How Much Do We Gain From Greater Personalisation?

Objective:
STAR (Stochastic TARgeted) is risk-based glycaemic control (GC) using prediction of future insulin sensitivity (SI) variability to safely dose insulin and nutrition, where SI variability is the key driver in GC difficulty and hypoglycaemia. Currently, STAR uses a 2D stochastic model where current identified patient-specific SI is used to predict future SI variability in a cohort-specific sense. This study assesses the impact on GC performance of a new, more patient-specific 3D stochastic model, using previous and current SI values to predict metabolic variability.
Method:
Bi-variate and tri-variate Gaussian kernel density methods are used to estimate conditional probability estimation of future SI knowing current SI (2D model) and also previous SI (3D model). Models are built randomly using 411 (70%) of retrospective GC episodes. They are tested using clinically validated virtual trials on the 176 (30%) remaining patients, repeating 3 times (N=528 episodes). Safety, performance, and workload are compared.
Results:
Out of the total 528 simulated episodes, workload was similar (11.6 measures/day). Performance was similar (90% in 80-145mg/dL band), but tighter for the 3D model (78% vs 74% in 80-125mg/dL band). Median BG level was lower for the 3D model (108 [99, 120] vs. 113 [103, 124]mg/dL), with higher insulin (3.0 [1.5, 5.0] vs 2.5 [1.5, 4.0] U/h) and nutrition (99 [66, 100] vs 92 [70, 100] % goal feed). Safety was very slightly better for the 2D model (2% vs 3% BG<72mg/dL; 1% vs 1.4% BG<40mg/dL).
Conclusions:
The new, more personalised 3D stochastic model provides moderately improved performance and similar safety and workload. Overall, results suggest greater personalization in predicting variability can improve STAR GC performance and justify implementation to see if it improves outcomes.