Machine Learning-driven Identification of the Honeymoon Phase in Pediatric Type 1 Diabetes and Optimizing Insulin...

Area of Science:

• Endocrinology
• Artificial Intelligence
• Data Science

Background:

• The honeymoon phase in Type 1 Diabetes (T1D) temporarily improves glycemic control, posing challenges for precise insulin management.
• Early and accurate identification of this phase is crucial for optimizing treatment and preventing complications.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for detecting the T1D honeymoon phase.
• To enhance personalized insulin therapy and improve patient outcomes through accurate phase identification.

Main Methods:

• Trained ML models including LSTM, Transformer, Random Forest, and Gradient Boosting on pediatric T1D patient data (ages 6-17).
• Utilized continuous glucose monitoring (CGM), GMI reports, HbA1c, and medical history for model training.
• Analyzed glucose trends and identified key features like glucose variability and insulin adjustments.

Main Results:

• The Transformer model achieved the highest accuracy (91%), followed by Gradient Boosting (89%), LSTM (88%), and Random Forest (87%).
• Glucose variability, insulin adjustments, GMI, and HbA1c were critical predictors.
• Accurate detection facilitated optimized insulin adjustments, improving glycemic control and reducing hypoglycemia.

Conclusions:

• The developed ML approach offers a robust method for identifying the T1D honeymoon phase.
• This technology holds significant potential for personalized insulin management and improved patient outcomes.
• Further validation and clinical integration are recommended for broader application.