Physical Activity Integration in Blood Glucose Level Prediction: Different Levels of Data Fusion

Area of Science:

• Biomedical Engineering
• Data Science
• Endocrinology

Background:

• Accurate blood glucose level (BGL) prediction is essential for effective diabetes management.
• Physical activity (PA) significantly impacts BGL, but its complex and variable effects pose challenges for prediction models.
• Optimal utilization of PA data is crucial for enhancing BGL prediction performance.

Purpose of the Study:

• To develop and evaluate novel PA-informed models for improved BGL prediction.
• To explore different data fusion strategies (signal, feature, and decision levels) for integrating PA and BGL data.
• To identify the most effective methods for incorporating PA information into BGL prediction.

Main Methods:

• Proposed several PA-informed BGL prediction models.
• Developed signal-level fusion by integrating automatically recorded PA data with BGL data.
• Implemented feature-level fusion using subjective/objective PA assessments and PA statistics.
• Utilized ensemble learning for decision-level fusion of predictions.
• Conducted comparative analysis on the Ohio dataset.

Main Results:

• Signal-level fusion of heart rate data with BGL data demonstrated effectiveness.
• Feature-level fusion using PA intensity categories significantly improved prediction accuracy.
• PA-informed models outperformed the no-fusion approach in BGL prediction.
• Specific fusion strategies showed superior performance compared to others.

Conclusions:

• Integrating PA data, especially heart rate and intensity categories, is a highly effective strategy for enhancing BGL prediction.
• Signal-level and feature-level fusion approaches offer promising avenues for leveraging PA information.
• The findings provide valuable insights for developing more accurate and personalized diabetes management tools.