Machine-Learning Based Model to Improve Insulin Bolus Calculation in Type 1 Diabetes Therapy

Area of Science:

• Endocrinology and Metabolism
• Artificial Intelligence in Healthcare
• Biomedical Data Science

Background:

• Type 1 diabetes (T1D) management relies on accurate mealtime insulin boluses (MIB).
• Current standard formulas (SF) for MIB do not account for glucose rate-of-change (ΔG), leading to glycemic variability.
• Continuous glucose monitoring (CGM) data offers potential for more precise MIB calculations.

Purpose of the Study:

• To develop and evaluate machine learning models for improved MIB calculation in T1D.
• To incorporate CGM-derived glucose rate-of-change (ΔG) into MIB dosage algorithms.
• To compare the performance of novel models against standard formulas and existing literature methods.

Main Methods:

• Four machine learning models, including multiple linear regression (MLR) and LASSO, were developed for MIB calculation.
• Models were validated using the UVa/Padova T1D simulator in various mealtime scenarios.
• Retrospective analysis of 218 real-world glycemic traces was conducted to confirm findings.

Main Results:

• All developed models outperformed existing MIB calculation techniques.
• The LASSO regression with quadratic terms (LASSO Q) model demonstrated superior performance.
• LASSO Q significantly reduced bolus estimation error (0.86 U vs. 1.45 U for SF) and hypoglycemia incidence (35.93% vs. 44.41% for SF).

Conclusions:

• Machine learning frameworks can effectively develop improved MIB calculation models using CGM and ΔG data.
• The proposed LASSO Q model offers enhanced glycemic control compared to SF and other methods.
• Implementing the LASSO Q model for MIB dosage may reduce adverse glycemic events in T1D therapy.