Data Based Prediction of Blood Glucose Concentrations Using Evolutionary Methods

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Endocrinology

Background:

• Daily glucose monitoring is crucial for diabetes management to prevent acute and chronic complications.
• Accurate prediction of glucose levels is challenging due to complex insulin and food intake dynamics.
• Machine learning shows promise in modeling and predicting glucose concentrations.

Purpose of the Study:

• To evaluate and enhance machine learning techniques for predicting glucose concentrations in individuals with diabetes.
• To develop novel algorithms for improved glucose prediction accuracy.
• To assess the performance of different machine learning models using clinical data.

Main Methods:

• Utilized genetic programming, random forests, k-nearest neighbors, and grammatical evolution for glucose prediction.
• Proposed enhanced algorithms: optimized grammar grammatical evolution and tree-based genetic programming with a three-compartment model.
• Trained and tested models using continuous glucose monitoring data and carbohydrate/insulin intake information from ten patients.

Main Results:

• 90% of forecasts achieved clinically acceptable accuracy (Clarke error grid categories A and B).
• The best-performing methods still produced 5-10% of serious errors (category D) and ~0.5% of very serious errors (category E).
• An enhanced genetic programming algorithm incorporating a three-compartment model improved time series smoothing for carbohydrate and insulin dynamics.

Conclusions:

• Machine learning models offer a viable approach for glucose prediction in diabetes management.
• Further refinement of algorithms is necessary to minimize prediction errors and enhance patient safety.
• The proposed enhanced genetic programming model shows potential for more accurate and reliable glucose forecasting.