Abstract

Objective: The honeymoon phase in Type 1 Diabetes (T1D) presents a temporary improvement in glycemic control, complicating insulin management. This study aims to develop and validate a machine learning-driven method for accurately detecting this phase to optimize insulin therapy and prevent adverse outcomes.

Methods: Data from pediatric T1D patients aged 6–17 years, including continuous glucose monitoring (CGM) data, Glucose Management Indicator (GMI) reports, HbA1c values, and patient medical history, were used to train machine learning models. These models Long Short-Term Memory (LSTM) networks, Transformer models, Random Forest, and Gradient Boosting Machines were designed to analyze glucose trends and identify the honeymoon phase in T1D patients.

Results: The Transformer model achieved the highest accuracy at 91%, followed by Gradient Boosting Machines at 89%, LSTM at 88%, and Random Forest at 87%. Key features such as glucose variability, insulin adjustments, GMI values, and HbA1c levels were critical in model performance. Accurate identification of the honeymoon phase enabled optimized insulin adjustments, enhancing glucose control and reducing hypoglycemia risk.

Conclusion: The machine learning-driven approach provides a robust method for detecting the honeymoon phase in T1D patients, demonstrating potential for improved personalized insulin management. The findings suggest significant benefits in patient outcomes, with future research focused on further validation and clinical integration.