Development of Insulin Bolus Calculators in Type 1 Diabetes using A Framework Based on Real-world Data, Digital Twins...

Area of Science:

• Biomedical Engineering
• Computational Biology
• Endocrinology

Background:

• Precise mealtime insulin bolus (MIB) dosing is critical for managing type 1 diabetes (T1D) and preventing glucose fluctuations.
• Traditional MIB formulas are limited as they do not integrate real-time continuous glucose monitoring (CGM) data.
• Existing CGM-integrated methods often rely on empirical rules or in-silico models, hindering real-world applicability.

Purpose of the Study:

• To investigate a framework combining machine learning (ML), digital twins (DTs), and real-world data to enhance MIB dosing algorithms for T1D.
• To assess, tune, and develop improved MIB dosing strategies beyond traditional and in-silico approaches.
• To leverage DTs for refining MIB formulas using real-world patient data.

Main Methods:

• Utilized ReplayBG, a DT for T1D, to evaluate a published linear ML model on real-world data from 30 subjects.
• Recalibrated the linear ML model using the real-world dataset to improve its performance.
• Trained and tested nonlinear gradient-boosting models (XGBoost, LightGBM) entirely on real data via DT simulations.

Main Results:

• DT-enhanced models demonstrated improved glucose control compared to the original in-silico model.
• Recalibrated linear and nonlinear models significantly increased time-in-range (up to 80.6% with LightGBM) and reduced time-above-range.
• Risk metrics associated with hypoglycemia and hyperglycemia showed notable improvement.

Conclusions:

• A DT-based framework utilizing real-world data effectively refines and develops MIB calculators for T1D.
• This approach achieves performance gains exceeding those of the original in-silico model.
• DTs facilitate the development, validation, and extension of MIB formulas, enabling practical ML-tailored solutions for T1D management.