Machine learning-based equations for improved body composition estimation in Indian adults

Bioelectrical impedance analysis (BIA) is commonly used as a lower-cost measurement of body composition as compared to dual-energy X-ray absorptiometry (DXA) in large-scale epidemiological studies. However, existing equations for body composition based on BIA measures may not generalize well to all populations. We combined BIA measurements (TANITA BC-418) with skinfold thickness, body circumferences, and grip strength to develop equations to predict six DXA-measured body composition parameters in a cohort of Indian adults using machine learning techniques. The participants were split into training (80%, 1297 males and 1133 females) and testing (20%, 318 males and 289 females) data to develop and validate the performance of equations for total body fat mass (kg), total body lean mass (kg), total body fat percentage (%), trunk fat percentage (%), L1-L4 fat percentage (%), and total appendicular lean mass (kg), separately for males and females. Our novel equations outperformed existing equations for each of these body composition parameters. For example, the mean absolute error for total body fat mass was 1.808 kg for males and 2.054 kg for females using the TANITA’s built-in estimation algorithm, 2.105 kg for males and 2.995 kg for females using Durnin-Womersley equations, and 0.935 kg for males and 0.976 kg for females using our novel equations. Our findings demonstrate that supplementing body composition estimates from BIA devices with simple anthropometric measures can greatly improve the validity of BIA-measured body composition in South Asians. This approach could be extended to other BIA devices and populations to improve the performance of BIA devices. Our equations are made available for use by other researchers.Author summary: Bioelectrical Impedance Analysis, or BIA, is a popular technology for estimating the amount of fat and lean mass in the body. BIA scales are low-cost and portable, meaning they are commonly used in clinical weight monitoring and in large-scale research studies in place of more accurate but more expensive whole-body DXA scans. However, BIA technology was developed mostly in Western populations and relies on assumptions about body shape, size and composition that may not apply to all populations globally. We investigated whether addition of simple body measurements such as waist circumference and skinfold thickness could improve the accuracy of BIA measurements in an Indian population. We applied machine learning methods to a large dataset from India with simple body measurements and whole-body DXA scans to find the most informative combinations of measures. Our resulting models were able to estimate body fat and lean mass in Indians over twice as accurately as BIA scales alone, suggesting that this could be a valuable approach to improve body composition estimation from other BIA devices and in other under-represented populations. Our equations are available via a web application for researchers to use.