Association of pro-inflammatory diet with increased risk of all-cause dementia and Alzheimer's dementia: a prospective study of 166,377 UK Biobank participants

UK Biobank is a large cohort study containing approximately 500,000 participants recruited between 2006 and 2010, ranging in age from 37 to 73 years. Participants provided baseline data by completing touch screen questionnaires, interviews, physical and functional measures, and genetic and biological assessments at 22 study centers in England, Scotland, and Wales [17]. The UK Biobank is approved by the North West UK Multicentre Research Ethics Committee, the National Information Management Board, and the UK Government. Studies involving human participants were reviewed. Patients/participants provided their written informed consent.

In this study, we screened 502,387 participants recruited by UK Biobank. We began by excluding 291,430 participants missing a 24-h estimated nutrient intake. We then excluded 79 participants who already suffered from dementia at the time of enrollment. Ultimately, after excluding an additional 44,501 participants without complete baseline information, 166,377 participants were included in this study (Fig. 1).

We used data from an Internet-based 24-h diet questionnaire (Oxford WebQ) in the UK Biobank to assess diet. Specific details about the Oxford WebQ have been explained in detail elsewhere [18]. The questionnaire was initially introduced at the end of the UK Biobank recruitment process and assessed on the last 70,000 participants. In addition, at four different time periods between February 2011 and April 2012, the UK Biobank invited those recruits who had provided their email address to complete this online questionnaire via email [19]. Oxford WebQ was able to automatically generate estimated energy and nutrient values for participant-reported foods by surveying participants on their consumption of 200 common foods and beverages in the previous 24 h. This prospective study obtained baseline nutrient intakes by averaging the estimated energy and nutrient value data reported by questionnaires for participants at different time periods during 2011–2012. The average nutrients obtained were then used in the calculation of the dietary inflammation index. A validation study based on blood and urine biomarkers has supported the reliability of Oxford WebQ [20].

The dietary inflammation index (DII) is a literature- and population-based tool developed with the capability to compare the dietary inflammatory potential in different populations [9]. The DII has been applied in a large number of studies [21, 22]. Its design was based primarily on the assessment of 45 food parameters (macronutrients, micronutrients, bioactive compounds, foods, and spices) associated with inflammation. Specifically, it assigned + 1 for increased pro-inflammatory markers or decreased anti-inflammatory markers, − 1 for decreased pro-inflammatory markers or increased anti-inflammatory markers, and 0 for no effect on inflammatory markers based on the effect of various food parameters on six inflammatory markers [interleukin (IL)-1β, IL-4, IL-6, IL-10, TNF-α, and C-reactive protein]. It calculated the weighted pro-inflammatory and/or anti-inflammatory scores of various food parameters based on the type of studied as well as the number of different types of studies from the literature included. The “food parameter-specific overall inflammatory effect scores” were then calculated from these weighted pro-inflammatory and anti-inflammatory scores. Subsequently, the global means and standard deviations of 45 food parameters were created based on 11 databases from different countries. Based on the global mean and standard deviation, the Z-score and centered percentile of the food parameters were calculated for participants from the various independent studies. After multiplying the central percentile of each food parameter by its respective “food parameter-specific overall inflammatory effect score”, a “food parameter-specific DII score” was obtained. Finally, the total DII score was obtained by summing all the “food parameter-specific DII scores”. In this study, the food or nutrient parameters used to calculate DII were excluded under the following conditions: (1) information on foods or nutrients was not collected or provided, (2) specific intake data were not provided, and (3) units of intake provided could not be directly used to calculate DII. Ultimately, 29 eligible food or nutrient parameters were determined in UK Biobank (Additional file 1: Table S1). Therefore, the DII in this study was derived based on 29 food or nutrient parameters. The median of total DII in this study was − 0.40 with a range of − 6.60 to + 5.45. For the calculation of the energy-adjusted DII (E-DII), we referred to previous studies [23, 24]. The food and nutrient intakes included were first adjusted for total energy intake (density method = nutrient (food)/total energy intake × 1000 kcal). Then, the steps similar to the DII calculation were repeated to obtain the E-DII.

Dementia was determined using “algorithmically defined outcomes (ADOs)” in UK Biobank [25]. These AODs were derived from coded information collected from the UK Biobank’s baseline assessment and admissions data (including data from participants’ self-reported medical conditions, procedures, and medications) as well as death registrations obtained from linked data from hospitals. These data were algorithmically combined and categorized using the International Classification of Diseases (ICD), 9th and 10th editions (ICD-9 and ICD-10) to identify dementia. This algorithm has been validated in a study based on 17,000 Biobank participants in England [26] and has also been used in several previous studies about dementia [27, 28].

The equipment used in this study for the measurement of brain deconvolution structures was the Siemens Skyra 3T scanner. Referencing previous studies based on UK Biobank [29], we assessed the association between DII and brain deconvolution structures with brain volume [30], total white matter hyperintensities volume [31, 32], total white matter volume [33], total gray matter volume [34], hippocampal volume [35], hippocampal gray matter volume [36], and lateral ventricle volume [37], which were measured during the imaging follow-up that started in 2014.

A touch screen questionnaire was used to capture age, gender, ethnicity (White, Black, Asian, Chinese, Mixed, and Other), smoking status (Never, Previous, and Current), drinking status (Never, Previous, and Current), education level (Obtained higher education qualifications (College/HND/HNC or equivalent), obtained professional qualifications, or not), physical activity (whether a person met the UK Physical activity guidelines of 150 min of walking or moderate activity per week or 75 min of vigorous activity), and family history of dementia (Yes or No) at baseline. The Townsend deprivation index (TDI) was derived from zip codes of residence, using aggregate data on automobiles, unemployment, homeownership, and household overcrowding. Diabetes was identified by reporting at the initial assessment visit that the participant had been diagnosed with diabetes by doctor or was regularly taking medication for diabetes. The blood pressure values in this study were obtained from the average of two seated measurements taken at baseline with the Omron HEM-7015IT digital blood pressure monitor. According to the American Heart Association (AHA) 2017 guidelines [38], we classified blood pressure into four groups: normal (systolic blood pressure (SBP) < 120 mmHg and diastolic blood pressure (DBP) < 80 mmHg; elevated (SBP 120–129 mmHg and DBP < 80 mmHg; stage 1 hypertension (SBP 130–139 mmHg or DBP 80–89 mmHg); and stage 2 hypertension (SBP ≥ 140 mmHg or DBP ≥ 90 mmHg). In addition, during the initial visit, trained professionals measured participants’ height and weight and obtained body mass index (BMI) by dividing weight (kg) by the square of height (m). BMI was then categorized into four groups: underweight (BMI < 18.5); healthy weight (BMI 18.5 to < 25); overweight (BMI 25 to < 30); and obese (BMI ≥ 30). Energy intake was obtained from 24-h dietary estimates.

First, we used a Cox proportional hazards model to quantify the association between DII and dementia. Results are expressed by hazard ratios (HRs) and 95% confidence intervals (CI). Three sets of models were applied. Model 1 adjusted for age, sex, ethnicity, education level, TDI, and energy intake. Model 2 further adjusted for diabetes, blood pressure status, alcohol consumption, smoking, BMI, and physical activity in addition to the factors adjusted for in Model 1. Model 3 further adjusted for family history of dementia based on Model 2. The Schoenfeld residual method tested the proportional risk hypothesis of the Cox model; all tested hypotheses passed. Additionally, the linear trend test was performed by using the median of each quartile of the DII as a continuous variable in the regression model. For observing the presence of nonlinear trends, the restricted cubic spline (RCS) model with the DII’s 5th, 27.5th, 50th, 72.5th, and 95th as the five knots was applied in this study. If the RCS revealed a nonlinear association, we further calculated the inflection point between DII and the risk of developing dementia and examined the relationship between DII and dementia using a two-segment Cox proportional hazards model on the two sides of the inflection point. Following previous studies [39], the inflection points were determined using the R package “segmented”, based on a likelihood-ratio test and the bootstrap resampling method.

We also performed subgroup analyses for the incidence of dementia, grouped by sex, BMI (< 25 or ≥ 25), smoking status, drinking status, physical activity, education level, TDI (above median or not), and family history of dementia, and tested for interactions between grouping variables and DII using likelihood ratio tests.

We used generalized linear regression to express the association between MRI-measured, brain-deconstructed structures, and DII with estimated β and 95% CI. β values demonstrated the amount of change in the volume of the brain deconstructed structure for one-unit increases in DII.

Additionally, we performed a series of sensitivity analyses. (1) We observed the association between the DII and the occurrence of all-cause dementia again by applying the competing risk models with death and lost follow-up as competing endpoints. For the occurrence of subtypes of dementia such as Alzheimer’s disease and vascular dementia, we explored the association with DII by using death, lost follow-up, and the occurrence of dementia other than itself as competing outcomes. (2) After excluding participants according to the standard of energy intake < 800 or > 4200 kcal/day for men and < 600 or > 3500 kcal/day for women, we reanalyzed the association between DII and dementia incidence using Cox proportional hazards regression to prevent bias caused by partially unreliable dietary data. (3) In order to further validate the validity of DII in this study, we analyzed the association of DII with C-reactive protein (CRP) (mg/L), which was measured by immunoturbidimetric-high sensitivity analysis on a Beckman Coulter AU5800. (4) Additional analyses with a set of 3 knots and 4 knots were performed to test the stability of the nonlinear correlations obtained by RCS. (5) We repeated the aforementioned analysis using E-DII to further validate and consolidate our results. (6) Considering that some participants’ diets may change frequently, we further replicated the analysis for those participants who indicated that their reported diet was typical, using information provided by UK Biobank on whether the reported diet was typical (Field: 100020). (7) Considering the large proportion of excluded population due to missing physical activity data, we reincluded this group of participants with missing data on physical activity in the analysis. We categorized them as “missing” for physical activity and then repeated the aforementioned analysis.

A total of 166,377 participants were included in this study, with a median follow-up time of 9.46 years. During the follow-up period, 1372 participants developed dementia, including 543 with Alzheimer’s disease (AD), 267 with vascular dementia (VD), and 54 with frontotemporal dementia (FTD). Table 1 presents the baseline characteristics of the participants in this study according to the quartiles of dietary inflammatory index (DII). Compared to the lower levels of DII, we found that participants with higher levels of DII tended to be younger, female, without family history of dementia, with lower energy intake, higher TDI, fewer recent drinkers, and more normal blood pressure; however, there was an increase in obesity and current smoking and a decrease in the number of people who had obtained high education or other professional qualifications and who achieved the recommended weekly level of physical activity. In addition, we also compared the characteristics of participants who completed the Oxford WebQ with those who had not completed it (Additional file 1: Table S2), as well as comparing the characteristics of participants who accepted and unaccepted MRI measurements in this study (Additional file 1: Table S3).

For the association between DII and dementia occurrence (Table 2), we found a significant association between higher levels of DII (Q4 vs. Q1) and increased risk of all-cause dementia (HR: 1.249; 95% CI: 1.052–1.483; P = 0.011) as well as other types of dementia (dementia other than AD, VD, and FTD) (HR: 1.539; 95% CI: 1.178–2.010; P = 0.002) when adjusting for multiple factors using proportional Cox regression models (Model 3). In addition, when DII was analyzed as a continuous variable, the HR was 1.046 for all-cause dementia and 1.096 for other types of dementia for each unit increase in DII, after adjusting for a range of influencing factors. In the trend test, a significant linear association was also observed between DII and all-cause dementia (P for trend = 0.021) as well as other subtypes of dementia (P for trend = 0.001).

For the nonlinear trend, we examined the association between DII and AD, VD, and FTD application of RCS (Fig. 2). We found a “J-shaped” association between AD and DII (P_nonlinear = 0.003). Specifically, a significant positive correlation was observed between DII and the incidence of AD when DII increased to a certain value. Based on the “segmented” package, we determined that the inflection point for AD was at 1.30. To further investigate this relationship, we used a segmented Cox proportional hazards model. Our analysis showed that for each unit increase in DII above the inflection point, the adjusted hazard ratio (HR) for the occurrence of AD increased by 39.1%. Conversely, for each unit increase in DII below the inflection point, the adjusted HR for AD decreased by 3.6%, although there was not statistically significant at this segment (Table 3). Besides, we did not observe any significant nonlinear relationship between DII and vascular dementia (P_nonlinear = 0.449) or frontotemporal dementia (P_nonlinear = 0.163).

In the present study, after adjusting for age, sex, ethnicity, education level, TDI, smoking, drinking status, physical activity, BMI, energy intake, blood pressure status, diabetes, and family history of dementia, we found a positive association between DII and total volume of white matter hyperintensities, with each unit increase in DII associated with 73.30 mm³ increase in the volume of white matter hyperintensities (P = 0.006) (Table 4). Meanwhile, we also observed a decreasing trend in the volume of hippocampal gray matter with increasing DII (β: − 7.617; 95% CI: − 13.986 to − 1.248; P = 0.019).

In subgroup analyses (Fig. 3), the results remained approximately consistent when grouped by sex (P value for interaction = 0.814), smoking status (P value for interaction = 0.540), drinking status (P value for interaction = 0.654), physical activity (P value for interaction = 0.343), education level (P value for interaction = 0.767), TDI (P value for interaction = 0.984), and family history of dementia (P value for interaction = 0.331). In contrast, a significant effect change was observed when the data were grouped by BMI (P value for interaction = 0.044), whereby increased DII was significantly associated with an elevated risk of developing dementia only in those with BMI ≥ 25.

In competing risk models (Additional file 1: Table S4), a significant linear association was still observed between DII and all-cause dementia and other types of dementia (dementia other than AD, VD, and FTD). In analyses conducted after excluding participants with unreliable energy intake, similar results to the original analysis were observed (Additional file 1: Table S5). When constructing RCS models with 3 knots and 4 knots, both of indicated a significant nonlinear association between DII and AD (Additional file 1: Fig. S1-2).

Regarding the association of DII and inflammatory markers, the generalized linear regression revealed a significant positive association between DII and C-reactive protein concentration (β: 0.075; 95% CI: 0.063 to 0.087; P < 0.001), and this association persisted after adjusting for a range of factors (Additional file 1: Table S6).

Further, when using the energy-adjusted dietary inflammation index (E-DII) for the analysis, we obtained similar results to those obtained when using the DII analysis. Specifically, elevated E-DII levels was also found to be associated with an increased risk of all-cause dementia or other types of dementia (dementia other than AD, VD, and FTD) (Additional file 1: Table S7). For AD, the risk analysis also showed an “J-shaped” nonlinear association with E-DII (Additional file 1: Fig. S3), with an inflection point of 0.56 (Additional file 1: Table S8).

When participants with a typical diet were analyzed separately, the results were generally consistent with the analysis when the total population was used (Additional file 1: Fig. S4 and Table S9-10). When participants with missing information on physical activity were reincluded for analysis, the results were also generally consistent with those initially obtained (Additional file 1: Fig. S5 and Table S11-12).