetes mellitus (T2DM) is increasingly recognised both as a metabolic disorder and also as a major risk factor of cognitive decline, making it a major public health challenge in the USA. In 2022, approximately 38.4 million Americans were living with T2DM at an estimated $412.9 billion in economic burden. 1 2 Previous epidemiological studies indicate that individuals with diabetes have a higher risk of cognitive impairment and dementia compared with non-diabetic individuals. A meta-analysis of longitudinal studies found that diabetes is associated with a 50% higher risk of developing Alzheimer’s disease, a 21% higher risk of mild cognitive impairment and a 51% higher risk of any dementia compared with normoglycaemic controls. 3 4 Similarly, a comprehensive review reported that diabetes confers 1.25- to 1.90-fold higher risk for cognitive impairment or dementia and suggested that even pre-diabetes and subclinical hyperglycaemia are linked to elevated risk of cognitive decline. 5

Multiple cognitive domains appear to be influenced by diabetes, including memory, executive function, attention and processing speed. 6 For instance, people with diabetes frequently perform worse on tests of executive function and psychomotor speed, even after accounting for age, education and other confounding factors. 7 Visuospatial abilities and orientation may also be affected. However, not all studies have observed uniform effects. 8 9 Some studies reported only subtle cognitive differences or no significant impairment associated with pre-diabetes. 10 This variability suggests that the impact of diabetes on cognition may depend on context, including the severity of hyperglycaemia, glycaemic control, disease duration and coexisting risk factors.

Thus, ongoing research is focused on clarifying the relationship between the degree of hyperglycaemia and cognitive function. 11 12 Previous studies have shown evidence of a dose-dependent association of higher haemoglobin A1C (HbA1c) and glucose levels with worse cognitive performance, even in individuals without overt diabetes. 13 14 In one study, older diabetic patients with inadequately controlled HbA1c ( > 7) experienced significantly higher incidence rates of cognitive impairment over 5 years. 15 This is in contrast to patients with well-controlled diabetes who had cognitive impairment rates closer to normoglycaemic individuals. Furthermore, longer duration of diabetes has also been associated with greater cognitive deficits, implying glycaemic control and burden as a key modifiable factor in diabetes-related neurodegeneration. 16

Given the links of glycaemic control, diabetes severity, duration of diabetes and metabolic dysregulation, the connection between diabetes and cognition appears heavily context dependent. The Montreal Cognitive Assessment (MoCA) is a widely used screening test used both in clinical and research settings to assess global cognitive function and early cognitive impairment, 17 with reported sensitivity of 90% and specificity of 87% for detecting mild cognitive impairment. 18 More importantly, the executive function and visuospatial components of MoCA make it more sensitive than the Mini-Mental State Examination (MMSE). 19

Despite substantial interest in diabetes-related cognitive decline, important gaps remain. First, few studies have compared cognitive function across clearly defined, clinically relevant diabetes severity groups, such as normoglycaemia, pre-diabetes, medication-controlled T2DM and insulin-dependent T2DM, within the same analytic framework. 20 Second, prior work has largely focused on global cognition or single domains, with limited assessment of domain level MoCA performance across the diabetes spectrum. 21 Third, findings for pre-diabetes remain inconsistent, and many earlier studies used smaller or less diverse samples that may limit generalisability. 22 Finally, few investigations have simultaneously adjusted for a comprehensive set of metabolic, vascular and neurodegenerative comorbidities that may confound the relationship. 23

Building on prior research linking diabetes and glycaemic control to cognitive function, this study compares cognitive performance, measured by the MoCA scores, across four glycaemic status groups: individuals without diabetes, pre-diabetes, medication-controlled T2DM and insulin-dependent T2DM. Data are drawn from the Artificial Intelligence–Ready and Equitable Atlas for Diabetes Insights (AI-READI) study, a large, National Institutes of Health (NIH)-supported, multi-centre study designed to collect and share a multimodal AI-ready data from 4000 US adults aged 40 or older to enable machine learning research on diabetes and its complications. 24 We hypothesise that worsening diabetes severity is associated with declining cognitive function, quantified by lower MoCA total and domain scores, even after accounting for confounding factors. By analysing the AI-READI data set, we seek to advance our understanding of the potential dose-dependent relationship between cognitive impairment and glycaemic status to inform early cognitive function screening and intervention strategies.

Research design and methods

Study setting and design

This study uses data from the AI-READI study, 25 a cross-sectional, NIH-funded, multicentre study designed to support diabetes research through the collection of multimodal health data. 26 Participants were recruited from three US sites (Birmingham, AL; San Diego, CA and Seattle, WA), with sampling stratified by diabetes status, sex and race/ethnicity. By study design, sex and race/ethnicity distributions were balanced across groups, though individual-level data on these variables were not available. Individual-level data on sex, race and ethnicity are only accessible through a separate-controlled access process to reduce reidentification risk and the potential for stigmatising secondary analyses. Enrolment began on 18 July 2023 and will continue until 30 November 2026 to enrol 4000 participants. All participants provided informed consent and institutional review board approval was obtained at each site. 26 Site identifiers are not included in the publicly released AI-READI data set. Therefore, we were unable to examine site-specific differences or adjust for study site in our analyses.

For this project, we downloaded the publicly available AI-READI data set from the AI-READI website on 30 May 2025. We analysed the data set including demographics, cognitive testing, clinical measurements and laboratory measures from a cohort of over 1000 US adults aged 40 and older who were already enrolled with data publicly available. Patients or members of the public were not involved in the design, conduct, reporting or dissemination of this research, as the study consisted of secondary analyses of de-identified data.

AI-READI uses a purposive, triple-balanced sampling design rather than probability-based sampling. Therefore, no sampling weights were available or applied in our analysis. Furthermore, because this analysis is based on the publicly released AI-READI data set, the sample size (1067 participants) was determined entirely by the number of participants available in this release, and no a priori power calculation was performed.

Definition of glycaemic status groups

Each participant’s diabetes status was determined by the AI-READI study team as a four-level categorical variable in the data set. Although the precise algorithm used for classification was not disclosed in the public release, group assignments reflect a combination of laboratory values (eg, HbA1c, fasting glucose), medication history and self-reported diagnosis, which align with standard American Diabetes Association definitions. Any residual misclassification is expected to be nondifferential with respect to MoCA performance and would likely bias results towards the null rather than generate spurious associations. 27 AI-READI participants were classified into four groups including: (1) control group without diabetes: individuals without diagnosis of diabetes and with laboratory values below diagnostic thresholds (eg, fasting plasma glucose<100 mg/dL and HbA1c<5.7%); (2) pre-diabetes group: individuals with impaired glycaemic control that did not meet criteria for diabetes, such as fasting glucose between 100 and 125 mg/dL or HbA1c between approximately 5.7%–6.4%, without the use of glucose-lowering medications; (3) medication-controlled T2DM group: individuals diagnosed with T2DM who were managed with oral antihyperglycaemics or non-insulin injectables, but not insulin and (4) insulin-dependent T2DM group: individuals using insulin therapy for glycaemic control, either alone or in combination with other medications.

Cognitive assessment

Cognitive function was assessed using MoCA, a 30-point screening tool for mild cognitive impairment. The MoCA was administered to participants as part of the AI-READI protocol using a standardised electronic format (the MoCA Test Duo App on a tablet device). Trained clinical research coordinators supervised the administration. The MoCA evaluates multiple cognitive domains, including short-term memory recall, visuospatial abilities (eg, clock drawing), executive function (trail-making and abstraction tests), attention (digit spans, serial subtraction), language (naming, fluency), and orientation to time and place.

From the AI-READI data set, which follows the OMOP Common Data Model, we extracted both total and domain-specific MoCA scores for each participant. The MoCA total score reflects the sum of all item scores and ranges from 0 to 30, with higher scores indicating better global cognitive function. In accordance with MoCA scoring guidelines, one point was added to the total score for individuals with 12 or fewer years of formal education to adjust for education effects on test performance. Scores>27 typically indicate normal cognition, 18–25 suggest mild cognitive impairment, 10–17 moderate impairment and <10 severe impairment. 28 We also extracted subdomain scores, including the Abstraction score (0–2), reflecting performance on the conceptual similarities task that assesses semantic reasoning and executive function, and the Abstraction Time, which captures the time in seconds required to complete this task. The Orientation score (0–6) evaluates awareness of time and place, a domain that often remains preserved until more significant cognitive decline occurs, while Orientation Time reflects the time taken to complete this portion. Finally, we included the Memory (Combined MIS) score (0–15), which reflects the total number of words recalled on the delayed recall test, MoCA’s five-word recall, including points for category cuing or multiple-choice recognition. The MoCA memory score provides a sensitive indicator of memory and retrieval impairment.

Statistical analysis

We first performed descriptive analyses of participant characteristics overall and by four glycaemic status groups. Continuous measures were reported as means and SD, and categorical measures were summarised as counts and percentages. Group differences in baseline characteristics were assessed using one-way analysis of variance (ANOVA) for continuous measures (eg, age) and χ 2 tests for categorical measures (eg, sex, comorbidities).

The primary analysis compared cognitive outcomes across the four glycaemic groups (controls without diabetes, pre-diabetes, medication-controlled T2DM or insulin-dependent T2DM). We first examined MoCA total and subdomain scores using one-way ANOVA to test for overall group differences. Because MoCA scores demonstrated the expected mild ceiling effect and were not perfectly normally distributed, we evaluated distributional characteristics visually. However, the large sample size makes ANOVA robust to moderate deviations from normality, and no extreme skewness or outliers were observed that would invalidate parametric testing.

We then conducted multivariable linear regression models for the same continuous MoCA outcomes, adjusting for age, education, Body Mass Index (BMI), HbA1c, hypertension, high cholesterol, kidney disease, circulatory problems and neurodegenerative conditions, including Alzheimer’s disease, Parkinson’s disease, multiple sclerosis or self-reported mild cognitive impairment.

In addition to analysing MoCA as a continuous measure, we evaluated cognitive impairment defined as MoCA<26. Univariable logistic regression was first used to compare the odds of cognitive impairment across the four glycaemic groups. We then applied the same set of covariates in multivariable logistic regression models to assess adjusted associations with cognitive impairment. We did not conduct post-hoc pairwise comparisons, as our interest was in overall trends across glycaemic groups.

All analyses were performed on participants with complete data for the variables of interest; cases with missing data were excluded from the relevant analyses. All statistical tests were conducted using Python and Microsoft Excel. Two-tailed p value<0.05 was considered statistically significant.

Patient and public involvement

Patients and members of the public were not involved in the design, conduct, reporting or dissemination plans of this research.

Results

Participant characteristics

The AI-READI data included 1067 participants in the full analytic sample consisting of 372 normoglycaemic controls, 242 individuals with pre-diabetes, 323 with medication-controlled T2DM and 130 with insulin-dependent T2DM. Of the 1067 participants in the analytic sample, 1060 individuals (99.3%) completed at least one domain of MoCA. Completion rates were consistently high across glycaemic status groups: 371 of 372 (99.7%) controls, 239 of 242 (98.8%) individuals with pre-diabetes, 321 of 323 (99.7%) with medication-controlled T2DM and 129 of 130 (98.5%) with insulin-dependent T2DM contributed at least one MoCA domain score. A total of 1032 participants (96.7%) contributed all six MoCA domain scores. By glycaemic group, full-domain completion was observed in 364 participants without diabetes, 235 individuals with pre-diabetes, 312 with medication-controlled T2DM and 121 with insulin-dependent T2DM. Only seven participants (0.7%) lacked all MoCA data, and 28 participants (2.6%) had incomplete data, most commonly missing one or both of the time-based measures (Abstraction Time and Orientation Time).

The characteristics of participants overall and in each glycaemic group are reported in table 1 . The mean (SD) age of the overall sample was about 60 (11.1) years and did not differ significantly across the four groups (mean age range 59.6–61.4 years, p=0.15; table 1 ). Education level, however, differed by group (p<0.001) with the control group being the most educated (46% graduate level, 10% high school or below), whereas the insulin-treated diabetes group had the lowest educational attainment (23% graduate level, and 23% with high school or below). Expectedly, comorbidities were common in diabetes groups (eg, hypertension, dyslipidaemia, kidney disease; all p<0.001). BMI increased progressively from controls (28.5) to insulin-treated (34.4). There were no significant group differences in history of circulation problems (p=0.43) or diagnosed neurodegenerative conditions (p=0.15; table 1 ). Specifically, the percentage with neurodegenerative disease was 2.4% in the control group, 3.7% in the pre-diabetes group, 5.3% in the medication-controlled group and 6.2% in the insulin-dependent group. Mean HbA1c levels increased stepwise from 5.5% (controls) to 7.1% (insulin-dependent T2DM), p<0.001, confirming that glycaemic exposure increased across the spectrum of the defined categories. C-reactive protein, a marker of systemic inflammation, also differed significantly among glycaemic groups (p<0.001), ranging from a mean 2.8 mg/L in the control group to 6.8 mg/L in the insulin-dependent group.

Table 1 Demographic and clinical characteristics of the participants with diabetes and controls Characteristics All participants (n=1067) Control (n=372) Pre-diabetes (n=242) Medication controlled (n=323) Insulin dependent (n=130) P value * Age (years): mean (SD) 60.32 (11.14) 59.56 (11.28) 60.42 (11.04) 61.40 (10.93) 59.62 (11.35) 0.15 Education level, n ( % ) <0.001  High school or below ( < 12) 139 (13.03) 39 (10.48) 21 (8.68) 49 (15.17) 30 (23.08)  College (13–16) 504 (47.24) 162 (43.55) 111 (45.87) 162 (50.15) 69 (53.08)  Graduate school (17–25) 411 (38.52) 171 (45.97) 110 (45.45) 111 (34.37) 30 (23.08)  Unknown or uncategorised (missing) 14 (1.31) 0 0 2 (0.62) 1 (0.77) BMI (kg/m 2 ): mean (SD) 30.7 (8.10) 28.54 (7.02) 30.24 (7.56) 32.05 (8.53) 34.39 (8.94) <0.001 HbA1c%, mean (SD) 6.08 (1.25) 5.51 (0.74) 5.80 (0.82) 6.53 (1.34) 7.13 (1.71) <0.001 CRP (mg/L): mean (SD) 3.77 (7.30) 2.83 (5.15) 3.33 (5.02) 4.01 (8.62) 6.80 (11.10) <0.001 Hypertension, yes ( % ) 571 (53.51) 124 (33.33) 121 (50) 224 (69.35) 102 (78.46) <0.001 Kidney problems, yes ( % ) 118 (11.06) 11 (2.96) 31 (12.81) 40 (12.38) 36 (27.69) <0.001 High blood cholesterol, yes ( % ) 544 (50.98) 147 (39.52) 108 (44.63) 209 (64.71) 80 (61.54) <0.001 Circulation problem, yes ( % ) 84 (7.87) 28 (7.53) 14 (5.79) 30 (9.29) 12 (9.23) 0.43 Neurodegenerative disease, yes ( % ) 43 (4.03) 9 (2.42) 9 (3.72) 17 (5.26) 8 (6.15) 0.15  Alzheimer’s disease 1 (0.09) 0 0 1 (0.31) 0  Mild cognitive impairment 30 (2.81) 6 (1.61) 8 (3.31) 10 (3.10) 6 (4.62)  Parkinson’s disease 7 (0.66) 3 (0.81) 0 3 (0.93) 1 (0.77)  Multiple sclerosis 6 (0.56) 0 1 (0.41) 3 (0.93) 2 (1.54) * P values were calculated using one-way ANOVA for continuous variables and χ2 test for categorical variables. ANOVA, analysis of variance; BMI, Body Mass Index; CRP, C-reactive protein; HbA1c, haemoglobin A1C.

Demographic and clinical characteristics of the participants with diabetes and controls

P values were calculated using one-way ANOVA for continuous variables and χ2 test for categorical variables.

ANOVA, analysis of variance; BMI, Body Mass Index; CRP, C-reactive protein; HbA1c, haemoglobin A1C.

Global cognitive performance (MoCA total score)

We observed a statistically significant difference in MoCA total scores across four glycaemic groups (p<0.001, table 2 , figure 1 ), with highest mean MoCA total score in the control group (mean=25.8) and lowest in the insulin-dependent T2DM group (mean=24.0), with intermediate values for pre-diabetes (mean=25.5) and medication-treated diabetes (mean=24.9). As shown in table 2 and figure 1 , these differences reflect a stepwise decline in global cognitive function with increasing diabetes severity. The significant difference remained in a multivariable analysis with adjustment by age, education, BMI, HbA1c and vascular comorbidities, and neurodegenerative diseases (adjusted p=0.02; table 2 ), with adjusted means of 25.8 (control group), 25.4 (pre-diabetes group), 25.3 (medication treated group), and 24.5 (insulin-dependent group).

-Distribution of MoCA total scores by glycaemic group, illustrating a stepwise decline in cognitive performance from individuals with no diabetes to those with insulin-dependent diabetes. MoCA, Montreal Cognitive Assessment.

Table 2 Univariate and multivariate analysis for comparison of Montreal Cognitive Assessment (MoCA) scores across glycaemic status groups Domain * Analysis † Control (n=371) Pre-diabetes (n=239) Medication controlled (n=321) Insulin dependent (n=129) P value Abstraction (0–2) Mean (SE) 1.88 (0.02) 1.88 (0.03) 1.86 (0.02) 1.73 (0.05) 0.004 Adjusted mean (SE) 1.88 (0.03) 1.87 (0.03) 1.89 (0.04) 1.76 (0.05) 0.03 Abstraction Time (s) Mean (SE) ‡ 43.14 (1.76) 42.86 (1.91) 43.91 (1.60) 50.51 (4.15) 0.15 Adjusted mean (SE) § 44.28 (2.43) 44.08 (2.80) 41.79 (2.90) 46.64 (4.02) 0.55 Combined MIS Score (0–15) Mean (SE) 12.63 (0.14) 12.25 (0.21) 11.93 (0.17) 11.64 (0.28) 0.003 Adjusted mean (SE) 12.72 (0.22) 12.44 (0.25) 12.41 (0.26) 12.23 (0.36) 0.51 Orientation (0–6) Mean (SE) 5.93 (0.02) 5.95 (0.02) 5.91 (0.02) 5.82 (0.04) 0.004 Adjusted mean (SE) 5.92 (0.02) 5.94 (0.03) 5.93 (0.03) 5.86 (0.04) 0.13 Orientation Time (s) Mean (SE) ‡ 28.52 (0.55) 31.12 (0.97) 30.85 (0.74) 30.63 (1.11) 0.04 Adjusted mean (SE) § 28.35 (0.92) 30.57 (1.07) 29.24 (1.11) 28.44 (1.53) 0.18 Total score (0–30) Mean (SE) 25.83 (0.16) 25.45 (0.22) 24.85 (0.18) 24.03 (0.37) <0.001 Adjusted mean (SE) 25.76 (0.23) 25.44 (0.27) 25.32 (0.28) 24.51 (0.38) 0.02 * Score range is shown in parenthesis, and higher scores indicate better cognitive function, except for time-based measures (Abstraction Time, Orientation Time), where shorter time indicates better cognitive function. † Adjusted means account for age, education, BMI, HbA1c, hypertension, kidney problems, high blood cholesterol, circulation problems and neurodegenerative diseases. Subjects (n=41) with missing data in any of these covariates were excluded from multivariable analysis. ‡ The number of subjects with data for analysis are 364 in control group, 235 in pre-diabetes group, 312 in medication-controlled group and 121 in insulin-dependent group. § The number of subjects with data for analysis are 351 in control group, 227 in pre-diabetes group, 298 in medication-controlled group and 115 in insulin-dependent group. BMI, Body Mass Index; HbA1c, haemoglobin A1C; MIS, Memory Index Score.

Univariate and multivariate analysis for comparison of Montreal Cognitive Assessment (MoCA) scores across glycaemic status groups

Score range is shown in parenthesis, and higher scores indicate better cognitive function, except for time-based measures (Abstraction Time, Orientation Time), where shorter time indicates better cognitive function.

Adjusted means account for age, education, BMI, HbA1c, hypertension, kidney problems, high blood cholesterol, circulation problems and neurodegenerative diseases. Subjects (n=41) with missing data in any of these covariates were excluded from multivariable analysis.

The number of subjects with data for analysis are 364 in control group, 235 in pre-diabetes group, 312 in medication-controlled group and 121 in insulin-dependent group.

The number of subjects with data for analysis are 351 in control group, 227 in pre-diabetes group, 298 in medication-controlled group and 115 in insulin-dependent group.

BMI, Body Mass Index; HbA1c, haemoglobin A1C; MIS, Memory Index Score.

MoCA cognitive domain scores

In univariable analysis, several specific cognitive domains showed significant differences, including the Memory Index Score (MIS), with mean MIS score of 12.6 (out of 15) in the control group and 11.6 in the insulin-dependent group (p=0.003; table 2 ). However, this difference was no longer significant after adjustment for covariates (p=0.51; table 2 ). In univariate analysis, Abstraction scores differed significantly by glycaemic status (p=0.004; table 2 ) with the lowest mean score (1.73) in the insulin-dependent group, while the other groups scored similarly (1.86–1.88). The group difference in abstraction score remained statistically significant in multivariable analysis (adjusted p=0.03; table 2 ). In univariate analysis, Orientation scores showed a significant difference across groups (p=0.004), with a slightly lower mean in the insulin-dependent group (5.82) compared with other groups (~5.91–5.95). However, this difference was no longer statistically significant after adjustment (p=0.13, table 2 ).

MoCA task completion times

We also examined the time-based measures recorded during MoCA administration. Abstraction task time was longest in the insulin-dependent group (51 s) and similar in other groups (43 s), though the group difference was not statistically significant in unadjusted analysis (p=0.15). In multivariate analysis, group differences in Abstraction Time remained non-significant (p=0.55), with adjusted means ranging from 41.8 to 46.6 s. The time on orientation response differed significantly across groups (p=0.04), with the pre-diabetes group taking the longest time (~31.1 s) and the shortest time in the control group (28.5 s). However, this difference was not statistically significant in multivariate analysis (p=0.18, table 2 ).

Cognitive impairment prevalence and severity

Figure 2 illustrates the prevalence and severity distribution of cognitive impairment across glycaemic groups. Presence and severity of cognitive impairment increased with severity of diabetes (p<0.001, figure 2 ). Cognitive impairment was present in 39% of controls, rising to 44% in pre-diabetes, 54% in the medication-controlled group and 58% in the insulin-dependent group. Conversely, mild cognitive impairment (MoCA total score 18–25) was found for 37% in controls, rising to 42% in pre-diabetes and 52% in the medication-treated group and 50% in the insulin-dependent group. Moderate cognitive impairment (MoCA total score 10–17) was highest in the insulin-dependent group (7%), and approximately 2% in other groups. Only one individual (0.8%) in the insulin-dependent group had severe impairment (MoCA score<10), and there were none in the other groups.

Stacked bar plot for the presence and severity of cognitive impairment by glycaemic group.

Association between diabetes status and cognitive impairment

In the univariate analysis, both the medication-treated group and insulin-dependent group showed a significantly elevated risk of cognitive impairment compared with the control group ( table 3 ). Participants with medication-controlled T2DM had an unadjusted OR of 1.83 (95% CI: 1.36 to 2.48, p<0.001) for being cognitively impaired compared with the control group. Those with insulin-dependent T2DM had an even higher unadjusted OR of 2.14 (95% CI: 1.42 to 3.22, p<0.001). In contrast, pre-diabetes was not associated with significantly higher risk of cognitive impairment than health controls (OR 1.21, 95% CI: 0.87 to 1.68, p=0.26).

Table 3 Univariable and multivariable analysis for association between cognitive impairment and glycaemic status Univariate analysis Multivariate analysis * n Cognitive impairment n (%) OR (95% CI) P value n Cognitive impairment n (%) Adjusted OR (95% CI) P value Control 371 146 (39.35) Reference 358 137 (38.27) Reference Pre-diabetes 239 105 (43.93) 1.21 (0.87 to 1.68) 0.26 231 100 (43.29) 1.20 (0.85 to 1.71) 0.30 Medication controlled 321 174 (54.21) 1.83 (1.36 to 2.48) <0.001 307 163 (53.09) 1.50 (1.05 to 2.15) 0.03 Insulin dependent 129 75 (58.14) 2.14 (1.42 to 3.22) <0.001 123 73 (59.35) 1.85 (1.12 to 3.07) 0.02 * Adjusted for age, education, BMI, HbA1c, hypertension, kidney problems, high blood cholesterol, circulation problems and neurodegenerative diseases (including Alzheimer’s disease, mild cognitive impairment, multiple sclerosis and Parkinson’s disease). Multivariable model included participants with complete data on these covariates (n=1019), participants with missing data (n=41) in any of these covariates were excluded from the multivariable analysis. BMI, Body Mass Index; HbA1c, haemoglobin A1C.

Univariable and multivariable analysis for association between cognitive impairment and glycaemic status

Adjusted for age, education, BMI, HbA1c, hypertension, kidney problems, high blood cholesterol, circulation problems and neurodegenerative diseases (including Alzheimer’s disease, mild cognitive impairment, multiple sclerosis and Parkinson’s disease). Multivariable model included participants with complete data on these covariates (n=1019), participants with missing data (n=41) in any of these covariates were excluded from the multivariable analysis.

BMI, Body Mass Index; HbA1c, haemoglobin A1C.

In multivariable analysis with adjustment for age, sex, education, HbA1c, BMI, hypertension, high cholesterol, kidney disease, circulatory disease and neurodegenerative conditions, the significant associations remained for the T2DM groups, with an adjusted OR of 1.50 (95% CI: 1.05 to 2.15, p=0.03) for medication-controlled diabetes and 1.85 (95% CI: 1.12 to 3.07, p=0.02) for insulin-dependent diabetes as compared with controls. The pre-diabetes group was not associated with a significantly higher risk of cognitive impairment (adjusted OR 1.20, 95% CI: 0.85 to 1.71, p=0.30). Older age was significantly associated with higher odds of cognitive impairment (OR=1.03 per year increase; 95% CI: 1.02 to 1.04, p<0.001). Educational attainment was also strongly associated with cognitive status. Compared with participants with a college education, those with graduate-level education had lower risk of cognitive impairment (OR=0.66; 95% CI: 0.50 to 0.88, p=0.004), whereas those with a high school education or less had substantially higher risk of cognitive impairment (OR=1.90; 95% CI: 1.26 to 2.88, p=0.002). BMI, HbA1c, neurodegenerative conditions and vascular comorbidities were not independently associated with cognitive impairment in the multivariable model (all p≥0.16).

Discussion

In this cross-sectional study of the AI-READI participants, we compared cognitive function across four glycaemic status groups and found that worse diabetes severity was associated with poorer cognitive function. Individuals with insulin-dependent T2DM had the lowest MoCA total scores, and these differences remained significant even after adjusting for confounders including age, education and neurodegenerative conditions. These results are consistent with a growing body of evidence that T2DM is a risk factor for mild cognitive impairment and various forms of dementia. The cognitive differences observed align with prior population-based studies that show a dose-dependent relationship between diabetes and cognitive decline. Different from most of the previous studies, our study compared specifically the early mild cognitive dysfunction assessed by the MoCA, a screening tool with reported high sensitivity (90%) and specificity (87%) for mild cognitive impairment. By using well-validated MoCA measures for cognitive function in this large sample of participants, our study extends these observations and provides insight into how glycaemic control is related to cognitive health.

Although several MoCA subdomains differed by glycaemic group in unadjusted analyses, most of these differences were attenuated and no longer statistically significant in multivariable analysis. One likely explanation for the loss of statistical significance in several cognitive domains is that many of the covariates included in the adjusted models, such as age, educational attainment, BMI and vascular comorbidities, are themselves strong determinants of cognitive performance. Because these factors are also unequally distributed across glycaemic groups, they may serve as both confounders and potential mediators in the relationship between diabetes and cognition. For example, lower educational attainment was more prevalent in the diabetes groups and is independently associated with reduced cognitive function. Similarly, vascular risk factors like hypertension and circulation problems, each of which can impair cerebral perfusion or promote microvascular damage, were more prevalent in individuals with advanced diabetes. Neurodegenerative disease diagnoses, although less frequent, also influence cognitive scores and may further dilute glycaemic group effects. Taken together, our findings therefore suggest that diabetes may act in part through these shared pathways to influence cognition, rather than acting entirely independently.

Despite this attenuation, total MoCA score and the abstraction subdomain remained significantly different across groups in multivariable analysis. These outcomes likely reflect broader and more persistent cognitive vulnerabilities. Total score captures cumulative deficits across domains, while abstraction reflects higher order frontal lobe functioning and may be an early marker of executive dysfunction. Their persistence after adjustment by confounders suggests these aspects of cognition may be particularly sensitive to long-term diabetes-related changes and less easily explained by demographic or vascular confounders. However, it is important to note that although these differences reached statistical significance, the absolute differences in scores were small and may not be clinically meaningful. Nonetheless, even small decrements in MoCA scores may represent early signals of cognitive vulnerability that accumulate over time. In large populations, such subtle deficits could translate into earlier onset of functional limitations or greater risk of progression to dementia.

When we examined cognitive impairment as a binary outcome, however, the relationship with diabetes severity became more pronounced. In multivariable analyses, individuals with insulin-dependent T2DM had 1.85-fold increased risk of cognitive impairment (95% CI: 1.11 to 3.07), and those with medication-controlled T2DM had 1.50-fold increased risk of cognitive impairment (95% CI: 1.05 to 2.15), compared with the control group. These findings persisted even after adjusting for key demographic and confounding factors, underscoring the potential cumulative burden of diabetes on cognitive health. This suggests that while domain-specific score differences may be subtle, the overall likelihood of cognitive impairment is meaningfully elevated with greater diabetes severity.

In our control group of non-diabetes participants with mean age ~60 years old, we found 39% participants had mild cognitive impairment, which is consistent with prior population-based studies in Norway (mean age of 64 years) and China (mean age of 62 years) that reported cognitive impairment prevalence rates of 49% and 38%, respectively, using the same cut-off of MoCA<26. 29 30 This supports the validity of our findings in the cognitive function measure across diabetes groups.

These results both reinforce the clinical relevance of our findings and also align with prior longitudinal studies linking diabetes to cognitive decline. For example, the Atherosclerosis Risk in Communities (ARIC) study found that poor glycaemic control (HbA1c≥7%) and longer diabetes duration were associated with greater cognitive decline over ~5 years. 15 We likewise observe that those presumably with longer standing or harder-to-control diabetes (the insulin-requiring group, mean HbA1c~7.1% despite treatment) have the worst cognitive outcomes. These results support the idea that it may not be diabetes itself, but rather poor glycaemic control and disease duration that elevate cognitive risk. In our cross-sectional data, the gradient of cognitive impairment (pre-diabetes<non-insulin T2DM<insulin-dependent T2DM) reinforces this conclusion. The persistence of cognitive differences even after adjusting for HbA1c suggests that cumulative exposure and treatment intensity (as reflected by insulin use) may contribute beyond a single HbA1c measurement. Additionally, the significant effect of age and education on cognitive impairment in our multivariable models supports the internal validity of our findings, indicating that the model captures established associations in cognitive ageing.

It is important to note that MoCA allowed us to efficiently capture these multi-domain cognitive deficits. The MoCA has greater sensitivity than the MMSE for detecting mild cognitive impairment. In fact, previous studies in diabetic populations have found MoCA to outperform the MMSE in detecting cognitive dysfunction. The use of tablet-based digital administration is also noteworthy in our study, as it ensured standardised timing and scoring across participants. Furthermore, our data extraction captured response times, providing valuable quantitative information. Although the relevance of a few-second difference in task completion times remains uncertain, these objective metrics represent a step towards more precise cognitive measures for use in large studies.

There are several plausible biological mechanisms linking diabetes to cognitive impairment. Chronic hyperglycaemia can damage small vessels and neurons through oxidative stress, advanced glycation end-products formation and other pathways, potentially leading to cerebral microvascular disease. 31 32 Insulin resistance may also impair brain insulin signalling and promote amyloid accumulation, contributing to neurodegeneration. 33 34 These mechanisms align with our findings of lower cognitive scores among diabetic participants, particularly those on insulin who also had more comorbid vascular risk factors. Together, these mechanisms support a multifactorial model in which poor glycaemic control, vascular disease and metabolic dysfunction converge to impact cognition. 16

This study has several limitations that must be acknowledged. First, due to its cross-sectional design, we cannot establish the temporal direction of the association between glycaemic status and cognitive function. It is possible that individuals with cognitive impairment have more difficulty managing their diabetes, leading to worse glycaemic control, rather than hyperglycaemia causing cognitive decline. Second, while we adjusted for education level and used the MoCA education correction, residual confounding may persist, particularly given the higher educational level in our control group. Third, we were unable to adjust for all potential comorbidities, medications and treatments that may be associated with cognitive function. While our models accounted for key vascular risk factors such as hypertension and the presence of neurodegenerative disease diagnoses, data were unavailable for other important risk factors, including sex, ethnicity, lifestyle factors, history of stroke and APOE-ε4 genotype. Fourth, we could not account for potential site-level differences because study-site identifiers are not available in the public AI-READI data set. Finally, our cohort may not fully represent all populations with or without diabetes. Higher educational attainment and healthcare access in our sample could limit applicability to lower resource settings.

Conclusions

Our findings suggest that more severe forms of T2DM, particularly insulin-treated T2DM, are associated with modest but measurable impairments in global and executive cognitive function. While many domain-specific differences did not remain statistically significant after adjustment, the elevated risk of overall cognitive impairment observed in insulin-dependent individuals reinforces the broader cognitive burden associated with advanced diabetes. Routine cognitive screening, particularly in patients with advanced or poorly controlled diabetes, may support timely interventions and promote treatment adherence. This is becoming more relevant as treatments for Alzheimer’s disease emerge. As T2DM prevalence continues to rise globally, integrating cognitive evaluation into standard diabetes care could play an important role in preserving cognitive function and quality of life.