Stochastic modelling of insulin sensitivity variability in critical care

Abstract

Tight glycemic control has been shown to reduce mortality by 29–45% in critical care. Targeted glycemic control in critical care patients can be achieved by frequent fitting and prediction of a patient's modelled insulin sensitivity index, S_I. This parameter can vary significantly in the critically ill due to the evolution of their condition and drug therapy.

A three-dimensional stochastic model of S_I variability is constructed using 18 long-term retrospective critical care patients’ data. Given S_I for an hour, the stochastic model returns the probability distribution of S_I for the next hour. Consequently, the resulting glycemic distribution 1 h following a known insulin and/or nutrition intervention can be derived. Knowledge of this distribution enables more accurate predictions for glycemic control with pre-determined likelihood based on confidence intervals.

Clinical control data from eight independent critical care glycemic control trials were re-evaluated using the stochastic model. The stochastic model successfully captures the identified S_I variation trend, accounting for 84% of measurements over time within the 0.90 confidence band, and 45% with a 0.50 confidence. Incorporating the stochastic model into the numerical glucose–insulin dynamics model, a virtual cohort was generated, imitating typical glucose–insulin dynamics in a critically ill population. Control trial simulations on this virtual cohort showed that the 0.90 confidence intervals cover 88% of measurements, and the 0.5 confidence intervals cover 46%. These results indicate that the stochastic model provides first order estimate of insulin sensitivity, S_I, variation and resulting glycemic variation in critical care.

Introduction

Critically ill patients often experience stress-induced hyperglycemia and high levels of insulin resistance, even given no history of diabetes [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [35]. The metabolic response to stress is characterised by major, highly variable changes in glucose metabolism. Increased secretion of counter-regulatory hormones leads to a rise in endogenously produced glucose and the rate of hepatic gluconeogenesis, and a concomitant reduction in insulin sensitivity. Tight glucose control has been shown to reduce intensive care unit (ICU) patient mortality by 45% if average glucose is less than 6.1 mmol/L for a cardiac care population [9], [10]. Krinsley [12] showed a 25–30% total reduction in mortality over a broader critical care population with a higher average glucose limit of 7.75 mmol/L. Therefore, control algorithms that provide tight regulation for glucose intolerant ICU patients would reduce mortality and the burden on time and medical resources.

Previous clinical glycemic control studies include [13], [14], [15], [16], [17], [18]. Chase et al. [13] and Doran et al. [16] focused on critical care patients, whose glucose–insulin dynamics are highly variable due to the stress of their illness and the impact of drug therapies. Chase et al. [13] developed a control algorithm that has been clinically verified in ICU to reduce elevated blood glucose levels in a controlled, predictable manner, while accounting for inter-patient variability and varying physiological condition. The overall approach is a targeted adaptive control scheme that identifies changes in patient dynamics, particularly with respect to insulin sensitivity.

Following Chase et al. [13], better understanding and modelling of patient variability in the critical care population can lead to better glycemic management in ICU. In particular, a common risk in any intensive insulin therapy is hypoglycemic events. Many current ad hoc intensive insulin therapy protocols have reported hypoglycaemic episodes up to 25% of incident rate [9], [19], [20], [21]. Many studies have shown that an episode of hypoglycaemia can lead to counter-regulation and severe rebound hyperglycemia, which is particularly difficult to control [22], [23], [24], [25]. Hypoglycaemic episodes therefore present a significant added risk in providing intensive insulin therapy in the ICU.

Understanding and modelling the variability in patient condition, or more specifically, the patients’ variable dynamic response to insulin, will thus assist clinical control intervention decision making, and minimise the associated risk. Currently, no intensive insulin therapy protocol offers the likelihood, or distribution, of glycemic response to an intervention, leaving clinicians partly blind in controlling such a highly dynamic system. Therefore, the ultimate goal of this study is to produce blood glucose distributions and confidence bands for control intervention decisions based on stochastic models of clinically observed parameter variations. Such bands will allow targeted control, with user specified confidence on the glycemic response outcome. The result will be added certainty and safety in providing tight glycemic control.