Comparative Evaluation of Machine Learning and Deep Learning Models for Blood Glucose Prediction on the OhioT1DM Dataset

Abstract Type 1 diabetes mellitus is a common condition among young individuals, highlighting the need for accurate blood glucose level (BGL) predictions for effective continuous glucose monitoring. Investigating and comparing the performance of extreme gradient boosting models using a data-driven approach is essential for improving BGL prediction accuracy. This study extends the analysis of the OhioT1DM dataset by evaluating and comparing the performance of traditional machine learning models, extreme gradient boosting models (XGBoost, CatBoost, and LightGBM), and deep learning models (LSTM and Bi-LSTM) in predicting BGL. The findings demonstrate that extreme gradient boosting models can achieve competitive performance compared to certain deep learning architectures while being less computationally expensive. In this study, the LSTM model achieves an RMSE of 13.65 for a 30-minute prediction horizon, while the Bi-LSTM model records an RMSE of 21.73 when using continuous glucose monitoring (CGM) as the sole feature for future predictions using all the 12 patients.