Three-dimensional human facial morphologies as robust aging markers

Abstract

Aging is associated with many complex diseases. Reliable prediction of the aging process is important for assessing the risks of aging-associated diseases. However, despite intense research, so far there is no reliable aging marker. Here we addressed this problem by examining whether human 3D facial imaging features could be used as reliable aging markers. We collected > 300 3D human facial images and blood profiles well-distributed across ages of 17 to 77 years. By analyzing the morphological profiles, we generated the first comprehensive map of the aging human facial phenome. We identified quantitative facial features, such as eye slopes, highly associated with age. We constructed a robust age predictor and found that on average people of the same chronological age differ by ± 6 years in facial age, with the deviations increasing after age 40. Using this predictor, we identified slow and fast agers that are significantly supported by levels of health indicators. Despite a close relationship between facial morphological features and health indicators in the blood, facial features are more reliable aging biomarkers than blood profiles and can better reflect the general health status than chronological age.

Figure 1

Visualization of facial aging. (A) The female and male average profiles of five age groups from 17 to 77 years old. N indicates the number of subjects in each age group. (B-D) Synthesized female and male average profiles between −2 and +2 SD of loading values of age-correlated PLS component 1 (B), component 2 (C) and combined components 1 and 2 (D). (E-G) Heat map of 3D effects displaying loading values of age-correlated PLS component 1 (E), component 2 (F) and combined components 1 and 2 (G) shown on female and male faces. The loading values were multiplied by 10 000. Red and blue denote, respectively, higher and lower values along x-, y- and z-axes.

Figure 2

Aging-related facial morphological phenotypes. (A) The 17 landmarks used to align all faces. (B) Clustering of all quantified facial features, blood serum indicators, blood cell indicators, body indexes and PLS components in females and males. (C) The correlation network of female facial features and PLS components. (D) The network of male facial features and PLS components. Node size is proportional to the correlation between the feature or component and age. (E) Distribution of correlations between chronological age and each quantitative feature in females and males.

Figure 3

Prediction of physiological age, slow and fast agers based on 3D facial images. (A) Correlation of age predicted by facial vertices-based SVR predictors with the actual age of the subjects, and the correlation between SVR- and PLSR-predicted ages. Predictors are trained separately in females and males using all but one sample to predict the age of the left-out sample. The predictors are generated based on ages accurate to the day. (B) The deviations and correlations between predicted and real ages reach saturation levels with > ∼40% of the samples. Curves of each predictor show MAD and PCC between predicted age and chronological age when 10%, 20%,..., or 100% of the data was used to build the predictors. (C) MAD between the chronological ages and predicted ages in each age group. Error bars denote SD. (D) The average profiles of the predicted fast agers, slow agers and well-predicted female and male subjects in age groups older than 40 years. The classification is based on the age difference > 6 years between predicted age and chronological age. N indicates the number of subjects in each class of each age group. (E) Levels of the most age-correlated health indicators in predicted fast agers, slow agers and well-predicted subjects. The classification is the same as in D. ^*, ^**, ^*** and ^**** denote unpaired one-sided Student's t-test P < 0.1, 0.05, 0.001 and 0.0001, respectively. RCC stands for Spearman's rank correlation coefficient between each sample class (rank 1, 2 and 3 for predicted fast agers, well-predicted subjects and slow-agers, respectively) and indicator level. FDR stands for false discovery rate, which is the fraction of times among 1 000 sample label permutations that give RCC ≥ the real RCC. Error bars denote SD. (F) The joint FDR for all age-associated blood indicators' RCC to slow- and fast-ager classification in females and males in each age group. FDR is calculated as the fraction of times among 1 000 sample label permutations that have larger than or equal to the real number of indicators whose absolute RCC are greater than the defined cut-off (female 0.4 and male 0.2). Support denotes the cases of positive RCCs for Cluster 1 indicators and negative RCCs in Cluster 6 indicators; oppose denotes the opposite.

Figure 4

Heat map of 3D effects showing loading values of PLS component 1 correlated with CHO, LDL-C, HDL-C or ALB level on female and male faces. Loading values were multiplied by 10 000. Red and blue denote, respectively, higher and lower values along x-, y- and z-axes.