Intake of whole grain foods and risk of coronary heart disease in US men and women

The Nurses’ Health Study (NHS) was initiated in 1976, when 121,700 female registered nurses aged 30–55 years answered a mailed questionnaire on their medical history and lifestyle characteristics. A parallel cohort study of younger women, the Nurses’ Health Study II (NHSII), was established in 1989 and included 116,340 eligible female nurses aged 25–42 years. A questionnaire similar to that used in NHS was administered at baseline to assess medical history and lifestyle factors. In 1986, the Health Professionals Follow-up Study (HPFS) was started and recruited 51,529 US male health professionals aged 40–75 years. The HPFS participants completed a baseline questionnaire that was similar to that used in the NHS and NHSII. In all three cohorts, participants were sent questionnaires biennially to update their diet and lifestyle information and identify newly diagnosed CHD and other diseases. The cumulative response rates in three cohorts exceeded 90% [19, 20].

The study baseline was set to be 1984 for NHS, 1991 for NHSII, and 1986 for HPFS, when a comprehensive semi-quantitative food frequency questionnaire (sFFQ) was administered to assess diet. The exclusion criteria included the presence of cardiovascular disease or cancer at baseline (n=5942 in HPFS, 3869 in NHS, 1104 in NHSII), not returning the sFFQ or had unusual total energy intake (<500 or >3500 kcal/day for women and <800 or >4200 kcal/day for men) (n=1595 in HPFS, 15,753 in NHS, 13,001 in NHSII), or completing baseline questionnaire only (n=694 in HPFS, 523 in NHS, 624 in NHSII). The final study population consisted of 74,244 participants in NHS, 91,430 in NHSII, and 39,455 in HPFS.

The study protocol was approved by the Human Research Committee of Brigham and Women’s Hospital and the Harvard T.H. Chan School of Public Health. Completion and return of study questionnaires implied informed consent of the participants.

In all three cohorts, diet was assessed using the validated sFFQ at baseline and updated every 2–4 years during the follow-up until 2014 in NHS, 2015 in NHSII, and 2016 in HPFS. For each food item listed in the sFFQ, the participants were asked their average consumption frequency of a pre-specified portion size during the previous year. There are nine possible responses for consumption frequencies ranging from never or <1 time/month to ≥6 times/day. In the current analysis, we focused on the consumption of seven commonly consumed whole grain foods/components, including cold breakfast cereals (serving size, 1 cup), dark bread (serving size, 1 slice), popcorn (serving size, 1 cup), oatmeal (serving size, 1 cup), bran added to food (serving size, 1 teaspoon), wheat germ (serving size, 1 teaspoon), and brown rice (serving size, 1 cup). Based on the brand names provided by the participants, we further classified cold breakfast cereal into whole grain-based breakfast cereal and refined grain-based breakfast cereal, depending on the relative contents of whole grain ingredients in the product (containing ≥25% whole grains or bran by weight classified as whole grain cold breakfast cereal). Since 2002 in the NHS and HPFS, and 2003 in the NHSII, an additional question regarding the types of popcorn was added to the sFFQ in which participants were asked whether they consumed regular or light/fat free popcorn. We estimated the total whole grain consumption by first transforming the reported serving size into grams and then summed up the weight of whole grain ingredients according to corresponding whole grain contributions in all grain-containing food [21]. The seven individual whole grain foods contributed on average 86% of total whole grain in NHS, 78% in NHSII, and 84% in HPFS during the follow-up. Validation studies showed the validity of whole grain foods assessments. For example, the FFQ assessments were significantly correlated with those assessed using multiple-day diet records; the correlation coefficients were 0.58 for dark bread and 0.73 for cold breakfast cereal [22].

In all three cohorts, a similar follow-up questionnaire was mailed to participants to assess and update information regarding smoking status, vitamin supplements use, alcohol consumption, menopausal status (women only), and years of postmenopausal hormone use (women only), physician-diagnosed hypertension and hypercholesterinemia, and other time-varying variables. Height was reported at baseline, and body weight was updated biennially. Body mass index (BMI) was calculated as weight in kilograms divided by the square of height in meters (kg/m²). Recreational physical activity was measured with a validated questionnaire asking about the average time spent on 10 common activities. Based on this information, we calculated weekly energy expenditure in metabolic equivalent (METs) hours weighting each activity by its intensity level [23]. Multiple validation studies demonstrated adequate validity of these self-reported variables [24‐28]. We used the alternative healthy eating index (AHEI) [29], after removing the whole grain component, to represent overall diet quality.

Total CHD including nonfatal myocardial infarction (MI) and fatal CHD was the primary disease outcome for the current analysis. In all three cohorts, permission was sought to access medical records of participants who reported having a nonfatal MI on a follow-up questionnaire. Study physicians who were blinded to exposure status reviewed the medical records and confirmed a reported MI according to the WHO criteria, which require the presence of symptoms, and either typical electrocardiographic changes or elevated cardiac enzyme levels [30, 31]. Deaths were identified through reports from the next of kin, the postal authorities, or by searching the National Death Index (NDI) [32]. Fatal CHD was confirmed by a review of hospital records or autopsy reports if CHD was listed as the underlying cause of death and if evidence of previous CHD was available from medical records [33]. Sudden deaths without cardiac causes were not considered as fatal CHD in the current analysis.

Cumulative averages of total whole grains and individual whole grain foods were calculated to represent long-term intake using the formula \({Y}_n=\frac{\sum_i^n{Y}_i}{n}\), where Y_n denotes the intake level at nth follow-up cycle [34]. For each participant, we counted their person-time from the return date of the baseline FFQ to the CHD diagnosis date, death date, date of last return of a valid follow-up questionnaire, or the end of follow-up (2014 for NHS, 2016 for HPFS, and 2017 for NHSII), whichever occurred first. To alleviate the potential reverse causality that participants with existing diseases might change their usual diet intake that is etiologically relevant, we stopped updating dietary information once the participants developed diabetes, stroke, coronary artery bypass graft, hypertension, hypercholesterinemia, or cancer during follow-up. The proportion of missing values of covariates ranged from 4.6% for smoking status to 5.1% for BMI. We replaced missing values with valid values in the preceding questionnaire for one follow-up cycle and then created missing indicators to handle subsequent missing values.

An age- (months) and calendar time-stratified multivariable-adjusted Cox proportional hazards model was used to estimate the hazard ratios (HRs) and 95% confidence intervals (CIs) for the association between total whole grains as well as individual whole grain foods and risk of CHD. The risk set was defined by both age and calendar year which minimized the potential confounding effects by age and time trend and accommodated the time-varying modeling at the same time. The proportional hazards assumption was evaluated by including product terms between each categorical whole grain variable and the duration of follow-up calculated as months from baseline to CHD diagnosis date, death, or the end of follow-up, whichever occurred first. The proportional hazards assumption was evaluated by including product terms between each categorical whole grain variable and the duration of follow-up calculated as months from baseline to CHD diagnosis date, death, or the end of follow-up, whichever occurred first. The likelihood ratio tests were used to compare models with and without interaction terms and none of the p values were statistically significant in pooled results, suggesting no violations of the proportional hazard assumption. Total whole grain intake was categorized into fifths using quintiles according to cohort-specific distributions in each follow-up cycle. We used three pre-specified categories (< 1 serving/month, 1 serving/month to 1 serving/week, and ≥ 2 servings/week) for oatmeal, brown rice, added bran, and wheat germ which had modest-to-low intake, whereas four pre-specified categories (< 1 serving/month, 1 serving/month to 1 serving/week, 1 serving/week to 4–6 servings/week, and ≥ 1 serving/day) were used for whole grain cold breakfast cereal, dark bread, and popcorn, for which the consumption level were on average higher. Covariates included in the models were ethnicity (white, African American, Asian, others), time-varying BMI (<21.0, 21.0–22.9, 23.0–24.9, 25.0–26.9, 27.0–29.9, 30.0–32.9, 33.0–34.9, or ≥35.0 kg/m²), smoking status (never smoked, past smoker, currently smoke 1–14 cigarettes per day, 15–24 cigarettes per day, or ≥25 cigarettes per day), alcohol intake (0, 0.1–4.9, 5.0–9.9, 10.0–14.9, 15.0–29.9, and ≥30.0 g/day), multivitamin use (yes, no), physical activity (quintiles), modified AHEI (quintiles), total energy (quintiles), family history of myocardial infarction (yes, no), baseline diabetes (yes, no), postmenopausal hormone use (women only; never, former, or current hormone use, or missing), and oral contraceptive use (yes, no; women only). The continuous variable of total whole grain consumption and each individual whole grain food was used to calculate p value for trend and HR (95% CI) of CHD per one serving/day of intake. We also evaluated the substitution effects of replacing 10 g of total refined grain with the same amount of total whole grain on CHD risk by mutually adjusting two variables in the same model. Differences in their β coefficients were used to estimate the HRs for the substitution association, and their variances and covariance matrix were used to derive the 95% CI [34].

We evaluated the heterogeneous associations among individual whole grain foods with CHD risk using a likelihood ratio test comparing two models after adjusting for other covariates: one with total whole grains only and the other with total whole grains plus seven individual whole grain foods in categorical terms. Moreover, in the subgroup analysis, we explored the effect modification by several lifestyle factors, including BMI, physical activity, smoking status, and family history of myocardial infarction. P values for interactions were calculated from likelihood ratio tests comparing two nested models: the full models with product terms between quintile of total whole grain and categorical lifestyle factors and the reduced models without the product terms. In a secondary analysis, we analyzed the associations for regular or fat-free/light popcorn consumption with CHD risk. Data from each cohort were analyzed separately, and results were pooled using a fixed-effects model. Finally, to delineate the dose-response relationship between whole grain and CHD risk, we combined data from three cohorts and fitted cubic spline regressions with the same covariate adjustment in the primary analysis (except for women only variables) for total whole grain and individual whole grain foods. The whole grain intake levels were truncated at 99.5th percentile to avoid the influence of extreme values. Total whole grain intake was converted to servings by dividing a factor of 16 according to the dry weight estimation of serving size [35, 36]. Corresponding to 5th, 35th, 65th, and 95th percentile of total whole grain intake, four knots at 0.17, 0.76, 1.4, and 2.9 servings/day were used for the cubic splines. We used likelihood ratio tests with 2 degree of freedom (knots number −2) to calculate p value for non-linearity by comparing models with linear term only and models with both linear and spline terms.

Several sensitivity analyses were performed. First, because the strategy of stopping updating dietary information upon occurrence of chronic conditions may introduce differential errors in dietary assessments between diseased participants and other participants, we conducted a sensitivity analysis by resuming to update diet 8 years after the incidence of these intermediate outcomes. For example, for participants who developed stroke in 1990, we stopped updating diet in 1994 and 1998 and then continued to update diet after 1998. Second, we used the baseline intake or the simple-updated time-varying intake instead of the cumulative averaged value to repeat the analyses. Finally, we conducted a 4-year latency analysis to address the potential reverse causation bias. All statistical tests were 2-sided with significant level of 0.05 and performed using SAS 9.3 (SAS Institute, Cary, NC).