This study was embedded within the first two sub-cohorts of the Rotterdam Study (RS), a prospective population-based cohort among inhabitants from the suburb Ommoord in Rotterdam, the Netherlands. Details regarding the study methodology have been published previously . Briefly, the first sub-cohort (RS-I) was established in 1989 and consisted of 7983 participants aged 55 years and older. In 1999, the study was expanded with the second sub-cohort (RS-II) consisting of 3011 participants who had turned 55 years of age or moved into the study area. Extensive follow-up examination rounds take place every 3–5 years through home interviews and various physical and laboratory checks at the research center.
For the current study, we consider two different baselines. The periods considered as baselines were dependent on the examination rounds in which dietary intake was assessed: between 1989 and 1993 in the first cohort (RS-I-1), which forms baseline I in the current study, and between 2009 and 2012 in the first and second cohort (RS-I-5 and RS-II-3), which forms baseline II in the current study.
Of the 7983 participants included in the study at baseline I, 5435 participants had dietary data available. We excluded 3 participants who had unreliable dietary data (i.e., an estimated energy intake of < 500 or > 5000 kcal/day), 22 participants who had dementia at the time of dietary assessment, and 35 participants who did not the sign informed consent to link the study database to their medical records. This leaves a total of 5375 participants eligible for follow-up from baseline I. Of the 4040 participants who participated at baseline II, 2998 participants had dietary data available. We excluded 110 participants who had unreliable dietary data, 23 participants who had dementia at the time of dietary assessment, 1 participant who had insufficient cognitive screening to assess dementia, and 3 participants who did not sign informed consent to link the study database to their medical records. This leaves a total of 2861 participants eligible for follow-up from baseline II. A schematic overview of the study population is provided in Additional file 1.
At baseline I, participants completed a 170-item food frequency questionnaire (FFQ). They first completed a checklist on which food items they consumed at least twice a month in the preceding year, after which information on frequencies and portion sizes was obtained in an interview by a trained dietician. At baseline II, dietary intake was assessed with a self-administered 389-item FFQ including questions on frequency and portion sizes of food item consumption in the last month. Both FFQs have been validated against other dietary assessment methods which showed that based on these FFQs, participants can be adequately ranked according to their food and nutrient intake [23,24,25]. From the FFQ data, we derived adherence scores for the MIND diet, Dutch dietary guidelines, and Mediterranean diet, as outlined below.
The MIND diet as described by Morris et al.  contains recommendations regarding 15 food components, including 10 food components considered to be healthy for the brain (i.e., green leafy vegetables, other vegetables, nuts, berries, beans, whole grains, fish, poultry, olive oil, and wine) and five unhealthy food components (i.e., red meat, butter and stick margarine, cheese, fast fried food, and pastries and sweets). An overview of food items on the different FFQs that summarizes these food components can be found in Additional file 1. If participants used olive oil as the primary cooking fat (> 50%), a 1 was assigned and a 0 otherwise. For each other food component, a 0 was assigned if participants did not adhere to the recommendations, a 0.5 for moderate adherence, and a 1 for good adherence. Scores assigned to each food component were summed, obtaining a total score ranging from 0 to 15.
Dutch dietary guidelines
We used a previously defined score to assess adherence to Dutch dietary guidelines . Briefly, participants received a score of 1 (adherence) or 0 (no adherence) for recommendations of 14 food components (i.e., vegetables, nuts, fruits, legumes, whole grains, whole grains of total grains, fish, dairy products, tea, coffee, unsaturated fats and oils of total fats, red and processed meat, sugar-containing beverages, alcoholic beverages, and salt). The sum score ranged from 0 to 14.
The Mediterranean diet is described by Panagiotakos et al.  containing recommendations regarding 11 food components (i.e., vegetables, fruits, legumes, whole grains, fish, full-fat dairy products, potatoes, olive oil, poultry, meat, and alcoholic beverages). Adherence was determined by assigning a score ranging from 0 to 5 to each food component with higher scores reflecting better adherence. The final sum score ranged from 0 to 55.
Participants were screened for dementia at baseline and every 3–5 years during follow-up examinations using the Mini-Mental State Examination (MMSE) and the Geriatric Mental Schedule (GMS) organic level. Those with an MMSE score of < 26 or a GMS organic level score of > 0 were further examined using the Cambridge Examination for Mental Disorders in the Elderly diagnostic interview. Additionally, participants were continuously under surveillance for dementia through the electronic linkage between the study database and medical records from general practitioners and the Regional Institute of Outpatients Mental health Care. The final diagnosis of dementia and its most common subtypes was made by a consensus panel led by a neurologist based on the standard criteria for all-cause dementia (DSM-III-R) and for sub-diagnosis of Alzheimer’s disease (NINCDS-ADRDA). Follow-up for dementia was completed until January 1, 2018.
Data on relevant covariates were obtained at both baselines I and II. Trained interviewers obtained information regarding education attainment (primary, lower, intermediate, higher), smoking status (never, former, current), and use of medication. Height and weight were measured, and body mass index (BMI) (kg/m2) was calculated. Physical activity was measured using a validated adapted version of the Zutphen Physical Activity Questionnaire at baseline I and the LASA Physical Activity Questionnaire at baseline II. Physical activity was expressed in metabolic equivalent of task (MET)—hours per week. Daily energy intake in kilocalories was calculated from the FFQ data using the Dutch Food Composition Tables (NEVO). Diabetes mellitus was defined as having a fasting serum glucose of ≥ 7.0 mmol/L, a random serum glucose level of ≥ 11.1 mmol/L, or the use of blood glucose-lowering medication. Hypercholesterolemia was defined as a serum total cholesterol concentration ≥ 6.2 mmol/L or the use of lipid-lowering medication. Systolic and diastolic blood pressure was measured twice on the right arm with the participant in a sitting position using a random zero sphygmomanometer of which the mean was used for analyses. Hypertension was defined as a systolic blood pressure of ≥ 140 mmHg, a diastolic blood pressure of ≥ 90 mmHg, or the use of blood pressure-lowering medication. Depressive symptoms were considered as a score of ≥ 16 on the validated Center for Epidemiology Depression Scale. History of stroke was obtained from interviews and verified through medical records. APOE genotype was obtained using polymerase chain reaction of coded DNA samples for RS-I and with bi-allelic TaqMan assay for RS-II.
Cox proportional hazard models were used to determine the association between the different diet scores per standard deviation (SD) increase and incidence all-cause dementia. Analyses were conducted from baselines I and II separately. Participants were censored when they were diagnosed with dementia, died, were lost to follow-up, or at the end of the follow-up (January 1, 2018), whichever came first. To test for potential non-linear relationships, we added natural cubic splines with three knots to the diet scores in het model and tested whether this significantly improved the fit of the model using likelihood ratio tests. To determine how associations changed over time, we performed analyses in cumulative follow-up intervals from the different baselines (i.e., performing analyses from the different baselines to 5 years, baselines to 7 years) . We constructed a basic model adjusted for sex, age, age^2, and educational attainment (model 1). Subsequently, we further adjusted for smoking status, physical activity, and energy intake (model 2). To minimize the risk of residual confounding, we considered an additional model in which we further adjusted for covariates that may act as confounders and/or mediators, which include BMI, diabetes, hypercholesterolemia, and hypertension (model 3). Missing data on covariates (29% for physical activity and < 5% for all other covariates) were imputed using five-fold multiple imputation. The distribution of the covariates in the imputed dataset was comparable to the original dataset (data not shown). Analyses were performed on each imputed dataset, and the results were presented as pooled hazard ratios (HRs) with 95% confidence intervals (95% CIs). All analyses are repeated considering Alzheimer’s dementia as the outcome variable. Possible effect modification by sex, educational attainment, smoking status, and APOE ε4 genotype (carrier vs. non-carrier) was investigated by including multiplicative interaction terms to the MIND diet score in model 2, and if the interaction term was statistically significant (p < 0.05), we performed stratified analyses.
To examine if single food components of the MIND diet drove the observed associations, we repeated the analyses with versions of the MIND diet score for which each individual food component was one at a time excluded from the total score and included as a covariate in the het model. Furthermore, as the MIND diet covers five unhealthy food components which are not covered in the Dutch dietary guidelines and Mediterranean diet, we excluded all unhealthy food components altogether from the total score and included these five components as a covariate in the model to determine whether these unhealthy food components together drove the association.
To ensure the robustness of our findings, we conducted several sensitivity analyses. First, we repeated the analyses for participants above and below the age of 75 years separately. Second, as cognitive impairment might have influenced the reliability of dietary recall, we repeated the analyses after excluding participants with an MMSE score of < 26 at the time of dietary assessment. Third, participants with a history of stroke at dietary assessment were excluded and censored at the date of incidence of stroke. Fourth, having depressive symptoms is an important confounder in the association between dietary intake and dementia risk, but we had no data on depressive symptoms for 58.2% of the participants. We therefore repeated the analyses after excluding all participants with depressive symptoms or missing data on depressive symptoms. Finally, we repeated the analyses after excluding the first 5 years of follow-up, to assess potential reverse causality.
All statistical analyses were conducted using the R Statistical Software version 4.0.3.