Predicting Hyperglycemia Using NIR Spectrum of Spent Fluid in Hemodialysis Patients

Abstract

We aimed to assess the near infrared spectroscopy as a method for non-invasive on-line detection of hyperglycemia from spent hemodialysis effluent. We used partial least squares regression and several machine learning algorithms: random forest (RF), logistic regression, K-nearest neighbor (KNN), support vector machine (SVM), decision tree classifier, and Gaussian naive Bayes (NB) to classify normoglycemia from hyperglycemia. These classifier methods were used on the same dataset and evaluated by the area under the curve. The serum glucose levels were presented in the form of a binomial variable, where 0 indicated a glucose level within reference range and 1 a glucose level beyond the normal limit. For this reason, the methods of machine learning were applied as more specific methods of classification. RF and SVM have shown the best classification accuracy in predicting hyperglycemia, while decision tree and NB showed average accuracy.