Identifiability Analysis of Three Control-Oriented Models for Use in Artificial Pancreas Systems

Area of Science:

• Biomedical Engineering
• Control Systems Theory
• Computational Biology

Background:

• Type 1 diabetes (T1D) management requires accurate glucose control models.
• Existing control-oriented models for T1D glucose control include SOGMM, ICING, and MMC.
• Model identifiability is crucial for reliable glucose control and artificial pancreas development.

Purpose of the Study:

• To analyze the structural and practical identifiability of three common control-oriented glucose control models for T1D.
• To evaluate the parameter influence and identify key parameters for model accuracy.
• To compare the performance and parameter variability of SOGMM, ICING, and MMC models.

Main Methods:

• Structural identifiability analysis using generating series (GS) and identifiability tableaus.
• Practical identifiability assessment via sensitivity analysis, Latin hypercube sampling (LHS), and collinearity analysis.
• Model identification using continuous glucose monitor (CGM), insulin pump, and meal records from T1D patients (n=5).
• Performance evaluation using root mean square (RMS) error.

Main Results:

• Identifiable parameter sets were established for all models.
• Insulin sensitivity was found not to be the most dynamically influential parameter, contrary to prior assumptions.
• Models showed comparable performance (RMS ~20 mg/dl), but MMC failed for one patient and exhibited higher parameter variability.
• SOGMM and ICING demonstrated better performance and parameter stability than MMC.

Conclusions:

• Both structural and practical identifiability analyses are essential before model identification/individualization in T1D.
• While all models represent CGM data, their suitability for artificial pancreas systems varies.
• SOGMM and ICING are more promising than MMC due to performance and parameter stability, though ICING's large parameter set poses overfitting risks.