Glycemic-aware metrics and oversampling techniques for predicting blood glucose levels using machine learning

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Endocrinology

Background:

• Accurate short-term blood glucose level prediction is crucial for managing Type 1 Diabetes (T1D).
• Effective artificial pancreas technology relies on reliable glucose forecasting for timely alerts, such as hypoglycemia alarms.
• Patients with T1D spend varying amounts of time in different glycemic ranges, posing challenges for predictive modeling.

Purpose of the Study:

• To investigate machine learning techniques for short-term blood glucose level prediction in T1D patients.
• To determine the optimal machine learning approach considering regression metrics and data preprocessing for different glycemic ranges.
• To inform the development of more accurate artificial pancreas systems.

Main Methods:

• Evaluation of various machine learning regression models for blood glucose prediction.
• Analysis of preprocessing techniques to address imbalanced glycemic range data.
• Comparison of model performance across normal, hyperglycemic, and hypoglycemic glycemic subranges using standard benchmark data.

Main Results:

• Different machine learning model and preprocessing combinations yield varying accuracies depending on the glycemic subrange.
• A linear Support Vector Regression model with polynomial features excels in predicting normal and hyperglycemic glucose levels.
• A Multilayer Perceptron trained on oversampled data is optimal for predicting hypoglycemic glucose levels.

Conclusions:

• The choice of machine learning technique for blood glucose prediction is contingent upon the specific glycemic range and the desired alert type.
• Tailored machine learning approaches are necessary for accurate glucose forecasting across diverse glycemic states in T1D management.
• This research provides insights for optimizing artificial pancreas algorithms for improved patient safety and diabetes control.