Features development in machine learning model for non-invasive blood-glucose measurement

Abstract

Diabetes mellitus is the result of changes in the body caused by the decrease of insulin performance which is characterized by an increase in blood sugar level. Detection of blood sugar can be done with an invasive method or a non-invasive method. However, the non-invasive methods are considered better because they do not hurt the body and can check earlier, faster, and more accurately. A non-invasive blood glucose meter has been developed by a research team at IPB University. The output of the non-invasive tool is the intensity of the residual spectrum data, which will be related to the result of invasive measurement of blood glucose using some classification models, i.e. support vector machine and random forest. This research is aimed to compare the features development methods from the output of non-invasive tools and get the best features for modeling that can provide better predictions. The result of feature development shows that the best feature in the output of the non-invasive device is the trapezoidal area method at period because it has a higher accuracy value than the other four methods. The validation process shows that the random forest method has a higher accuracy value compared to the support vector machine.