Body Mass Index Classification based on Facial Features Using Machine Learning Algorithms for Utilizing in Telemedicine

Abstract

Objectives: Given the benefits of controlling Body mass index (BMI) on the quality of life, BMI classification based on facial features can be used for developing telemedicine systems and eliminate the limitations of existing measuring tools especially for paralyzed people, that enable physicians to help people online when faced with situations like the COVID-19 pandemic.
 Methods: In this study, new features and some previous-work features were extracted from face photos of white, black and Asian people, ages 18 to 81, with normal and overweight BMI. Faces were evaluated in three different steps. First, all faces were considered as one group. Second, they were divided into elliptical, round and square shape groups and third, they were separated based on gender. Then for each step, the performances of Random Forest (RF) and Support Vector Machine (SVM) were evaluated with all of the facial features and with selected features based on Pearson correlation coefficient. Matlab R2015b was used for implementation.
 Results: The results revealed that features with higher correlation improved the accuracy of both algorithms. RF best performance using highly correlated features for 97 women and 92 men was in women and square-face groups (91.75% and 87.30% respectively), and SVM best performance was in women group (94.84%), square-face and round-face groups (84.12% and 84% respectively).
 Conclusion: Accuracy of BMI classification based on facial features can be improved by categorizing faces into shapes and gender, and selecting appropriate features. The findings can be used for performance enhancement of telemedicine applications, especially for helping the differently-abled.