Study on diabetes prediction based on discrete and continuous mean amplitude of glycemic excursions using machine learning methods

Abstract

Chronic Hyperglycemia and Acute Glucose Fluctuations are the two main factors that trigger complications in Diabetes Mellitus (DM). Continuous and sustainable observation of the two triggering factors is significant to be done to reduce the potential of cardiovascular problems in the future. This potential can be reduced by minimizing the occurrence of glycemic variability (GV). At present, observations on GV are based on the Mean Amplitude of Glycemic Excursion (MAGE), which is measured based on continuous blood glucose data from patients using particular devices. Continuous blood glucose uptake has technical constraints, especially in the provision of equipment. Thus, this study aims to calculate the value of MAGE based on discrete blood glucose observations from 43 volunteer patients. The resulting MAGE value is used to predict the diabetes status of volunteers. The hypothesis is that the discrete data is relevant to be used to predict the value of MAGE. Experiments were carried out by calculating MAGE values ​​from original discrete data (21 observations over three days) and continuous data obtained using Spline Interpolation. This study also utilizes the Machine Learning algorithm, especially k-Nearest Neighbor with Dynamic Time Wrapping (DTW) as a technique to measure the distance between time series data. From the classification test, discrete data and continuous data from the interpolation results show precisely the same accuracy value that is equal to 82.85%. Furthermore, there are variations in the MAGE value for each patient where the diabetes class has the most significant difference, followed by the pre-diabetes class, and the typical class. The conclusion that can be drawn is that patients in pre-diabetic and diabetic classes tend to have more varied blood glucose values ​​than patients from normal classes.