Testing the Utility of a Data-Driven Approach for Assessing BMI from Face Images

Area of Science:

• Computer Vision
• Human Physiology
• Social Psychology

Background:

• Facial cues are implicated in social interactions and may signal body mass index (BMI).
• Previous methods for quantifying facial BMI cues used limited facial proportions with low predictive accuracy.

Purpose of the Study:

• To develop and evaluate a data-driven approach for predicting body mass index (BMI) from facial characteristics.
• To compare the predictive power of new models based on principal components (PCs) of facial shape and color with traditional facial proportion metrics.

Main Methods:

• Employed a data-driven strategy using principal components (PCs) derived from objective facial shape and color features in images.
• Built statistical models utilizing these PCs to predict BMI.
• Compared the performance of PC-based models against models using established facial proportions (perimeter-to-area ratio, width-to-height ratio, cheek-to-jaw width).

Main Results:

• Models based on 2D shape-only PCs, color-only PCs, and combined 2D shape and color PCs demonstrated significantly superior BMI prediction compared to traditional facial proportion models.
• A non-linear principal component model incorporating both 2D facial shape and color emerged as the most effective predictor of BMI.

Conclusions:

• A data-driven, "bottom-up" approach utilizing facial shape and color principal components offers substantially improved accuracy in assessing body mass index (BMI) from face images.
• This methodology provides a more robust and quantifiable method for understanding facial indicators of adiposity in social contexts.