Digital single-image smartphone assessment of total body fat and abdominal fat using machine learning

Abstract

Background: Obesity is chronic health problem. Screening for the obesity phenotype is limited by the availability of practical methods. Methods: We determined the reproducibility and accuracy of an automated machine-learning method using smartphone camera-enabled capture and analysis of single, two-dimensional (2D) standing lateral digital images to estimate fat mass (FM) compared to dual X-ray absorptiometry (DXA) in females and males. We also report the first model to predict abdominal FM using 2D digital images. Results: Gender-specific 2D estimates of FM were significantly correlated (p < 0.001) with DXA FM values and not different (p > 0.05). Reproducibility of FM estimates was very high (R2 = 0.99) with high concordance (R2 = 0.99) and low absolute pure error (0.114 to 0.116 kg) and percent error (1.3 and 3%). Bland–Altman plots revealed no proportional bias with limits of agreement of 4.9 to -4.3 kg and 3.9 to -4.9 kg for females and males, respectively. A novel 2D model to estimate abdominal (lumbar 2–5) FM produced high correlations (R2 = 0.99) and concordance (R2 = 0.99) compared to DXA abdominal FM values. Conclusions: A smartphone camera trained with machine learning and automated processing of 2D lateral standing digital images is an objective and valid method to estimate FM and, with proof of concept, to determine abdominal FM. It can facilitate practical identification of the obesity phenotype in adults.

Peer Reviewed

Postprint (published version)