Long-Term Glucose Forecasting for Open-Source Automated Insulin Delivery Systems: A Machine Learning Study with...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Diabetes Technology

Background:

• Glucose forecasting is crucial for diabetes management, particularly for real-time insulin dosing and optimizing physical activity.
• Open-source automated insulin delivery (AID) systems generate large datasets valuable for developing predictive models.

Purpose of the Study:

• To evaluate machine learning (ML) and deep learning (DL) methods for predicting glucose variability (GV) in individuals using open-source AID systems.
• To assess the accuracy, performance, and resource costs of various ML/DL models for glucose forecasting up to 48 hours ahead.

Main Methods:

• A three-stage framework: data collection, data preparation/analysis, and ML/DL model development/evaluation.
• Systematic implementation and fine-tuning of ML/DL models on a large-scale diabetes dataset.
• Evaluation of 17 GV metrics, including mean absolute error (MAE) and root mean square error (RMSE), alongside model execution time and memory consumption.

Main Results:

• Fine-tuned ML/DL models demonstrated considerable accuracy in glucose forecasting and variability analysis up to 48 hours in advance.
• Long Short-Term Memory (LSTM) models achieved the lowest average MAE (2.50 mg/dL) and RMSE (3.7 mg/dL), outperforming Autoregressive Integrated Moving Average (ARIMA) models.
• Model execution time correlated with training data size; LSTM models had the lowest execution time but highest memory usage.

Conclusions:

• ML/DL methods, particularly LSTM, show significant potential for accurate glucose forecasting and variability analysis in AID systems.
• The study highlights the importance of considering model performance, resource costs, and sustainable implementation for real-world applications.
• This research supports the integration of advanced computational tools for enhanced diabetes management through AID systems.