The UK Biobank study is an ongoing population-based prospective cohort that enables comprehensive investigation of multifactorial disease etiology, incorporating genomic, environmental, and behavioral determinants (14). Between 2006 and 2010, approximately 500,000 participants were recruited across the United Kingdom. Comprehensive baseline data were collected, including questionnaire information, physical measurements, biological samples (blood, urine, and saliva), and imaging assessments (e.g., brain MRI scans). In addition, the UKB has established long-term follow-up through linkage with National Health Service health records, including hospital admissions, outpatient visits, cancer registries, and death registries (15). The Northwest Multi-centre Research Ethics Committee approved the UK Biobank study protocol, and written informed consent was obtained from all participants at baseline. The application number for the present study is 470,494.

In this study, we initially screened a total of 501,988 participants from the UKB. We first excluded 123 individuals who had been diagnosed with dementia at baseline. Subsequently, 74,967 participants were excluded due to missing data on the modified TyG indices (including blood glucose, triglyceride (TG), BMI, and WC). An additional 70,444 individuals were excluded due to missing key covariate information on age, sex, apolipoprotein E (APOE) genotype, or ethnicity. Ultimately, 356,454 participants met the inclusion criteria and were included in the final analyses (Figure 1).

The modified TyG indices, including TyG-BMI and TyG-WC, are extensions of the traditional TyG index, incorporating BMI and WC into their respective calculations. The traditional TyG index is derived from fasting blood glucose (FBG) and TG levels measured at baseline. The formulas are as follows: TyG index = log [TG (mg/dL) × FBG (mg/dL)/2]; TyG-BMI index = TyG index × BMI; TyG-WC index = TyG index × WC.

The primary outcomes of this study are AD and VaD. The determination of dementia types is based on multiple data sources, including hospitalization and diagnostic records from England’s hospital incidence statistics, Scotland’s incidence records, and Wales’ patient database. Additionally, death registration data was collected from the National Health Service digital records in England and Wales, as well as from the Information Services Division of Scotland, to identify more cases. The classification of dementia types was based on both the International Classification of Diseases Ninth Revision (ICD-9) and Tenth Revision (ICD-10) codes. AD was defined by ICD-10 codes F00 and G30, and ICD-9 code 331.0. VaD was defined by ICD-10 codes F01 and I67.3, and ICD-9 code 290.4.

The covariates included in this study were: age, sex, ethnicity, APOE genotype, educational level, Townsend Deprivation Index (TDI), household income, smoking status, alcohol consumption, dietary habits, sleep patterns, hypertension, diabetes, dyslipidemia, history of stroke, history of atrial fibrillation (AF), and the use of glucose-lowering and lipid-lowering medications. For imaging outcomes, we also adjusted for total brain volume.

Race was dichotomized (White/non-White). Smoking and alcohol use were categorized as current, former, or never users (16). Diet quality was assessed using a dietary score based on seven components (vegetables, fruits, fish, whole grains, refined grains, processed meats, and unprocessed meats), with a total score ranging from 0 to 7; higher scores indicated greater adherence to healthy dietary guidelines (17, 18). Sleep patterns were evaluated using a composite sleep score derived from five factors: chronotype, sleep duration, insomnia, snoring, and daytime sleepiness. The total score ranged from 0 to 5 and was used to categorize sleep patterns as healthy (4–5), intermediate (2–3), or poor (0–1) (19).

The MRI imaging protocol used in this study was designed by the UK Biobank Imaging Working Group, and the detailed image acquisition protocol and data processing methods have been thoroughly described in previously published literature (20). All neuroimaging data were standardized and acquired using the Siemens Skyra 3 T MRI scanner, and image segmentation and quantification were performed based on the automated processing workflow established by the UKB. This study focused on three key neuroimaging measures: total brain volume, white matter hyperintensity (WMH) volume, and hippocampal volume.

This study, based on data from the UKB, systematically evaluates the association between modified TyG indices and dementia risk, as well as its impact on brain structure. Descriptive statistics present categorical variables as numbers (%) and continuous variables as means (standard deviations (SD)). Between-group differences were evaluated using chi-square tests for proportions and ANOVA for continuous measures. Missing covariate data were addressed using multiple imputation by chained equations (MICE) to minimize potential bias.

To quantify the relationship between modified TyG indices and dementia risk, we constructed a series of Cox proportional hazards models with multiple levels of adjustment: Model 1 adjusted for age, sex, ethnicity, APOE genotype, educational level; Model 2 further adjusted for TDI, household income, smoking, alcohol consumption, diet, and sleep patterns; Model 3 additionally included hypertension, diabetes, dyslipidemia, history of stroke, history of atrial fibrillation (AF), and the use of glucose-lowering and lipid-lowering medications.

To reveal the nonlinear relationship between modified TyG indices and dementia risk, we used a restricted cubic spline (RCS) model (4 knots) to fit the dose–response curve, adjusting for the covariates in Model 3, and evaluated the statistical significance of the nonlinear trend using likelihood ratio tests. For the cross-sectional associations between modified TyG indices and brain structures, we used generalized linear regression to calculate the standardized β coefficient (reflecting the change in brain volume per 1 SD increase in modified TyG indices) and its 95% CI.

In the subgroup analysis, we stratified participants by age, sex, ethnicity, APOE ε4 carrier status, hypertension, hyperlipidemia, diabetes, and fasting duration, and evaluated the interactions between these subgroup variables and the modified TyG indices using likelihood ratio tests. Two sensitivity analyses were conducted to verify result robustness: (1) excluding those who developed dementia during the initial 2 years of follow-up to address potential reverse causality bias; (2) using the original TyG index to replace modified TyG indices and re-evaluating its association with dementia to assess whether the relationship between modified TyG indices and dementia is stable.

All analyses were conducted using R (v4.4.2) with the “mice,” “survival,” and “rms” packages. A two-sided p value < 0.05 was considered statistically significant. To account for multiple comparisons, false discovery rate (FDR)-adjusted p values were calculated using the Benjamini-Hochberg procedure.

Table 1 describes the baseline characteristics of 356,454 participants, including 352,474 non-dementia controls and 3,980 dementia cases (2,594 AD, 1,386 VaD). The overall population had a mean age of 56.50 ± 8.11 years, 54.33% female, 94.09% white, and a mean follow-up of 13.27 ± 2.05 years.

Compared to non-dementia controls, dementia groups were older, had fewer females, higher white ethnicity, and greater socioeconomic disadvantages (lower college education rates and higher deprivation scores). Genetically, dementia cases—particularly those with AD—had a markedly higher prevalence of APOE ε4 carriers. Unhealthy lifestyle factors were also more common among dementia cases, including lower healthy diet scores, higher rates of smoking and less favorable sleep patterns (p < 0.001). Medically, histories of hypertension, diabetes, and stroke were significantly more prevalent in both dementia subtypes (p < 0.001), while dyslipidemia and AF histories did not differ significantly across groups (p = 0.12 and 0.5, respectively). Use of glucose- and lipid-lowering medications was also higher among dementia groups.

In Cox proportional hazards models, both TyG-BMI and TyG-WC were independently associated with AD and VaD risk after full adjustment (Table 2). In the Model 3, a 1-unit increment in TyG-BMI was linked to 0.17% decrease in AD risk (HR = 0.9983, FDR-adjusted p = 0.001) and a 0.3% increase in VaD risk (HR = 1.003, FDR-adjusted p < 0.001). Similarly, each 1-unit increase in TyG-WC predicted a 0.04% reduction in AD risk (HR = 0.9996, FDR-adjusted p = 0.042) and a 0.12% increase in VaD risk (HR = 1.0012, FDR-adjusted p < 0.001). The sextile-based analysis revealed a clear dose–response relationship. In the fully adjusted model using Q5 as the reference, both TyG-BMI and TyG-WC exhibited a significant U-shaped association with the risk of AD. Participants in the lowest sextile (Q1) had a significantly increased risk of AD (TyG-BMI: HR = 1.47, FDR-adjusted p < 0.001; TyG-WC: HR = 1.23, FDR-adjusted p = 0.019), while those in the highest sextile (Q6) also showed a trend toward increased risk, although the associations were not statistically significant (TyG-BMI: HR = 1.1, FDR-adjusted p = 0.177; TyG-WC: HR = 1.07, FDR-adjusted p = 0.309). In contrast, the risk of VaD showed a monotonically increasing pattern, with participants in the highest sextile (Q6) having significantly elevated risks (TyG-BMI: HR = 1.32, FDR-adjusted p = 0.029; TyG-WC: HR = 1.45, FDR-adjusted p = 0.011) (Table 2).

The multivariable-adjusted RCS analysis revealed significant nonlinear dose–response relationships between TyG-BMI, TyG-WC, and the risks of AD and VaD (Figure 2). For AD, both indices displayed a broad U-shaped relationship. The nadir of risk occurred at TyG-BMI = 286.10, with the protective range spanning 232.62–388.18. TyG-WC showed protective effects within the range of 776.76 to 1114.72, with the lowest risk at 911.94 (P for nonlinear = 0.023). Risk gradually increased outside of these optimal ranges. In contrast, the association with VaD followed a J-shaped pattern: risk remained stable within the normal range, but increased significantly when TyG-BMI exceeded 235.05 (P for nonlinear < 0.001) or TyG-WC exceeded 868.93 (P for nonlinear < 0.001).

Based on the above results, we further investigated relationships between modified TyG indices and region-specific brain volumetric changes using structural MRI. After further adjusting for total brain volume on the basis of Model 3, multiple linear regression analysis revealed that both TyG-BMI and TyG-WC were independently associated with brain structural alterations. As shown in Table 3, higher levels of TyG-BMI and TyG-WC were significantly associated with increased hippocampal volume (TyG-BMI: β = 0.64, FDR-adjusted p < 0.001; TyG-WC: β = 0.23, FDR-adjusted p < 0.001), suggesting that elevated modified TyG indices may offer protective effects against hippocampal atrophy, a hallmark of AD.

Meanwhile, both TyG-BMI and TyG-WC were also significantly associated with greater WMH volume (TyG-BMI: β = 8.85, FDR-adjusted p < 0.001; TyG-WC: β = 3.22, FDR-adjusted p < 0.001), indicating that higher modified TyG indices are closely linked to increased small vessel disease burden, which might play a pivotal role in the development and progression of VaD.

Owing to the limited number of dementia subtype events, formal survival-based mediation analysis was not feasible in the present study. To further explore whether modified TyG indices may influence cognitive outcomes through brain structural alterations, disease-relevant brain structural measures were additionally included as covariates in the multivariable models. After further adjustment for hippocampal volume in the analyses of AD or WMH volume in the analyses of VaD, the associations between modified TyG indices and dementia risk were substantially attenuated and no longer statistically significant (Supplementary Table 1). These findings suggest that brain structural alterations may partly account for the association between modified TyG indices and dementia outcomes.

We performed stratified analyses to assess whether the associations between modified TyG indices and dementia risk varied across different subgroups (Supplementary Table 2). Significant interactions were found with age (TyG-BMI: p = 0.003; TyG-WC: p < 0.001) and APOE genotype (TyG-BMI: p = 0.003; TyG-WC: p < 0.001) for AD. In participants aged ≥65 years, both TyG-BMI and TyG-WC were significantly and inversely associated with AD risk, whereas no such associations were found in those under 65 years. Similarly, among APOE ε4 non-carriers, higher modified TyG indices were associated with lower AD risk, but this association was absent in ε4 carriers.

In addition, significant interactions were observed between TyG-BMI and hypertension (p = 0.019), TyG-BMI and hyperlipidemia (p = 0.019), and between TyG-WC and diabetes (p = 0.007), indicating stronger associations in individuals with these comorbidities.

For VaD, both TyG-BMI and TyG-WC showed positive associations with VaD risk, and interaction effects were observed with age (TyG-BMI: p < 0.035; TyG-WC: p < 0.001) and APOE genotype (TyG-BMI: p = 0.005; TyG-WC: p < 0.001). These associations were more prominent among younger individuals and APOE ε4 carriers.

No statistically significant interactions were detected for sex, ethnicity, or fasting duration in either AD or VaD models (all P for interaction > 0.05), suggesting that the observed associations were largely consistent across these subgroups.

To assess result robustness, we conducted several sensitivity analyses. First, we excluded participants who developed dementia within the first 2 years of baseline assessment, and the associations between TyG-BMI, TyG-WC, and dementia risk remained largely unchanged, further supporting the reliability of our results (Supplementary Table 3). Second, we re-evaluated the association between the original TyG index (without incorporating BMI or WC) and the risk of dementia (Supplementary Table 4). In the continuous variable model, each one-unit increase in TyG was associated with a 12% reduced risk of AD (HR = 0.88, 95% CI: 0.82–0.94, FDR-adjusted p = 0.001), but an 14% increased risk of VaD (HR = 1.14, 95% CI: 1.03–1.26, FDR-adjusted p = 0.029). RCS analysis revealed a U-shaped association between TyG and AD risk (P for nonlinear = 0.011) and a J-shaped association with VaD risk (P for nonlinear < 0.001) (see Supplementary Figure 1). In the sextile-based model, elevated AD risk was primarily observed in the lowest TyG sextile (Q1; HR = 1.3, FDR-adjusted p = 0.001), while the highest sextile (Q6) showed a trend toward increased risk, though not statistically significant (HR = 1.02, FDR-adjusted p = 0.772). For VaD, this association was attenuated and no longer significant in Model 3 (HR = 1.13, FDR-adjusted p = 0.406). Taken together, these findings suggest that while the original TyG index may offer some predictive value for AD and VaD, its associations are relatively unstable. In contrast, the composite indices incorporating BMI or WC demonstrated more robust and consistent associations across analyses, indicating superior predictive performance.