Neural networks to predict Body Mass Index according to height and waist circumference

Introduction: The Body Mass Index (BMI) is an indicator of risk of suffering from diseases associated with excess body fat. Objective: To evaluate the neural network as a predictor of BMI according to height and waist circumference. Methods: Analytical, and cross-sectional study of 2004 Venezuelan adults belonging to the Prevalence of Metabolic Syndrome study of the City of Maracaibo. The variables were BMI, height and waist circumference. Multilayer perceptron-type neural networks were used, evaluated using scatter-plots, receiver operating characteristic curve and classification tables. Results: For quantitative BMI, the relative error was 0.191 and 0.180 in training and testing, respectively. For BMI categories, the percentage of incorrect predictions in training and testing were 25.50% and 20.80%, respectively. The neural network developed to quantitatively predict the BMI from height and waist circumference had an R² coefficient of 0.812 and qualitatively, an area under the curve of 0.968, 0.919, 0.844 and 0.950 for low weight, normal weight, overweight, and obesity, respectively. The neural network to predict BMI categories had percentages of correct predictions of 37.50%, 74%, 79.80% and 84.40% for underweight, normal weight, overweight, and obesity, respectively. Conclusions: The use of multilayer perceptron-type neural networks is efficient to predict the BMI quantitatively and qualitatively from height and waist circumference.