We translated self-reported food intake (using the commonly used German DEGS1-FFQ) of different samples into detailed nutrient intake per day for various macro- and micronutrients and dietary components, assessed test-retest and between-timeframe reliability, and made the scoring scripts openly available. Although we assume a large proportion of under- and overreporting on the individual’s level, the nutrient values were normally distributed and mostly met recommended reference intake on average, except for higher sugar and less fiber intake then recommended. This is in line with studies across different European countries showing that sugar intake can make up to 115 g/d or 20% of overall energy intake [37] and fiber intake only 14–21 g/d not reaching recommended levels [38]. As expected, males consumed higher amounts of nutrients than females, and differences in nutrient intake were present across groups of different dietary adherence or eating habits, such as between omnivorous and vegetarian dieters or dietary fat and sugar indices.
Reliability between last 7 days and 24 hours FFQ was medium for all macro- and micronutrients. In the validation study of the (original) DEGS1-FFQ on the food group level, reliability was assessed between 28 days FFQ and two 24h recalls by phone and ranged from low to good across 53 food items. Most validation studies use multiple 24 hour recalls per participant to assess dietary intake of different days. This way, day-to-day variations of food consumption can be recorded [39] and the imprecisions of FFQs at the individual level can be adjusted [40, 41]. Combining FFQ and 24hR has been shown to increase the correlation with actual dietary intake. While repeated administration of 24hR improves consistency with actual dietary intake and decreases random error, the benefit is marginal, with four 24hRs sufficing to reach a maximum correlation of 90% [40]. While the combination of FFQs with 24hR is beneficial, the use of 24hR to validate FFQs is not advised [42]. In comparison to the item-wise correlation, we computed reliability for nutrient intake summed over all food groups for 7 days and 24 hours assessed with the same online tool. The overall better performance of our nutrient scoring compared to the original food group scoring can be attributed to the reduction of outcome nutrient variables compared to food groups and the assumed higher consistency for overall nutrient intake patterns compared to single food item intake. The provided nutrient scoring of the DEGS1-FFQ can thus be considered valid in terms of assessing relatively similar nutrients and food groups when reporting 7 days or, to a lesser degree, 24h.
Assessing reliability over two timepoints within one month, we found moderate agreement for FFQ7d (kappaall ≥ 0.40, kappamax = 0.73), similar to previous validation studies [43, 44], with highest reliability for protein and lowest for fat. For FFQ24h, reliability was poor (kappaall ≥ 0.08, kappamax = 0.34), with highest reliability for carbohydrates and lowest for fiber. Indeed, the difference between FFQ7d and FFQ24h could reflect individual variance in eating habits that are more consistent over a time course of one week than on single days, including weekend days, when actual food intake might change quite drastically. To achieve highest correlation the administration of four 24hR has been recommended [40]. Therefore, the use of only two FFQ24h in our study might have led to underpowered results and lower consistency. Overall, FFQs relating to shorter time periods (e.g. 24h) may be helpful to assess diet as a confounder variables, e.g. for microbial sampling, yet reliability is poor. FFQs with longer time periods (e.g. 7 days or more) have higher reliability and should be used for assessing more habitual dietary habits or intake or adapted to the length of the intervention period.
We further evaluated the sensibility of the computed nutrient data and extended our nutrient scoring for DEGS1-FFQ data to two additional samples. The male-only sample with vegetarians and omnivores showed similar nutrient results as the sample with a wider BMI range and omnivores, with higher reported intake for the high fat and sugar group (for comparison: total energy intake: GUT-BRAIN: 1490–1740 kcal; Mensa: 1450-1640kcal, GREADT: 1710–2530 kcal), indicating that our results reflect common variance in actual eating habits. In particular for the male sample, the grouping into high vs. low dietary fat and sugar intake, showed distinct group differences for nutrient classes, sugar intake was much higher than recommended in both groups and fiber similarly lower than recommended. Further grouping into high vs. low fat and sugar intake showed distinct group differences for reported energy and nutrient intake (i.e. protein, fat, sugar, carbohydrates, saturated fats and others).
The DFS grouping might be limited in terms of comparing plant-based derived fiber intake as the questionnaire is by design focusing on processed and animal-based products. Sensible DFS grouping was confirmed by higher HbA1c levels in the high DFS group, reflecting a risk for diabetes development, coronary heart disease or stroke in the long-term [45].
Associations of anthropometrics and biomarkers with nutrient intake showed mixed results. First, all nutrients were collinearly linked. As expected, in the omnivorous, overweight sample higher total fiber intake across 7 days was weakly linked to lower BMI, body fat mass and mean diastolic blood pressure, pointing towards healthier diets high in fiber intake relating to lower weight in that sample. A link between fiber and lower weight has been shown before for diets restricted in animal-based foods [46] and systematically reviewed for whole-grain and fiber-rich foods [8]. All other nutrients were positively linked to BMI. However, when running linear models accounting for interdependency of datapoints and with fiber adjusted for caloric intake, these associations were not significant. When looking at the larger merged sample spanning from normal-weight to obese, higher total fiber and its subclasses (insoluble, soluble, cellulose, lignin) intake remained negatively linked to BMI and WHR. Yet again, linear mixed models showed no significant association. The fiber-BMI link was more distinct in sex-stratified analyses. All nutrients, including fiber, were linked to higher BMI in males. In females, however, higher fiber intake was strongly linked to lower BMI. Sex-specificity was further shown by an anti-correlation of body fat and fat-free mass in males, with the inverse relation in females, who showed high accordance of BMI and body fat mass.
Some studies reported inverse findings compared to ours: whole-grain intake was associated with lower BMI for both sexes, yet fiber in particular was inversely correlated with BMI only in men, not in women [47, 48] and likewise with immune function [49]. In contrast to previous studies [50, 51], fiber intake was not different between males and females in our sample, therefore fiber intake differences might have not played a strong role in sex-specific metabolic mechanisms. Proposed mechanisms of fiber intake in women may be metabolic benefits, i.e. reduced lipids in the blood, mediated by estradiol levels [52] and even blunted hormonal signaling during the reproductive cycle [53]. The picture on sex-specific associations of fiber intake on anthropometrics seems rather inconclusive and more studies are needed to disentangle sex-specific effects of fiber intake on metabolic, immune or reproductive markers.
Although we found links between nutrient intake and anthropometric markers, most associations were non-significant when adjusting for computed total energy intake. Although this might not reflect real total intake, proportions of nutrients may be relevant and telling within the same nutrient scoring pipeline within one dataset. We recommend adjusting nutrient-of-interest variables for overall calorie intake as has been done previously [54–56].
A strength of this analysis is the pooling of data from two human studies with deeply phenotyped samples with research questions focusing on eating habits resulting in a larger dataset with 189 data points. By increasing power, we could show that negative links between nutrient intake and BMI and WHR are strongest for fiber and its subclasses compared to all other nutrients. Sex-stratified analyses showed stronger, but opposite links for fiber and anthropometric measures in men.
These results encourage researchers to regard calculated nutrient intake as a putative measure of interest to be extracted from FFQ data. Such calculation pipelines are rarely if at all available. Therefore we publish all scripts open access and open code. Overall, we propose FFQs along with automated nutrient scoring as a powerful tool to assess dietary intake. Our automated pipeline may contribute to developing nutrient scoring further and to advance nutrition sciences. In which context fiber intake may be a powerful tool for weight management and dietetic treatments as proposed before [57, 58] remains to be investigated further. Interestingly, as the sample consists of young to middle-aged, healthy overweight individuals, the observed associations already indicate moderate effects already in a healthy population. Linear mixed model estimates suggested that every 10 grams of dietary fiber per day was associated with around 0.5 to 1.0 kg/m2 lower BMI. Dietary intake, in particular high fiber diets, have a large potential in preventing obesity-related states and comorbidities on a societal level [8, 59]. We suggest to increase educational efforts on fiber content of foods (as it is oftentimes not printed on food packaging, or available in experimental datasets, e.g. Food-pics database [60]) and to ameliorate policy making in the food sector (public and private) [61] and nutrition communication [62] to enhance fiber-rich diets and food items.
Overall, nutritional epidemiology will benefit from more advanced nutrient assessment and future studies revealing more insights on the impact of nutrient intake to provide more reliable and comparable evidence, which may inform public policy-making in the long-term.
Limitations
Firstly, due to inherent structure of the FFQ used, imprecision in the results might remain. For instance, only 53 food items are covered in the DEGS1, which means that a variety of different food items is not taken into account leading to gaps in data acquisition (e.g. legumes/ soy products,...). Another inaccuracy might stem from the fact that only certain FFQ questions are accompanied by a visual prompt, such as a picture of the portion size of a certain food (as provided by the Robert Koch Institute). As a result, the lack of a benchmark when estimating food intake might have led to deviations in the assessment.
Secondly, large ranges of values were present for some nutrients. This can be attributed to occasional over- or underreporting, however, we refrained from excluding datapoints to show truly reported values showcasing potential reporting biases. This shows the variability of self-reported FFQ data at the nutrient level and leaves room for defining data curation strategies depending on the research question, such as curating implausible data entries by consensus decision of different raters.
Thirdly, as the main sample is partly from a dietary intervention study, characteristics of this sample may reflect a selection bias in favour of omnivorous eaters with some awareness of the study goals to influence eating behaviour. Yet, we cross-validated data from the main sample with two other samples, albeit also from studies focusing on eating behaviour. Overall, self-reported dietary data is never blinded and neutral, since participants may reflect on social desirability and therefore report in a biased way. Yet, self-reported FFQ data in combination with recalls are a valid tool to assess nutrient intake [40] and relative nutrient intake can be compared.