Data-driven blood glucose level prediction in type 1 diabetes: a comprehensive comparative analysis.

Area of Science:

• Biomedical Engineering
• Data Science
• Endocrinology

Background:

• Accurate blood glucose level (BGL) prediction is vital for managing type 1 diabetes.
• Selecting appropriate models and input data remains a challenge in BGL prediction.

Purpose of the Study:

• To comprehensively investigate data-driven time series forecasting approaches for BGL prediction.
• To compare the impact of different input variables (univariate vs. multivariate) on prediction accuracy.

Main Methods:

• Compared classical time series, traditional machine learning, and deep neural networks for BGL prediction.
• Evaluated univariate (BGL only) versus multivariate inputs (BGL, carbohydrates, insulin, physical activity).
• Utilized two public Ohio datasets and assessed performance using regression and clinical metrics.

Main Results:

• Traditional machine learning models demonstrated the fastest training times and superior BGL prediction performance, particularly with multivariate inputs.
• The inclusion of additional variables did not always lead to significant performance improvements, suggesting the need for advanced data fusion techniques.

Conclusions:

• Traditional machine learning models are effective for BGL prediction in type 1 diabetes management.
• Multivariate data can enhance BGL prediction, but effective data fusion is key to fully leveraging additional information.