Implementation of neural networks for predicting metabolic syndrome: a multinational data study

Background: Metabolic syndrome (MS) is composed of abnormalities such as central obesity, insulin resistance, hypertension and dyslipidemia. Objective: To implement a neural network to predict MS from cholesterol, triglycerides, high density lipoproteins (HDL), obesity and hypertension. Material and methods: Analytical and cross-sectional study with 1878 patients from databases in Venezuela, Thailand and Indonesia. Variables such as MS, hypertension, obesity, HDL, triglycerides and total cholesterol were included. Multilayer perceptron neural networks were used, evaluated with classification tables, area under the curve (AUC) and performance metrics (sensitivity, specificity, positive and negative predictive values). Results: The neural network showed a high capacity to predict MS, with a low percentage of incorrect predictions both in the training set (15.80%) and in the test set (18.20%). In training, the overall accuracy was 84.20%, with higher accuracy for cases without MS (88.30%) than for cases with MS (77.80%). In testing, the overall accuracy was 81.80%, also with higher accuracy for cases without MS (86.60%) than for cases with MS (74.80%). The AUC was 0.911, indicating an outstanding predictive capacity. Regarding the model performance, the sensitivity was 81.25% in training and 79.26% in testing, while the specificity reached 85.92% and 83.33%, respectively. The positive predictive value was 77.80% in training and 74.78% in testing, and the negative predictive value was 88.30% and 86.57%, respectively. Conclusions: The multilayer perceptron neural network is an effective tool to predict MS, showing an outstanding predictive capacity.