The present study was based on data from the fifth wave (June 2012 to February 2013) of an occupation-based dynamic cohort established in Tokushima Prefecture in Japan. Details of the occupation-based annual examinations in Tokushima Prefecture have been reported elsewhere [23]. The participants were followed up every year. For dietary intake and physical activity assessment, the participants were followed up with an interval of 5 years from the fifth wave. The study population in the fifth wave included 1398 men and women aged 20–63 years who participated in the fifth wave, and participants for whom there were incomplete data for lipid markers (n = 1), physical activity (n = 1), and drinking habits (n = 3) were excluded. From the 1393 participants, subjects who had nonfasting blood tests (n = 82) and subjects with dyslipidemia (n = 413) at the baseline survey were excluded. Next, we excluded subjects who only participated in the baseline survey (n = 153) were excluded. Out of 745 participants with 3701 observations for the follow-up term, we also excluded seventeen observations with a non-fasting blood test in the follow-up period (the number of participants was therefore not changed). The remaining 745 subjects (508 men and 237 women) with 3,684 observations were used for analysis (Fig. 1).

Overview of the participants. For 1398 participants aged 20–63 years, we excluded 5 subjects due to missing information on lipid profile, drinking habit, and physical activity. We excluded subjects who had nonfasting blood tests (n = 82) and subjects with dyslipidemia (n = 413) at the baseline survey. Next, we also excluded subjects who only participated in the baseline survey (n = 153). The remaining 745 subjects (508 men and 237 women) were used for analysis.

The participants were asked about meals taken in the past month using a food frequency questionnaire (FFQ), “FFQg ver 2.0” (Kenpakusha Inc.), to determine the frequency of food intake. The amount of food intake was calculated as the product of the frequency and amount consumed at each meal for the following 17 food groups: cereals, potatoes and starches, deep yellow vegetables, other vegetables and mushrooms, algae, pulse, fish, mollusks and crustaceans, meat, eggs, milk and milk products, fruits, confectionaries, beverages, sugar and sweeteners, nuts and seeds, fats and oils, and seasoning and spices [24]. The validity of this FFQg was verified by comparing food intake amounts using the weighting method for seven consecutive days. The correlation coefficient between the FFQg and 7-day records for energy was 0.47 and the correlation coefficient between the FFQg and the intakes of cereals, meat, fish, milk, dairy products, green-yellow vegetables, other vegetables, and fruits were 0.76, 0.27, 0.27, 0.72, 0.58, 0.46, 0.53, and 0.64, respectively, and there was a significant correlation between the two methods at the p < 0.05 level for 22 of the 29 food groups [25]. The questionnaire included questions on how many times and how much of each food group was consumed per week. The amount of food consumed (per week) was calculated as the frequency of eating and the amount of the product at each meal.

The dietary diversity score was calculated using the Quantitative Index for Dietary Diversity (QUANTIDD) developed by Katanoda et al. [26]. The QUANTIDD score is calculated by the proportion of foods that contribute to total energy or the amount of food and the number of food groups using the following formula:

$$}=\frac\nolimits_}}^}}}(})}^\,}}}},$$

where prop [j] is the proportion of each food group (s) j contributing to total energy or nutrient intake and n is the number of food groups. In this study, we estimated the score based on the amounts of 16 food groups excluding the beverages group. The possible score ranges from 0 to 1. A high score indicates consumption of many kinds of food groups and a low score indicates consumption of only a few kinds of food groups.

The participants were further divided into 3 tertiles of dietary diversity score according to their total diversity score: tertile 1 (lowest dietary diversity) to tertile 3 (highest dietary diversity). QUANTIDD scores were: 0–0.7982 for T1, 0.7983–0.8619 for T2, 0.8620–1.00 for T3 in men and 0–0.8499 for T1, 0.8500–0.8991 for T2, 0.8992–1.00 for T3 in women.

Dietary patterns that consist of a correlation matrix for 17 food groups were assessed using principal component factor analysis. The principal components were selected on the basis of eigenvalues > 1.3 and interpretability (Supplementary Table 1). Principal component scores were saved for each individual and were used as continuous variables. The first principal component was named the healthy pattern because it contained higher factor loading of ten food groups: other vegetables and mushrooms, deep yellow vegetables, pulses, algae, fish, mollusks and crustaceans, potatoes and starches, sugars and sweeteners, fruits, seasonings and spices, and fats and oils. This pattern explained 22.8% of the variance. The second principal component, named the western pattern, consisted of higher factor loading of four food groups including beverages, meat, cereals, and eggs and it explained 9.7% of the variance. The third component, named the sweetener pattern, consisted of higher factor loading of three food groups including confectionaries, nuts, and milk and milk products and it explained 7.7% of the variance.

Data for serum concentrations of total cholesterol (TC), triglycerides (TG), LDL-cholesterol, and HDL-cholesterol were obtained during medical health check-ups. Fasting venous blood samples after an overnight fast (at least 10 h) were obtained from each subject in the morning for serum biochemical measurements in both the baseline survey and follow-up surveys. Non-HDL cholesterol was calculated by the following formula: total cholesterol minus HDL cholesterol.

Body height was measured to the nearest 0.1 cm with participants standing without shoes or sandals. Body weight was measured to the nearest 100 grams with participants wearing light clothing. Body mass index (BMI) was calculated using the following formula: BMI = weight (kg)/[height(m)]2. Daily values of physical activity (MET-hours/week) were calculated using the International Physical Activity Questionnaire [27]. Information on education level (categorical; elementary, junior high and high school, tertiary college, career college and junior college, college, and graduate school or other), drinking habit (categorical; current, former, never), and smoking habit (categorical; current, former, never) was obtained by a self-administered questionnaire. Subjects completed the questionnaire before the day of physical examination, and the questionnaire was checked and collected at baseline. The diagnosis of dyslipidemia was based on the criteria of the Japan Atherosclerosis Society [28]. Subjects were diagnosed as having dyslipidemia at baseline when they fulfilled at least one of the following conditions: TG ≥ 150 mg/dl (hypertriglyceridemia), LDL-cholesterol ≥140 mg/dl (high LDL-cholesterol), HDL-cholesterol <40 mg/dl (low HDL-cholesterol), non-HDL-cholesterol ≥170 mg/dl (high non-HDL-cholesterol) and having a history of dyslipidemia.

The basic characteristics of participants according to dietary diversity are shown for each gender. Continuous variables were presented as means ± standard deviation (SD) or medians (25 percentile, 75 percentile), and comparisons of mean values were made using ANOVA or Kruskal-Wallis’s test. The Jonckheere-Terpstra test was used to calculate the p for trend. Categorical variables were expressed as numbers (percentages, %), and comparisons of proportions were made using the chi-square test. The Mantel-Haenszel test was used to calculate the p for trend.

The cumulative means of serum lipid markers during the follow-up period were analyzed using generalized estimating equations (GEEs). A GEE takes into account the dependency of repeated observations within participants. An additional advantage of a GEE is that missing values can also be used during analysis. Thus, subjects who were lost to follow-up surveys after the early wave examination were also included in the analyses. GEE models were fitted by the GENLIN syntax in SPSS. This procedure corresponds to generalized linear models. In the present analyses, compound symmetry was specified for the correlation structure. GEE analyses were used to estimate the adjustment cumulative mean of lipid markers in follow-up surveys according to the tertile of dietary diversity at baseline after controlling for the following variables. The confounding variables that were adjusted in Model 1 were age, energy intake (continuous, kcal/day), body mass index (continuous, kg/m2), physical activity (continuous, MET-hours/week), beverage consumption (continuous, grams/day), smoking habit (categorical; current, former, never), education (categorical; elementary, junior high and high school, tertiary college, career college and junior college, college, and graduate school or other), and follow-up time (continuous, years). Model 2 consisted of the adjustments in Model 1 plus healthy dietary pattern score, Model 3 consisted of the adjustments in Model 1 plus western dietary pattern score, and Model 4 consisted of the adjustments in Model 1 plus sweetener dietary pattern score.

All statistical analyses were performed separately by gender using SPSS (IBM Corporation, Tokyo, Japan) version 29.0 for Windows. All statistical tests were based on two-sided probabilities, and all p values < 0.05 were considered statistically significant.