The present study was approved by the Ethics Committees at Mukogawa Women’s University (No. 12-30). All experiments were performed in accordance with relevant guidelines and regulations.
Subjects
The participants were 115 female university students with an average age of 20.6 ± 0.48 (range 19.3–23.2) years old. We recruited participants from School of Human Environmental Sciences, Mukogawa Women’s University (Nishinomiya, Japan) which is located in an urban area, to eliminate variations in educational status and residential area. The data were collected from May 2013 and May 2014. A written informed consent form was distributed to and signed by all participants. Informed consent was approved by the Ethics Committees at Mukogawa Women’s University (No. 12-30). No relatives were included among the subjects.
Data recording
3D facial shapes
For the morphological characteristics of the soft tissue faces, 3D facial images were noninvasively recorded for each participant, using a 3D image-capturing device (3dMDcranial Systems; 3dMD LLC, Atlanta, GA, USA). Each participant sat upright in a chair with no back support. While recording, each participant was asked to keep the teeth in light intercuspation and the lips in repose.
Holistic dietary conditions
In the present study, we examined the participants’ holistic dietary conditions by combining information from the nutrition contents in daily food intake and eating behaviors.
Nutrition contents in daily food intake
Daily food intake was recorded for each participant using a food frequency questionnaire on food groups75, which comprised items from 29 food groups and 10 types of cooking, and calculated the average intake per week of each food or food groups in commonly used units or portion sizes. In detail, the participants were instructed to fill out questionnaires in which respondents described their consumption of 29 food groups and 10 types of cooking with regard to the amount (“A little”, “Normal”, and “Plenty”) and frequency (how many times a week). Regarding the amount of foods, visual aids (illustration and food models) were used to avoid bias concerning the actual portions. To estimate the intake of sugar, salt, and oils, questions about cooking methods (e.g. boiled, deep-fried, fried, and soup) were also collected. The survey time was 30 min for each participant. After the participants completed answering the questionnaire, three nationally registered dieticians (MK, NT, and MT) interviewed each participant to review the completed questionnaire. The nutrient intake was then estimated using an Excel software program add-in (Excel Eikun Ver. 6.0; Kenpakusha, Tokyo, Japan) based on the Revised Standard Tables of Food Composition in Japan76 as follows:
Nutrient intake = Portion size per food group (gram [g]) × Food amount category value × Number of times consumed in a week/7 days × Amount of each nutrient per gram of food group in the average composition table76.
In this formula, the food amount category value was defined as follows: “No” = 0, “A little” = 0.5, “Normal” = 1, “Plenty” = 1.5. To estimate salt consumption, a question comparing the taste of each household’s condiments to that found in restaurants was also included, and the salt amount was adjusted by multiplying by 1.15 (equal to restaurant food) and 1.3 (saltier than restaurant food). Validation of this estimation approach was confirmed in a previous study75.
Estimated nutrients included the following 53 nutrient items: energy [E] (kcal), grain (%E), water (g), protein (g), protein (%E), animal protein ratio (% total protein), lipid (g), lipid (%E), FA total amount (g), saturated FA (g), cholesterol (mg), monounsaturated FA (g), polyunsaturated FA (g), n-3 polyunsaturated FA (g), n-6 polyunsaturated FA (g), carbohydrate (g), carbohydrate (%E), deep-yellow-vegetable ratio, dietary fiber water solubility (g), dietary fiber insolubility (g), dietary fiber total amount (g), retinol (μg), alpha carotene (μg), beta carotene (μg), cryptoxanthin (μg), beta carotene equivalent (μg), retinol equivalent (μg), vitamin D (μg), alpha tocopherol (mg), beta tocopherol (mg), gamma tocopherol (mg), delta tocopherol (mg), tocopherol equivalent (mg), vitamin K (μg), vitamin B1 (mg), vitamin B2 (mg), Niacin (mg), vitamin B6 (mg), vitamin B12 (μg), folic acid (μg), pantothenic acid (mg), vitamin C (mg), mineral (g), sodium (mg), salt (g), potassium (mg), calcium (mg), magnesium (mg), phosphorus (mg), iron (mg), zinc (mg), copper (mg), and manganese (mg). In addition, six ratios (animal oil ratio, vegetable oil ratio, fish oil ratio, saturated FA ratio, monounsaturated FA ratio, polyunsaturated FA ratio, and n-6/n-3 FA ratio) were calculated based on the Revised Standard Tables of Food Composition in Japan76.
Eating behaviors
Eating behaviors or dietary habit was also recorded using the questionnaire, based on the guidelines for obesity issued by the Japan Society for the Study of Obesity77. The questionnaire consisted of 55 questions consisting of 7 major categories (i.e. recognition of weight and constitution, external eating behavior, emotional eating behavior, sense of hunger, eating style, food preference, regularity of eating habits). Questions are shown in Supplementary Table S5. All items were rated on a four-point scale ranging from 1 (i.e. “seldom”) to 4 (i.e. “very often”).
Physical body composition
Four body composition parameters (body weight, total skeletal muscle mass, total body fat mass, and BMI) were measured using a multifrequency bioelectrical impedance analysis device (InBody 720; Biospace, Inc., Tokyo, Japan)78. Further, body fat percentage (fat mass/weight), fat-free mass index (FFMI, fat-free mass/height2) and fat mass index (FMI; fat mass/height2)33 was determined from measured body compositions.
The samples’ mean weight and BMI were 50.97 ± 4.92 (range 37.64–63.38) kg and 20.29 ± 1.67 (range 16.94–25.40) kg/m2, respectively. The reported average weight and BMI in Japanese women (18–29 years old) is 50.0 kg and 20.0 lg/m2, respectively76, indicating that the mean of the sampled populations was almost equal to the national values. Based on the BMI classification of ‘normal weight’ as 18.5–24.9 kg/m279, the proportions of participants categorized as underweight, normal weight, and pre-obesity were 15%, 86%, and 1%, respectively. These data indicate that the sample was slightly thinner but otherwise representative of young Japanese women.
Relationships between body compositions and nutritional intakes or eating behaviors were shown in Supplementary Text S5.
Analyses
Data were analyzed via two methods: (1) direct comparisons between 3D facial shapes and eating behaviors and (2) indirect comparisons between these items. (1) was achieved by a multivariate analysis of covariance (MANCOVA) and canonical variate (CV) analysis, while (2) was conducted by clustering participants based on their eating behaviors and visualizing the averaged 3D faces of each cluster. All analyses were conducted using the statistical software program included with R (http://www.r-project.org/) and MATLAB (MathWorks, Ltd., Natick, MA, USA).
Direct comparisons between 3D facial shapes and the dietary condition
We directly compared the 3D facial shapes and dietary conditions (i.e. nutritional conditions and eating behavior) in three steps: a geometric morphometrics analysis of 3D faces, a summary of the nutritional conditions, and an examination of the relationships between facial shapes and eating behaviors. The details are described below.
Geometric morphometrics analyses of 3D faces
A 3D coordinate system identical to that employed in our previous study (Supplementary Fig. S1; Informed consent was obtained to publish this image in an online open-access publication.)34 was used in the current study. In short, the sagittal plane was defined by exocanthions and endocanthions, and the axial plane was defined by exocanthions, the porion, and the subnasale. The nasion was set as the origin.
For each facial surface, fitting of high-resolution template meshes34,80 was performed using a commercial software program (HBM-Rugle; Medic Engineering Co., Kyoto, Japan) based on the landmarks assigned to each 3D image (Supplementary Table S10). This method automatically generated a homogeneous model that consisted of 6017 Cartesian semi-landmarks on the wire mesh for each model. This technique permits the extraction of relevant surface anatomy from face data while removing non-relevant data, yielding 3D surface data that provide enough detail to facilitate a quantitative assessment while minimizing file sizes to still be sufficient to represent the facial shape (Supplementary Fig. S2; Informed consent was obtained to publish this image in an online open-access publication). The 6017 Cartesian semi-landmark coordinates were analyzed by geometric morphometrics. In brief, after separating the shape from the overall size, position, and orientation of the landmark configurations, the resulting Procrustes shape coordinates were used for subsequent statistical analyses, where the overall size was calculated as the centroid size (CS). To examine the variance in facial shapes of the subjects, we performed a principal component analysis (PCA) for the 6017 coordinates of the aforementioned surface model. The dimensionality of the shape principal components (sPCs) was determined to include sPCs with eigenvalues greater than 1, namely the Kaiser criterion. The facial morphospace (new dimensional space representing facial shape) consisted of these significant sPCs and used in the following process:
Summarization of the nutritional conditions
To examine the variance in nutritional conditions, the estimated nutrient items included in the daily food intake were dimensionally reduced using the PCA. The dimensionality of the nutritional principal components (nPCs) was also determined using the Kaiser criterion, as above. For both sPCs and nPCs, if the cumulative contribution rate based on the Kaiser criteria was greater than 90% variations our of all the variations, then we used the 90% cumulative contribution rate of total variance as the cutoff to determine the number of PCs.
Relationships between facial shapes and dietary conditions
To examine the direct relationship between facial shapes and dietary conditions, sPCs that corresponded to nPCs, eating behavior variables, and physical conditions (body weight, total skeletal muscle mass, total body fat mass, and BMI) were examined with a stepwise regression analysis, as follows:
$$V = w_{1} times {text{sPC1}} + w_{2} times {text{sPC2}} + cdots + w_{k} times {text{sPCk}}$$
where k indicates the number of the sPCs; w1,2,…k indicates the coefficient values of the regression analysis, and V is each nPC, eating behavior score, and physical condition (dependent variables). The inclusion criteria for the stepwise analysis were set as items with p-values of 0.05. To control multiple testing, significant dependent variables and selected sPCs were entered into the MANCOVA. The regression coefficient was represented using a heat map.
Further, a CV analysis was conducted to examine if correlation is significant between the facial characteristics (sPCs) and the dietary intake amount of nutritional contents when showing maximum correlation between them. Faces showing maximum correlations were visualized and corresponding nutrition contents were determined.
Indirect comparisons between 3D facial shapes and eating behaviors using clustering of the subject groups based on holistic eating behaviors
To visualize the 3D faces corresponding to the holistic nutritional patterns, we subcategorized the participant samples into k groups using the k-means clustering method, when the input was a combined vector of nPCs and the sum scores of the eating behavior. In this instance, k-means clustering is a method of dividing n observations (n = the number of the participants) into k groups, with each observation belonging to the cluster with the closest mean (cluster centroid) that serves as the cluster prototype. k-means clustering starts from randomized initial assignments of the observation and calculation of the cluster centroids. The method then continues iterations to reassign observations and calculate the centroids, so that movement minimizes the variance (squared Euclidean distance) within a cluster. Once no movement of the centroids is observed, the iterations are considered finished. The number of clusters was determined by the elbow method. Each cluster was represented by the cluster centroid of the categorized group (hereafter referred to as the “code”). To visualize the 3D facial shapes for each code, the 3D-averaged and accentuated faces (Supplementary Text S6) for each code were calculated using the aforementioned fitted meshes on the 3D face.
Calculation of SShD as an index for feminine or masculine facial shapes
The SShD of the individual face was measured by projection of the individual facial configurations onto a male–female axis that were determined a previous study30. In short, the male–female axis was calculated from 272 Turkish and Japanese men and women and defined the vector between the average facial configurations of males and females in the facial morphospace using the following equation31,81
$$mathrm{SShD}left(overrightarrow{{F}_{i}}right)=frac{( overrightarrow{{F}_{i}}cdot overrightarrow{{F}_{(m-f)}})}{{|overrightarrow{{F}_{(m-f)}}|}^{2}}$$
where (overrightarrow{{F}_{i}}) is the vector in the facial morphospace corresponding to an individual face i, and (overrightarrow{{F}_{(m-f)}}) is the vector between male and female facial configurations (male minus female previously determined30. If SShD < − 1, the face is hyperfeminine, and if SShD > 1, the face is hypermasculine. We used mixed samples of Japanese and Turkish individuals to calculate the SShD because we aimed to extract vectors to represent sexual dimorphism with population affinity.
Measures of SShD were mathematically decomposed to allometric and non-allometric components. That is, variations in SShD due to an individual’s size (allometric) and variations that were independent of size (non-allometric) were examined in the overall variation in SShD in each population group using a multivariate regression analysis. CS was used as a measure of an individual’s size.
Inter-code subject group comparisons
The analysis of variance (ANOVA) and the Tukey–Kramer post-hoc tests were conducted to examine whether or not any nutrient content variable was specific to each code (i.e. pattern) and whether or not any eating behavior major scales were specific to each code.
In addition, the faces were examined by using a total of 17 variables were extracted as inter-landmark distances of the face (|Ac-Ac|, |Zy-Zy|, |Ch-Ch|, |Go-Go|, |Go-Go”https://www.nature.com/”Zy-Zy|, |N-En|, |N-Sn|, |N-Zy|, |N-Prn|, |N-Ls|, |Sto-Gn|, |Ls-Li|, |Gla-Sn”https://www.nature.com/”Sn-Gn|, |N-Gn”https://www.nature.com/”Zy-Zy|, |Gla-ls”https://www.nature.com/”Zy-Zy|, |Gla-ls”https://www.nature.com/”Zy-Zy|, |Ps-Pi|; For definition, please see Supplementary Fig. S1 and Supplementary Table S7). Facial size differences between individuals were also standardized by normalizing the values of all linear variables to the distance between right and left exocanthions. ANOVA and the Tukey–Kramer post-hoc tests were conducted to these facial variables in addition to CS, SShD, allometric SShD, non-allometric SShD to examine if there were any differences between the codes.
Further, adjunctive data for health history, which were assumed as relevant to the dietary conditions (i.e. physical activity level, duration of physical exercises, birth weight, 2-year change in weight, sleep duration, and menarche age) were also collected from each participant for inter-code subject group comparisons. Physical body composition (e.g. weight, total fat mass) and adjunctive data for health history were also compared between codes.
p values were adjusted for multiple comparisons using the Benjamini–Hochberg method in order to control the false discovery rate82. Adjusted values of p < 0.05 were significant for all statistical tests.
