Machine Learning for the Classification of Obesity from Dietary and Physical Activity Patterns

Abstract

Conventional epidemiological analyses in health-related research have been successful in identifying individual risk factors for adverse health outcomes, e.g. cigarettes’ effect on lung cancer. However, for conditions that are multifactorial or for which multiple variables interact to affect risk, these approaches have been less successful. Machine learning approaches such as classifiers can improve risk prediction due to their ability to empirically detect patterns of variables that are “diagnostic” of a particular outcome, over the conventional approach of examining isolated, statistically independent relationships that are specified a priori. This chapter presents a proof-of-concept using several classifiers (discriminant analysis, support vector machines (SVM), and neural nets) to classify obesity from 18 dietary and physical activity variables. Random subsampling cross-validation was used to measure prediction accuracy. Classifiers outperformed logistic regressions: quadratic discriminant analysis (QDA) correctly classified 59% of cases versus logistic regression’s 55% using original, unbalanced data; and radial-basis SVM classified nearly 61% of cases using balanced data, versus logistic regression’s 59% prediction accuracy. Moreover, radial SVM predicted both categories (obese and non-obese) above chance simultaneously, while some other methods achieved above-chance prediction accuracy for only one category, usually to the detriment of the other. These findings show that obesity can be more accurately classified by a combination or pattern of dietary and physical activity behaviors, than by individual variables alone. Classifiers have the potential to inform more effective nutritional guidelines and treatments for obesity. More generally, machine learning methods can improve risk prediction for health outcomes over conventional epidemiological approaches.