Generalized multi task learning framework for glucose forecasting and hypoglycemia detection using simulation to...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Endocrinology

Background:

• Continuous glucose monitoring is vital for type 1 diabetes mellitus (T1DM) management.
• Current automated insulin delivery (AID) systems lack integrated multi-task learning for glucose prediction and hypoglycemia detection.
• Existing single-task models limit practicality and adaptability in AID.

Purpose of the Study:

• To introduce a domain-agnostic continual multi-task learning (DA-CMTL) framework.
• To simultaneously forecast glucose levels and classify hypoglycemia events within a unified model.
• To enhance the safety and reliability of automated insulin delivery systems.

Main Methods:

• Developed a DA-CMTL framework for simultaneous glucose forecasting and hypoglycemia event classification.
• Employed Sim2Real transfer learning for training on simulated datasets.
• Utilized elastic weight consolidation for domain generalization and adaptation.
• Validated the model on public datasets (DiaTrend, OhioT1DM, ShanghaiT1DM) and in vivo with diabetic rats.

Main Results:

• Achieved a root mean squared error of 14.01 mg/dL and mean absolute error of 10.03 mg/dL for glucose prediction.
• Demonstrated high accuracy in hypoglycemia event classification with 92.13% sensitivity and 94.28% specificity at 30-minute prediction.
• Reduced time below range in diabetic rats from 3.01% to 2.58% in real-world validation.
• Showcased cross-domain generalization capabilities.

Conclusions:

• The DA-CMTL framework offers a robust and scalable solution for integrated glucose management in T1DM.
• This approach has the potential to significantly improve the safety of automated insulin delivery systems.
• DA-CMTL provides a unified and adaptable platform for continuous glucose monitoring and event prediction.