LSTMs and Neural Attention Models for Blood Glucose Prediction: Comparative Experiments on Real and Synthetic Data

Area of Science:

• Biomedical Engineering
• Machine Learning
• Diabetes Technology

Background:

• Long short-term memory (LSTM) networks have previously shown promise in predicting blood glucose levels for type 1 diabetes management.
• Existing models often rely on manually engineered features, which may not capture complex physiological dynamics.

Purpose of the Study:

• To extensively evaluate LSTM network performance for blood glucose prediction across diverse scenarios (what-if, agnostic, inertial) using real and synthetic datasets.
• To assess the impact of incorporating additional physiological features and a novel attention module on prediction accuracy.
• To investigate the transferability of model performance between synthetic and real patient data.

Main Methods:

• Conducted large-scale experiments using LSTM networks on both real and synthetic time-series data.
• Incorporated additional physiological features (skin temperature, heart rate, skin conductance) and a time-of-day feature into LSTM models.
• Implemented and evaluated a novel neural attention module designed for time-series prediction.

Main Results:

• LSTM networks demonstrated robustness to noise and adaptability to new features on real patient data.
• A neural attention module improved performance on synthetic data but not on real data.
• The inclusion of time-of-day as a feature consistently enhanced LSTM performance on real data, but not synthetic data.

Conclusions:

• LSTM networks are a robust and adaptable tool for blood glucose prediction in type 1 diabetes.
• Performance observed on synthetic data does not reliably predict performance on real patient data, emphasizing the need for validation with clinical datasets.
• Careful consideration of input features, like time of day, is essential for optimizing real-world physiological model performance.