Data Augmentation via Digital Twins to Develop Personalized Deep Learning Glucose Prediction Algorithms for Type 1...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Endocrinology

Background:

• Type 1 diabetes (T1D) management requires accurate glucose level prediction.
• Deep learning models show promise but need extensive, varied datasets.
• Collecting comprehensive T1D datasets is challenging and resource-intensive.

Purpose of the Study:

• To propose a data augmentation strategy using digital twins for T1D (DT-T1D).
• To generate personalized synthetic data for improved glucose prediction.
• To overcome limitations of data scarcity in training deep learning models.

Main Methods:

• Adapted ReplayBG, an open-source tool, to create DT-T1D from retrospective patient data.
• Developed a two-step strategy: generating DT-T1D and simulating patient-specific synthetic data.
• Trained personalized deep networks for glucose prediction using original and synthetic data.

Main Results:

• Integrating synthetic data consistently improved model performance.
• Models trained with synthetic data and a small fraction of original data matched full dataset performance.
• Demonstrated enhanced accuracy in glucose level prediction.

Conclusions:

• DT-T1D-driven data augmentation effectively addresses data scarcity in T1D research.
• This approach enhances deep learning model performance for precise glucose prediction.
• Highlights the potential of digital twin technology for robust, personalized T1D management tools.