Stochastic Model Predictive (STOMP) glycaemic control for the intensive care unit: Development and virtual trial validation. (February 2015)

• Record Type: Journal Article
• Authors: Stewart, Kent W.; Pretty, Christopher G.; Tomlinson, Hamish; Fisk, Liam; Shaw, Geoffrey M.; Chase, J. Geoffrey
• Is Part Of: Biomedical signal processing and control. Volume 16(2015)
• Page Start: 61
• Page End: 67
• Publication Date: 2015-02
• Subjects: Glycaemic control -- Intensive care -- Model predictive control -- Virtual patients; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2014.09.011
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Stochastic Model Predictive (STOMP) is a glycaemic control protocol that combines the probabilistic, stochastic forecasting methods of previous methods (STAR) with model predictive control, for ease of tuning. Clinically validated virtual trials were used to evaluate the relative performance of STOMP. Results showed STOMP was able to obtain results very similar to STAR with both protocols maintaining approximately 85% of time within 4.4–8.0 mmol/L glycaemic band. STOMP was able to attain similar results to STAR while further increasing ease of controller tuning for different clinical requirements and reducing the number of BG measurements required by 35%. Abstract: Critically ill patients often experience stress-induced hyperglycaemia, which results in increased morbidity and mortality. Glycaemic control (GC) can be implemented in the intensive care unit (ICU) to safely manage hyperglycaemia. Two protocols SPRINT and STAR, have been implemented in the Christchurch ICU, and have been successful in treating hyperglycaemia while decreasing the risk of hypoglycaemia. This paper presents a new GC protocol that implements the probabilistic, stochastic forecasting methods of STAR, while formalizing the control methodology using model predictive control (MPC) theory to improve the ability to tune the dynamic response of the controller. This Stochastic Model Predictive (STOMP) controller predicts the response to a given insulin/nutrition intervention, and attributes weighted penalty values to several key performance metrics. The controller thus chooses an intervention at each hour that minimizes the sum of these penalties over a prediction window of 6 h, which is twice as long as the 3-h window used in STAR. Clinically validated virtual trials were used to evaluate the relative performance of STOMP. Results showed STOMP was able to obtain results very similar to STAR with both protocols maintaining approximately 85% of time within 4.4–8.0 mmol/L glycaemic band, and only 4–5 patients of the 149 patient STAR cohort having blood glucose (BG) <2.2 mmol/L. STOMP was able to attain similar results to STAR while further increasing ease of controller tuning for different clinical requirements and reducing the number of BG measurements required by 35%.