Perceived age estimation from facial image and demographic data in young and middle-aged South Korean adults

Abstract

Biological age is an indicator of whether an individual is experiencing rapid, slowing, or normal aging. Perceived age is highly correlated with biological age, which reflects health appraisal and is often used as a clinical marker of aging. Perceived age has been reported as an important indicator of biological age and general health status, not only in older adult populations but also in young and middle-aged adults. However, there is a lack of objective methods for quantifying perceived age in these younger age groups. Thus, this study aimed to propose a novel perceived age estimation algorithm to meet the need for an objective method to predict perceived age. This cross-sectional study included 609 healthy men and 1388 healthy women (29.02–57.91 years, average 44.4 years) from 2017 to 2019 using data from the Korean Medicine Daejeon Citizen Cohort Study. The proposed algorithm comprised two steps. First, the initial predicted perceived age was estimated from facial images using a convolutional neural network (CNN) ensemble model. Then, the final perceived age was estimated using regression from the chronological age, sex, BMI, and initial predicted perceived age obtained in the first step. Better performance results were obtained by model averaging and model stacking generated from various basic regression models. The averaging models of Lasso, XGBoost, and CatBoost showed a mean absolute error of 2.2944, indicating that this algorithm can be used as a screening method for general health status in the population.