Predicting Metabolic Syndrome with Neural Networks

Abstract

Metabolic syndrome (MetS) is a condition that predisposes individuals to acquire diabetes and cardiovascular disease. The prevalence of MetS among young Mexicans (17-24 years old) is high (14.6%), and is an important risk factor to develop more serious impairments. Thus, it is crucial to detect MetS in young as they could be alerted to modify their life habits to revert or delay further health complications. One barrier to identify the MetS in large young populations is the high costs and complex logistics involved. The aim of this study was to build a tool to predict MetS in young Mexicans using noninvasive data, such as anthropometrics (waist circumference, height, weight, or body mass index), family-inherited background (diseases of parents), or life and eating habits, but excluding laboratory data such as blood glucose, cholesterol or triglycerides that implies to withdraw blood samples incurring in costs and nuisance for people. We evaluated 826 Mexican undergraduate students collecting both, invasive and noninvasive data and determined whether each one bears or not MetS. Then we build neural networks (NN) using only noninvasive data, but with the output class known (with and without MetS). Noninvasive data were classified into six groups, and arranged into ten sets as input layer. We generated 10 NN’s taking 70% of record as training set, and the 30% as validation records. We used the positive predictive value (PPV) as classifier efficiency of the NN’s. The PPV of the NN’s vary from 38.2% to 45.4%, the last from a NN including those anthropometrics variables (sex, waist circumference, height, weight, body mass index) as input variables, and the hours per week of physical exercise. Comparing percentage of true positive (students with MetS) detected with the NN vs. a random selection, i.e., 45.4% of PPV vs. 14.6% of MetS prevalence in the objective population, it is expected to improve the MetS identification by three fold.