Virtual patient trials of a multi-input stochastic model for tight glycaemic control using insulin sensitivity and blood glucose data

Objective: Safe, effective glycaemic control (GC) requires accurate prediction of future patient insulin sensitivity (SI), balancing the risk of hyper- and hypo-glycaemia. The stochastic targeted (STAR) protocol combines a clinically validated metabolic model and SI metric with a risk-based stochastic approach to optimise patient specific insulin and feed rates. Validated virtual trials comparing a novel 3D stochastic model for prediction of future patient SI using current patient SI and current blood glucose (BG) to an existing 2D stochastic model for SI prediction were conducted. Methods: The virtual trials involved 1477 retrospective patients across two hospitals and two GC protocols. They were conducted using five-fold cross-validation to build each stochastic model, ensuring independent test data. Results: The 3D stochastic model shifted BG from the 4.4-8.0 mmol/L target band towards the lower 4.4-6.5 mmol/L band, providing a decrease from 12.31 % to 11.19 % in hyperglycaemic hours (BG > 8.0 mmol/L), but only a 0.24 % increase, from 1.01 % to 1.25 %, in light hypoglycaemic hours (BG < 4.0 mmol/L). Simultaneously, the 3D stochastic model enabled greater patient nutrition, and required negligible increase in computational or clinical workload. Conclusions: The 3D stochastic model provided greater personalisation and better realised STAR's design philosophy of minimising hyperglycaemic events for an acceptable clinical risk of 5.0 % BG < 4.4 mmol/L. Thus, this model could provide better clinical conformity to design targets if implemented within the STAR protocol. (C) 2020 Elsevier Ltd. All rights reserved.