Advances in digital anthropometric body composition assessment: neural network algorithm prediction of appendicular lean mass

• Published in: European journal of clinical nutrition Vol. 78; no. 5; pp. 452 - 454
• Main Authors: Marazzato, Frederic, McCarthy, Cassidy, Field, Ryan H., Nguyen, Han, Nguyen, Thao, Shepherd, John A., Tinsley, Grant M., Heymsfield, Steven B.
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 01.05.2024; Nature Publishing Group
• Subjects: 13/56; 631/114/2397; 692/308/53; 82/56; 9/10; Absorptiometry, Photon - methods; Adolescent; Adult; Aged; Algorithms; Anthropometry; Anthropometry - methods; Body Composition; Brief Communication; Clinical Nutrition; Datasets; Dual energy X-ray absorptiometry; Epidemiology; Ethnic factors; Female; Humans; Internal Medicine; Machine learning; Male; Mathematical models; Medicine; Medicine & Public Health; Metabolic Diseases; Middle Aged; Minority & ethnic groups; Neural networks; Neural Networks, Computer; Optical scanners; Predictions; Public Health; Statistical analysis; Statistical methods; Statistical models; Young Adult
• ISSN: 0954-3007; 1476-5640; 1476-5640
• DOI: 10.1038/s41430-023-01396-3

Currently available anthropometric body composition prediction equations were often developed on small participant samples, included only several measured predictor variables, or were prepared using conventional statistical regression methods. Machine learning approaches are increasingly publicly available and have key advantages over statistical modeling methods when developing prediction algorithms on large datasets with multiple complex covariates. This study aimed to test the feasibility of predicting DXA-measured appendicular lean mass (ALM) with a neural network (NN) algorithm developed on a sample of 576 participants using 10 demographic (sex, age, 7 ethnic groupings) and 43 anthropometric dimensions generated with a 3D optical scanner. NN-predicted and measured ALM were highly correlated (n = 116; R 2 , 0.95, p < 0.001, non-significant bias) with small mean, absolute, and root-mean square errors (X ± SD, −0.17 ± 1.64 kg and 1.28 ± 1.04 kg; 1.64). These observations demonstrate the application of NN body composition prediction algorithms to rapidly emerging large and complex digital anthropometric datasets. Clinical Trial Registration: NCT03637855, NCT05217524, NCT03771417, and NCT03706612.