Blood glucose concentration prediction based on VMD-KELM-AdaBoost.

Area of Science:

• Biomedical Engineering
• Data Science
• Machine Learning

Background:

• Diabetic blood glucose concentration time series are inherently time-varying, nonlinear, and non-stationary, posing challenges for accurate prediction.
• Existing prediction models often struggle with the complex dynamics of glucose levels, impacting clinical decision-making.

Purpose of the Study:

• To develop a novel, accurate, and robust short-term blood glucose prediction model for diabetic patients.
• To enhance the reliability of glucose monitoring by improving prediction accuracy and lead time.

Main Methods:

• Decomposition of blood glucose time series into intrinsic mode functions (IMFs) using Variational Mode Decomposition (VMD).
• Integration of Kernel Extreme Learning Machine (KELM) with the AdaBoost algorithm for predicting individual IMF components.
• Superposition of individual IMF predictions to generate the final cumulative blood glucose concentration forecast.

Main Results:

• The VMD-KELM-AdaBoost model demonstrated superior prediction accuracy compared to ELM, KELM, SVM, and LSTM models.
• Achieved high prediction accuracy 60 minutes in advance, with RMSE of 10.1422, MAPE of 4.8629%, and CC of 0.8737.
• Clarke error grid analysis showed 95.7% of predictions in region A; hypoglycemia early alarm sensitivity was 94.8% with a 7.7% false alarm rate.

Conclusions:

• The VMD-KELM-AdaBoost model effectively addresses the non-stationary and nonlinear characteristics of blood glucose data.
• This advanced model offers a significant improvement in short-term blood glucose prediction accuracy and reliability for diabetes management.
• The model's performance indicates its potential for real-time glucose monitoring and early hypoglycemia detection systems.