Deep Learning and Machine Learning Models for Predicting Day-Time Glucose in Different Ranges in Patients with Type 1 Diabetes

Background and Aim: Accurate prediction of glucose levels (episodes of hyperglycemia and hypoglycemia) is essential for improving diabetes management. We developed approaches to predict daytime glucose levels in the target range, above-target range, and below-target range in patients with type 1 diabetes. Materials and Methods: A set of continuous glucose monitoring data from 406 patients managed with multiple daily insulin injections was used. Convolutional Neural Networks and Multilayer Perceptron was used for deep learning. The models were compared with those built using machine learning algorithms, Random Forest and XGBoost. Results: Both DL and ML models provide high accuracy when predicting glucose within the target range and the above-target range at 30-minute prediction horizon and 30-minute lookback window (AUC >0.992 for all models). The performance of the models in predicting below-target glucose was poorer. A model based on Convolutional Neural Network outperformed other ones when predicted glucose in all ranges. Conclusion: The data obtained indicate that deep learning (CNNs, MLP) and machine learning (RF, XGBoost) algorithms provide accurate prediction of day-time glucose in the target and above-target ranges in people with type 1 diabetes treated with multiple daily insulin injections. Further research is needed to improve glucose prediction in the below-target range.