Predicting body fat percentage from anthropometric and laboratory measurements using artificial neural networks

Abstract

Purpose of the researchObesity is a major public health problem with rapidly growing prevalence and serious associated health risks. Characterized by excess body fat, the accurate measurement of obesity is a non-trivial question. Widely used indicators, such as the body mass index often poorly predict actual risk, but the direct measurement of body fat mass is complicated. The aim of the present research is to investigate how well can body fat percentage be predicted from easily measureable data: age, gender, weight, height, waist circumference and different laboratory results. For that end, linear regression, feedforward neural networks and support vector machines are applied on the data of a representative US health survey (n=862) using adult males. Optimal parameters are chosen and bootstrap validation is used to get realistic error estimates. ResultsNo methods can well predict the body fat percentage, but support vector machines slightly outperformed feedforward neural networks and linear regression (root mean square error 0.0988±0.00288, 0.108±0.00928 and 0.107±0.012 respectively). ConclusionEven this best performance means that soft computing methods had an R2 of 44%, but this slight advantage is balanced by the fact that regression models are clinically interpretable.