Deep learning for blood glucose level prediction: How well do models generalize across different data sets?

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Data Science for Diabetes Management

Background:

• Accurate blood glucose level prediction is crucial for diabetic patient care and closed-loop therapy systems.
• Existing deep learning models for glucose forecasting may exhibit biases due to varied methodologies and datasets.
• A comprehensive comparison is needed to assess model applicability and generalizability.

Purpose of the Study:

• To compare the performance and generalizability of various deep learning models for blood glucose level prediction.
• To identify optimal models for different objectives in diabetes management.
• To provide insights into model selection based on accuracy and generalization capabilities.

Main Methods:

• Evaluated feed-forward neural network (FFN), convolutional neural network (CNN), Long Short-Term Memory network (LSTM), temporal convolutional neural network (TCNN), and self-attention network (SAN).
• Utilized four diverse datasets varying in size, age groups, and conditions to test generalization.
• Employed Root Mean Square Error (RMSE), Mean Absolute Difference (MAD), Coefficient of Determination (CoD), Clarke Error Grid (CEG), and Kolmogorov-Smirnov (KS) test for analysis.

Main Results:

• LSTM demonstrated the best performance with the lowest RMSE and highest generalization capability.
• SAN closely followed LSTM in predictive performance and generalization.
• FFN showed potential for capturing trends despite lower overall predictive accuracy.

Conclusions:

• LSTM and SAN models are highly effective for blood glucose forecasting due to their ability to capture long-term dependencies.
• Model selection should align with specific goals, balancing accuracy and generalization requirements.
• This comparative analysis aids in choosing appropriate deep learning models for diabetes care and research.