Improved Methods for Mid-Term Blood Glucose Level Prediction Using Dietary and Insulin Logs

Area of Science:

• Endocrinology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Accurate blood glucose level prediction is crucial for daily diabetes management.
• Existing models require improvement, particularly for mid-term prediction horizons (120-180 min).

Purpose of the Study:

• To enhance the accuracy of blood glucose level predictions for insulin-dependent diabetes patients.
• To improve mid-term prediction accuracy (120-180 min horizons).

Main Methods:

• An absorption model-based method was used to train an artificial neural network.
• Parameters included insulin dosing, timing, baseline glucose, glucose infusion rate, and carbohydrate intake.
• Validated using data from a continuous glucose monitoring clinical trial.

Main Results:

• Achieved root mean square error of 1.72 mmol/L at 120 min and 1.95 mmol/L at 180 min.
• Demonstrated superior accuracy compared to existing comparable models on clinical data.

Conclusions:

• A short training period (approx. two weeks) using continuous glucose monitoring and logs is sufficient for accurate predictions.
• A hybrid model combining mid-term and short-term prediction methods is proposed for practical outpatient application.