Blood glucose concentration prediction based on kernel canonical correlation analysis with particle swarm...

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Time Series Analysis

Background:

• Continuous glucose monitoring systems (CGMS) track blood glucose fluctuations.
• Accurate short-term blood glucose prediction is crucial for diabetes management to prevent hyperglycemia and hypoglycemia.
• Existing methods often rely on linear models or complex multi-dimensional inputs.

Purpose of the Study:

• To develop a novel blood glucose prediction model using only historical glucose data.
• To incorporate non-linear relationships into canonical correlation analysis (CCA) for improved prediction accuracy.
• To establish a personalized hypoglycemic warning threshold based on prediction outcomes.

Main Methods:

• Utilized kernel canonical correlation analysis (KCCA) to capture non-linear dependencies in glucose data.
• Employed particle swarm optimization (PSO) to optimize kernel function parameters, reducing manual adjustment deviations.
• Implemented an error compensation (EC) technique for CCA to enhance prediction precision (EC-CCA).

Main Results:

• The PSO-KCCA model achieved an average Root Mean Square Error (RMSE) as low as 8.01 mg/dL for a 5-minute prediction horizon.
• High average R-squared values (up to 0.98) indicate strong model performance across various prediction horizons.
• EC-CCA demonstrated a 33.45% reduction in RMSE compared to standard CCA, with average hypoglycemic warning sensitivity and specificity of 94.37% and 92.25%, respectively.

Conclusions:

• PSO-KCCA is an effective method for accurate short-term blood glucose prediction.
• EC-CCA significantly reduces prediction delay and improves time-series forecasting.
• The personalized hypoglycemic warning system enhances patient safety by considering model accuracy.