Artificial neural network model effectively estimates muscle and fat mass using simple demographic and anthropometric measures

Summary Background and aims Lean muscle and fat mass in the human body are important indicators of the risk of cardiovascular and metabolic diseases. Techniques such as dual-energy X-ray absorptiometry (DXA) accurately measure body composition, but they are costly and not easily accessible. Multiple linear regression (MLR) models have been developed to estimate body composition using simple demographic and anthropometric measures instead of expensive techniques, but MLR models do not explore nonlinear interactions between inputs. In this study, we developed simple demographic and anthropometric measure-driven artificial neural network (ANN) models that can estimate lean muscle and fat mass more effectively than MLR models. Methods We extracted the demographic, anthropometric, and body composition measures of 20,137 participants from the National Health and Nutrition Examination Survey conducted between 1999 and 2006. We included 13 demographic and anthropometric measures as inputs for the ANN models and divided the dataset into training and validation sets (70:30 ratio) to build and cross-validate the models that estimate lean muscle and fat mass, which were originally measured using DXA. This process was repeated 100 times by randomly dividing the training and validation sets to eliminate any effect of data division on model performance. We built additional models separately for each sex and ethnicity, older individuals, and people with underlying diseases. The coefficient of determination (R2) and standard error of estimate (SEE) were used to quantify the goodness of fit. Results The ANN models yielded high R2 values between 0.923 and 0.981. These values were significantly higher than those of the MLR models (p Conclusions We developed and validated an inexpensive but effective method for estimating body composition using easily obtainable demographic and anthropometric data.