Personalised Blood Glucose Time Series Forecasting in Type 1 Diabetes: Deep Collaborative Adversarial Learning.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Data Science

Background:

• Blood glucose prediction (BGP) is crucial for type 1 diabetes (T1D) management and personalized medicine.
• Glycaemic fluctuations in T1D present significant challenges for accurate time series forecasting (TSF).
• Conventional adversarial learning methods have limitations in capturing long-term temporal dependencies for BGP.

Purpose of the Study:

• To develop an advanced BGP technique addressing limitations in conventional adversarial learning.
• To enhance temporal awareness and clinical relevance in blood glucose forecasting models.
• To improve personalized diabetes management through more accurate glucose predictions.

Main Methods:

• Introduced collaborative augmented adversarial learning to improve temporal awareness in BGP.
• Incorporated collaborative interaction optimization to capture extended time dependencies.
• Evaluated four learning systems (independent, adversarial, collaborative, adversarial collaborative) for 30-min and 60-min prediction horizons.

Main Results:

• Collaboratively augmented learning frameworks demonstrated statistically significant superior performance.
• The approach showed improvements in both clinical relevance and overall predictive performance.
• Validated using the Ohio T1D datasets, confirming enhanced BGP accuracy.

Conclusions:

• The proposed approach advances BGP accuracy and clinical reliability for T1D management.
• It supports personalized medicine by improving subject-specific glucose forecasting from continuous glucose monitoring (CGM) data.
• Opens new avenues for TSF in other complex domains.