A Multitask Learning Approach to Personalized Blood Glucose Prediction.

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Diabetes Technology

Background:

• Accurate blood glucose prediction is crucial for diabetes management systems.
• Deep learning models excel but require substantial personalized data.
• Multitask learning offers a way to train accurate personalized models using data from multiple subjects.

Purpose of the Study:

• To compare the effectiveness of multitask learning against sequential transfer learning and subject-specific learning for blood glucose prediction.
• To evaluate the performance of neural network and support vector regression models using these different learning approaches.
• To assess the impact of multitask learning on both short-term and long-term glucose prediction horizons.

Main Methods:

• Implemented and compared multitask learning, sequential transfer learning, and subject-specific learning using neural networks and support vector regression.
• Evaluated predictive performance using root mean square error (RMSE) and Clarke Error Grid Analysis (EGA).
• Assessed model performance across various prediction horizons (30-120 minutes) and in scenarios with limited training data.

Main Results:

• Multitask learning demonstrated consistently superior performance across all prediction horizons.
• Achieved predictive accuracy (RMSE) of 18.8 ±2.3 mg/dL at 30 min and 47.2 ±4.6 mg/dL at 120 min.
• Obtained ≥93% clinically acceptable predictions (EGA) and robust performance during adverse glycemic events with minimal training data.

Conclusions:

• Multitask learning is an effective strategy for developing accurate personalized blood glucose prediction models.
• This approach enables the deployment of effective models even with limited subject-specific data, overcoming a key limitation of traditional deep learning methods.
• Multitask learning holds significant promise for advancing diabetes decision support systems and closed-loop insulin delivery.