Blood Glucose Prediction Method Based on Particle Swarm Optimization and Model Fusion.

Area of Science:

• Biomedical Engineering
• Data Science
• Endocrinology

Background:

• Blood glucose stability is critical for diabetic patient health outcomes.
• Accurate blood glucose monitoring is essential for effective diabetes control.
• Existing single regression models have limitations in predicting volatile blood glucose levels.

Purpose of the Study:

• To develop an advanced method for predicting blood glucose values in diabetic patients.
• To address the challenge of high blood glucose concentration volatility.
• To overcome the limitations of individual predictive models through fusion.

Main Methods:

• Utilized Kalman filtering for sensor signal noise reduction.
• Employed particle swarm optimization for hyperparameter tuning of Extreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM).
• Implemented a stacking model fusion approach with XGBoost and LightGBM as base learners and Bayesian regression as the meta-learner.

Main Results:

• The proposed stacking fusion model demonstrated accurate blood glucose value prediction.
• Achieved an average absolute percentage error of 13.01% in blood glucose prediction.
• Significantly outperformed six other models in prediction accuracy.

Conclusions:

• The developed stacking fusion model offers a superior method for diabetes blood glucose prediction.
• This approach effectively handles blood glucose volatility and improves prediction accuracy.
• The method holds promise for enhanced diabetes management through precise monitoring.