Predicting obesity using machine learning techniques: Classification trees and random forests.

<p>Poster presented at the 1st Symposium on Big Data and Public Health, Fundação Getúlio Vargas (FGV), Rio de Janeiro, October, 2013.</p> <p>The present study investigates the prediction of obesity (BMI> 29.9 kg / m²) by waist (WC) and hip circumference (HC), and waist hip ratio (WHR) using a machine learning technique named Classification Tree. Data were collected from 400 college students (56.3% women) from 16 to 63 years old (M = 23:14, SD = 6:03). The sample was divided into two sets of each sex (training and test) for cross-validation. Seven trees were calculated in training group for each sex, using different numbers and combinations of predictors. The result shows that for women WC and WHR is the combination that produces the best prediction, since it has the lowest deviance (16.46) and misclassification (.035), and the higher pseudo R2 (.69). This model had an accuracy of 96.5% in the classification of obese and non-obese people in the training set, and presented a sensitivity of 60% and a specificity of 100% in the test set. For men CQ showed the best prediction with the lowest deviance (9:53) and misclassification (.022), and the higher pseudo R2 (.84). This model had an accuracy of 97.8% in the classification of obese and non-obese people, and presented a sensitivity of 75% and a specificity of 96.20% in the test set. Since classification trees are subject to high variability when applied to new data sets (e.g. the difference between the accuracy in the training set and in the test set) due to overfit. So, sample and variable bootstrap can provide a better approximation of the true accuracy of the prediction, avoiding overfiting. Random forest is a set of computational and statistical procedures that bootstrap both the sample and the variables, growing a large number of prediction trees, in order to discover which variable are important in the prediction, and what is the “true” accuracy (i.e. the accuracy that is expected using the given predictors over all possible sample). Applying the random forest analysis in the men’s training set using WC, HC and WHR as predictors led to an accuracy of 95.45% in the classification, with a specificity of 99% and a sensibility of 66%. The importance analysis, provided by the Random Forest technique, shows how much each variable contributes to the prediction of obesity. Waist circumference and hip circumference are, respectively, 21.19% and 30.96% more helpful than chance to successfully classify men’s obesity, while WHR is only 0.8%. Inclusion of the latter reduces, in average, 4% of the classification error, while the former reduces 5%. The result of the random forest analysis in the women’s training sample showed an accuracy of 95.58%, with a specificity of 99.99% but a sensibility of 43%. In terms of importance, WC is 11% more helpful than chance to successfully classify women’s obesity, while HC is 21% and WHR is only 1% better than random. Finally, we’ll introduce an innovative way to visualize prediction accuracy by applying the Fruchterman Reingold algorithm in the proximity matrix generated by the random forest analysis. This algorithm computes a network layout in which the length of edges (cases) depends on their absolute weight, which in our case is the proximity value for each pair of cases. Published at http://emap.fgv.br/BDPH2013/posters.html