Impact of Model Parameterization and Basis Function Order on Predictive Accuracy in Model-Based Glycaemic Control

Abstract

Glycaemic control (GC) is crucial in critical care settings to reduce mortality and improve clinical outcomes. Model-based GC algorithms offer personalized and effective predictive control. This study investigates the influence of model parameterization and basis function order on the accuracy of future predictions of insulin sensitivity (SI) variability, a key determinant of glucose dynamics. Glycaemic data from 30 critically ill patient episodes were utilized to fit a glucose dynamics model, with SI identified using b-spline basis functions. Stochastic maps of SI changes over time were created based on the identified SI profiles, enabling the estimation of SI prediction distributions for future time steps. The impact of varying the number of basis functions (M) relative to data points (N) and the basis function order (d) on predictive accuracy was evaluated. Results showed that increasing the model parameterization led to wider prediction distributions, while higher basis function order resulted in tighter prediction distributions and more accurate predictions. The Akaike Information Criterion analysis suggested an optimal M:N ratio of approximately 0.45 for all basis function orders. These findings highlight the potential of improved model parameterization to enhance predictive capability in GC algorithms, ultimately benefiting patient outcomes in critical care settings.