Glycaemic Control in ICU: Stable Patients Tend to Remain Stable

Objective:
STAR is a glycaemic control (GC) protocol with proven safety and performance. It uses a cohort-based 2D stochastic model of model-based, patient-specific insulin sensitivity (SI). Given current SI, it predicts a range of future SI values to dose insulin based on specified risk of hypoglycaemia. This study examines whether considering the prior change in SI (%SI) as an input to a 3D stochastic model can reduce the conservatism and provide more accurate estimates.
Method:
Metabolic data from 3 clinical ICU cohorts (819 episodes and 68629 hours) in Christchurch (SPRINT and STAR) and Hungary (STAR) are used. Triplets (%ΔSIn, SIn, SIn+1) are created for every hour to create a 3D stochastic model with inputs (%ΔSIn, SIn) and output SIn+1. The 5-95th percentile prediction width of the 3D model is compared at every %ΔSIn value to the 2D model 5-95th width. A narrower band for the 3D model indicates the 2D model used is over-conservative and GC could be more aggressive, while a wider bound indicates increased risks.
Results:
The 2D model is over-conservative for 77% of hours, mainly where %ΔSI is within an absolute 25% change, with 25-40% narrower prediction ranges. Predictive power is similar for both models, but much closer to the ideal value of 90% for the 3D model, indicating greater patient-specificity. Cross-validations show these results generalise well to different ICU populations.
Conclusions:
By reducing prediction range for 77% of hours across 3 ICU cohorts, predominantly where SI is stable, the new 3D model shows stable patients tend to remain stable in terms of %ΔSI. The new model better characterises patient-specific response to insulin, allowing more optimal dosing while increasing performance and safety.