External validation and sub-cohort analysis of stochastic forecasting models in NICU cohorts
Hyperglycaemia is a prevalent complication in the neonatal intensive care unit (NICU) and is associated with worsened outcomes. It occurs as a result of prematurity, under developed endogenous glucose regulatory systems and clinical stress. The stochastic targeting (STAR) framework provides patient-specific, model-based glycaemic control with a clinically proven level of confidence on the outcome of treatment interventions, thus directly managing the risk of hypo- and hyper- glycaemia. However, stochastic models that are over conservative can limit control performance. Retrospective clinical data from 61 episodes (25 retrospective and 36 from a prospective blood glucose control study) of insulin therapy in very-low birth weight (VLBW) and extremely-low birth weight (ELBW) neonates are used to create a new stochastic model of model-based insulin sensitivity (SI [L/mU/min]). Sub-cohort models based on gestational age (GA) and birth weight (BW) are also created. Performance is assessed by the percentage of patients who have 90% of actual intra-patient variability in SI captured by the 90% confidence bands of the cohort based (inter-patient) stochastic variability model created. This assessment measures per-patient accuracy for any given cohort model. Per-patient coverage trends were very similar between prospective and retrospective cohorts, providing a measure of external validation of cohort similarity. Per-patient coverage was improved though the use of BW and GA dependent stochastic models, which ensures that the stochastic models more accurately capture both inter- and intra- patient variability. Stochastic models based on insulin sensitivities during insulin treatment periods are tighter and give better and safer glycaemic control. More patient specific methods, particularly in the modeling of endogenous insulin and glucose production, will be required to further improve forecasting and glycaemic control.