Real-Time Model-Based Fault Detection of Continuous Glucose Sensor Measurements

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Diabetes Technology

Background:

• Continuous glucose monitoring (CGM) is crucial for artificial pancreas (AP) systems in type 1 diabetes (T1D) management.
• Faulty CGM readings can compromise AP system accuracy, leading to potentially dangerous glucose level fluctuations (hypoglycemia or hyperglycemia).
• Reliable fault detection in CGM data is essential for safe and effective AP operation.

Purpose of the Study:

• To develop and validate a novel algorithm for detecting faults in subcutaneous glucose concentration readings from CGM devices.
• To assess the algorithm's ability to mitigate the impact of CGM failures on insulin infusion rate calculations in AP systems.
• To enhance the safety and efficacy of AP systems for individuals with T1D.

Main Methods:

• Development of a nonlinear first-principle model for glucose-insulin-meal dynamics.
• Utilized an unscented Kalman filter for state and parameter estimation.
• Employed Principal Component Analysis (PCA) for dynamic change detection and a K-nearest neighbor (KNN) algorithm for fault classification, integrating data-driven and model-based approaches.

Main Results:

• The proposed fault detection algorithm demonstrated 84.2% sensitivity in identifying CGM failures.
• Successfully detected 155 out of 184 (84.2%) CGM failures.
• Achieved an average detection time of 2.8 minutes for CGM failures.

Conclusions:

• A novel integrated algorithm effectively detects CGM failures with high accuracy.
• The developed method significantly reduces the impact of CGM faults on insulin dosing.
• This approach promises to decrease the risk of hypoglycemia and hyperglycemia in T1D patients using AP systems.