Comparing hyperparameter optimized support vector machine, multi-layer perceptron and bagging classifiers for diabetes mellitus prediction

Abstract

Diabetes Mellitus (DM) is a chronic metabolic disorder that affects the way body processes blood glucose levels. Within the medical field, Machine Learning (ML) has significant potential for accurately forecasting and diagnosing a range of chronic conditions. If an accurate prognosis is achieved early, the risk to health and intensity of DM can be significantly mitigated. In this study, a robust methodology for DM prognosis was proposed, which included anomaly replacement, data normalization, feature extraction, and K-fold cross-validation. Three machine learning methods, Support Vector Machine, Multilayer Perceptron and Bagging, were employed to predict Diabetes Mellitus using the National Health and Nutritional Examination Survey (NHANES) 2011-2012 dataset. Accuracy, AUC and Recall were chosen as the evaluation metrics and subsequently optimized during hyperparameter tweaking. From all the comprehensive tests, Bagging outperformed the other two models with an Accuracy of 96.67, AUC score of 99.2 and Recall of 97.0. The proposed methodology surpasses other approaches for forecasting DM.