Nonlinear Machine Learning Models for Insulin Bolus Estimation in Type 1 Diabetes Therapy.

Area of Science:

• Biomedical Engineering
• Computational Biology
• Endocrinology

Background:

• Type 1 diabetes (T1D) management necessitates daily insulin injections due to beta-cell destruction.
• Standard insulin formulas (SF) lack glucose dynamics, leading to potential hypo/hyperglycemic events.
• Previous linear models integrated glucose rate-of-change (ROC) but nonlinearities remain a challenge.

Purpose of the Study:

• Investigate nonlinear machine learning models for improved insulin bolus estimation in T1D.
• Compare Random Forest (RF) and Gradient Boosting Tree (GBT) performance against linear models.
• Assess the impact on glycemic control and accuracy of optimal bolus estimation.

Main Methods:

• Utilized a dataset of 100 virtual subjects with simulated single-meal scenarios.
• Incorporated preprandial ROC, blood glucose (BG), and meal data.
• Trained and tested nonlinear RF and GBT models, comparing results to a validated linear model.

Main Results:

• RF and GBT models demonstrated statistically significant improvements in glycemic control.
• Time spent in hypoglycemia was reduced from 32.49% with linear models to 27.57% (RF) and 25.20% (GBT).
• Nonlinear models outperformed previously developed linear approaches in bolus estimation and glycemic control.

Conclusions:

• Nonlinear machine learning techniques, specifically RF and GBT, enhance insulin bolus estimation accuracy in T1D.
• These methods offer a promising approach to reduce adverse glycemic events.
• This study provides preliminary evidence for the clinical relevance of nonlinear ML in T1D therapy.