Personalized Blood Glucose Forecasting From Limited CGM Data Using Incrementally Retrained LSTM

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Endocrinology

Background:

• Accurate blood glucose (BG) forecasting is essential for artificial pancreas (AP) systems in Type 1 diabetes (T1D) management.
• Current deep learning models like Long Short-Term-Memory (LSTM) for BG prediction using continuous glucose monitor (CGM) data often require extensive personalized training data and struggle with diverse glucose variability (GV).
• Existing methods face challenges in achieving consistent forecast accuracy across individuals with T1D.

Purpose of the Study:

• To introduce a novel deep learning framework, Incrementally Retrained Stacked LSTM (IS-LSTM), designed to improve BG forecasting accuracy for T1D.
• To address the limitations of existing methods regarding data requirements and individual glucose variability.
• To enhance the efficiency and adaptability of personalized BG prediction models.

Main Methods:

• Developed the Incrementally Retrained Stacked LSTM (IS-LSTM) framework, which utilizes gradual adaptation to individual data and parameter-transfer for efficient learning.
• Compared the IS-LSTM approach against three benchmark methods, including Stacked LSTM, using two independent CGM datasets (OpenAPS and Replace-BG) from individuals with T1D.
• Evaluated prediction accuracy using root mean square error (RMSE) and clinical feasibility via Clarke error grid analysis at various prediction horizons (PHs).

Main Results:

• IS-LSTM significantly reduced RMSE compared to Stacked LSTM: from 14.55 to 10.23 mg/dL (OpenAPS) and 17.15 to 13.41 mg/dL (Replace-BG) at a 30-minute PH.
• Clarke error grid analysis confirmed clinical feasibility, with over 98.81% (30-min PH) and 97.25% (60-min PH) of predictions falling within the clinically safe zones.
• The method demonstrated effectiveness in cold-start scenarios, providing accurate predictions for new CGM users without extensive prior data.

Conclusions:

• The IS-LSTM framework offers a significant advancement in personalized BG forecasting for T1D management, outperforming existing state-of-the-art methods.
• This approach enhances the reliability and safety of AP systems by providing accurate and clinically feasible BG predictions, even with limited data.
• IS-LSTM's adaptability and efficiency make it a promising tool for improving the daily lives of individuals with T1D, including new CGM users.