Evaluation of Multiple Machine Learning Algorithms and Glycemic Variability Indices for Predictions of Glucose Levels in Type 2 Diabetes Mellitus Based on Continous Glucose Monitoring and Activity Data

Glycemic events and glycemic variability play a key role in the onset of complications in Diabetes Mellitus type 2 (DM2). Predicting these can be an important tool in self-management, and so preventing complications relating to DM2. Using Machine Learning, predictive models can be made. Continuous glucose monitoring (CGM) and activity data from the DIALECT (Diabetes and Lifestyle Cohort Twente) dataset were used. Features based on time, CGM, activity data, and clinical information were extracted. Various glycemic variability features were included, such as Mean Amplitude of Glycemic Excursion (MAGE), J-index, Time in Range (TIR), and High and Low Blood Glucose Index (HBGI and LBGI). Models for predicting the next glucose level were made using the following machine learning algorithms: Linear Regression (LR), Decision Tree (DT), Random Forest (RF), and XGBoost (Extreme Gradient Boosting). These models were compared with regard to prediction accuracy with the following metrics: Clarke Error Grid Analysis, root mean square error (RMSE), and mean absolute error (MAE). First, personal models were made based on data from three patients with different variability. Variability was assessed by calculating the average daily risk range (ADRR). Second, population models were made based on the full data. All models were compared with a baseline: assuming the previous glucose value as the current value. Results show LR performed best with personal models and RF performed best with population models. Feature importance showed that the most important feature categories in most models were previous glucose measurements and basic glucose calculations. Glycemic variability features were not very high ranked, except in the personal LR models, and SD features in the population models.