Identification of physiological parameters and missing inputs in glucose dynamics via PINN-constrained inference

Area of Science:

• Physiological modeling
• Biomedical engineering
• Systems biology

Background:

• Accurate glucose-insulin modeling is crucial for personalized diabetes management.
• Challenges include incomplete data, noisy glucose readings, and individual variability.
• Existing methods struggle with reliable personalization and system identification.

Purpose of the Study:

• Develop a physiologically grounded framework for glucose-insulin system identification.
• Estimate subject-specific physiological parameters and unobserved inputs from partially observed data.
• Enhance the accuracy and robustness of glucose-insulin models.

Main Methods:

• Introduced the Glucose Latent Input and Parameter Inversion (GLU-INVERT) framework.
• Extended the Bergman minimal model with additional physiological states.
• Employed a physics-informed learning mechanism and alternating optimization for parameter and input estimation.

Main Results:

• GLU-INVERT yielded parameter estimates within physiological ranges with reduced inter-subject variability.
• Achieved the lowest mean absolute relative difference (MARD) and root mean squared error (RMSE) in forecasting.
• Demonstrated statistically significant performance improvements and enhanced stability under data-limited conditions.

Conclusions:

• GLU-INVERT offers a robust and interpretable framework for physiological system identification.
• Addresses parameter uncertainty and missing input information in real-world data.
• Shows potential for personalized glucose monitoring and decision support through improved forecasting.