Generative deep learning for the development of a type 1 diabetes simulator.

Area of Science:

• Biomedical Engineering
• Computational Biology
• Endocrinology

Background:

• Existing Type 1 diabetes (T1D) simulators struggle with physiological complexity due to imprecise models.
• There is a need for more accurate virtual patient models to advance diabetes treatments.

Purpose of the Study:

• To develop a simulation approach using a conditional deep generative model for Type 1 diabetes (T1D).
• To synthesize virtual patients that more accurately represent the glucose-insulin system physiology, overcoming limitations of current T1D simulators.

Main Methods:

• Utilized a sequence-to-sequence generative adversarial network to simulate virtual T1D patients causally.
• Embedded causality by training with shifted input-output pairs (90-min shift) to model insulin and carbohydrate impact on blood glucose.
• Validated the model using three distinct T1D patient datasets and for closed-loop therapy with a state-of-the-art controller.

Main Results:

• Generated virtual patients showed statistical similarity to real patients in time-in-range, means, and variability.
• Identified authentic causal links between insulin, carbohydrates, and blood glucose levels in virtual patients.
• The generative model exhibited more realistic behavior than conventional simulators during closed-loop insulin therapy.

Conclusions:

• The developed approach accurately captures physiological dynamics and establishes genuine causal relationships in Type 1 diabetes (T1D) simulation.
• This method holds significant promise for enhancing the development and evaluation of diabetes therapies.