Including Aerobic Exercise Into Data-Based Virtual Twins for Glycemic Simulation.

Area of Science:

• Biomedical Engineering
• Computational Biology
• Endocrinology

Background:

• Data-driven models are crucial for virtual patient simulations in diabetes management.
• Personalized therapies necessitate precise, individualized models compatible with machine learning.
• Deep generative models, like GANs, offer potential for advanced physiological simulations.

Purpose of the Study:

• To develop an extended conditional Wasserstein generative adversarial network (GAN) model for diabetes metabolism.
• To incorporate aerobic exercise intensity data alongside insulin and carbohydrate data.
• To improve the accuracy and applicability of virtual patient models in diabetes research.

Main Methods:

• An extended conditional Wasserstein GAN was developed, integrating aerobic exercise intensity, insulin, and carbohydrate data.
• An aerobic physical activity model was employed to simulate immediate and prolonged exercise effects on glycemia.
• The model was trained and validated using 1479 days of data from 56 patients, including 308 exercise sessions.

Main Results:

• The model accurately replicates real-world data from the T1DEXI study, including mean blood glucose and time in/above/below range metrics.
• Performance was validated for both aggregate data and separated active versus sedentary days.
• The model successfully reproduces exercise-induced glucose reductions.

Conclusions:

• The enhanced model offers a more reliable framework for in silico trials incorporating physical activity.
• This advancement has significant potential for designing and validating automated insulin delivery systems.
• The model improves virtual patient representation for diabetes research and treatment development.