After-meal blood glucose level prediction using an absorption model for neural network training

Area of Science:

• Biomedical Engineering
• Computational Biology
• Endocrinology

Background:

• Diabetes Mellitus patients require tools for effective lifestyle management.
• Short-term Blood Glucose Level (BGL) predictions are crucial for outpatient diabetes care.

Purpose of the Study:

• To develop an accurate BGL prediction method for diabetes outpatients.
• To leverage baseline BGL, insulin dosage, and dietary intake for improved predictions.

Main Methods:

• A novel neural network training method incorporating a meal nutrient absorption model.
• Utilizing numerical absorption curve characteristics, insulin doses, and Continuous Glucose Monitoring (CGM) data.
• Comparison with a training method using only carbohydrate values.

Main Results:

• The proposed method demonstrated superior performance in 60- and 120-minute BGL prediction horizons.
• Achieved a Root Mean Square Error (RMSE) of 1.12 mmol/l and 1.75 mmol/l, respectively.
• Over 96% of predictions were within clinically acceptable ranges, outperforming the carbohydrate-only approach.

Conclusions:

• Integrating an absorption model into neural network training enhances BGL prediction accuracy.
• The developed method shows significant promise for improving diabetes management tools.
• Further clinical trials with larger patient cohorts are recommended to validate these findings.