Mean glucose during ICU admission is related to mortality by a U-shaped curve in surgical and medical patients: a retrospective cohort study

We collected information about patients admitted between January 2004 and December 2007 in a 20-bed medical/surgical ICU in a teaching hospital (Onze Lieve Vrouwe Gasthuis [OLVG], Amsterdam, The Netherlands) (the OLVG cohort). All data were anonymous and collected retrospectively, so no ethical approval was necessary. On average, one nurse took care of two patients, depending on the severity of disease. All beds were equipped with a clinical information system (MetaVision; iMDsoft, Tel Aviv, Israel) from which all clinical and laboratory data were extracted. The glucose regulation algorithm was implemented successfully in 2001 [15], targeting for glucose values of between 4.0 and 7.0 mmol/L. The glucose protocol was started for every patient at the time of arrival at the ICU. Insulin infusion was started when admission blood glucose exceeded 7.0 mmol/L. When admission glucose was lower than 7.0 mmol/L, blood glucose was further measured every 2 hours and insulin was started when necessary (that is, when blood glucose exceeded 7.0 mmol/L). The nursing staff was instructed to use a dynamic computerized algorithm to adjust the insulin infusion rate, depending on the current glucose value and the rate of glucose change (based on the previous five measurements). The software also provided the time the next glucose measurement was due, which could vary from 15 minutes to 4 hours. Routinely, enteral feeding was started within 24 hours after admission, aiming at 1,500 kcal per 24 hours, and subsequently adjusted to the patient's requirements, except for the uncomplicated cardiac surgery patients, who do not receive enteral feeding if extubated within 24 hours. A duodenal feeding tube was inserted in case of persistent gastric retention. The tight glucose algorithm was deactivated when patients resumed normal eating.

We excluded readmissions, patients with a withholding care policy, and patients with only one glucose value measured during admission. From the clinical information system, we collected demographic variables, mortality rates in the ICU, and glucose values. For severity of disease measures, we used the Acute Physiology and Chronic Health Evaluation II (APACHE II) score [16]. Informed consent was not required according to Dutch Ethical Review Board regulations, because a retrospective analysis of anonymous data was performed.

For each patient, we calculated the mean overall glucose during admission from all glucose values measured during admission and the mean morning glucose from the first value available between 5 and 7 a.m. per patient per day. Glucose values mentioned in this paper stand for mean overall glucose unless stated otherwise. We calculated the standard deviation (SD) and the mean absolute glucose (MAG) change [6] per patient as markers of glycemic variability. Glucose was obtained from arterial blood samples by means of a handheld glucose measurement device (AccuChek; Roche/Hitachi, Basel, Switzerland). Results were automatically stored in the clinical information system.

The cohort characteristics are presented as mean ± SD or as median and interquartile range (IQR), depending on the distribution of the data. The mean glucose values and SDs were divided into five strata with equal numbers of patients per group. For each stratum, the ICU mortality was calculated. Subsequently, we performed a logistic regression analysis to calculate the odds ratio (OR) with 95% confidence intervals for ICU mortality per glucose stratum. The stratum with the lowest mortality incidence was used as a reference. In this model, we adjusted for age, sex, severity of disease (APACHE II score), occurrence of severe hypoglycemia (≤2.2 mmol/L), and admission duration (that is, ≤ or >24 hours). The last adjustment was done because glucose values are higher and have a wider range in the first 24 hours of admission, biasing the patients with longer admission times and corresponding lower mean glucose values. In a second model, adjustment for occurrence of mild hypoglycemia (≤4.7 mmol/L), which is also independently associated with mortality [17], was made.

The overall mean (SD) glucose values of the medical and surgical populations were 7.9 (2.7) and 8.1 (1.6) mmol/L, respectively (Table 1). The mean glucose values of the first 24 hours of admission were higher and had a wider range than did the mean glucose values after 24 hours (medical: mean [SD] 8.4 [3.3] mmol/L, range 3.7 to 40.2 mmol/L and 7.0 [1.4] mmol/L, range 3.2 to 31.1 mmol/L; surgical: mean [SD] 8.3 [1.9] mmol/L, range 0.6 to 27.5 mmol/L and 7.6 [1.7] mmol/L, range 3.2 to 15.7 mmol/L). The mean morning glucose values were 7.4 [2.6] mmol/L in the medical population and 7.7 [2.3] mmol/L in the surgical population. After division of the mean glucose of both populations into five equally sized strata, the lowest mean glucose stratum ranged from 6.7 mmol/L and lower in the medical cohort and from 7.0 mmol/L and lower in the surgical cohort. The highest stratum ranged from 8.5 mmol/L and higher in the medical cohort and from 9.5 mmol/L and higher in the surgical cohort. Mean glucose ranges per stratum and corresponding mortality rates per cohort are displayed in Figure 1. This results in a U-shaped curve relationship between mean glucose and mortality in both cohorts, with high ICU mortality in the lowest and highest glucose strata (medical: 26.9% and 35.6%; surgical: 3.6% and 1.4%). Logistic regression analysis showed that in both populations mean glucose values in the lowest and highest strata were associated with a significantly higher OR for ICU mortality compared with the stratum with the lowest mortality (Figure 2). This results in 'safe ranges' of 6.7 to 8.5 mmol/L in the medical cohort and 7.0 to 9.5 mmol/L in the surgical cohort. The non-linear U-shaped relationship between mean glucose and ICU mortality was supported by significance of the quadratic transformation of the mean glucose levels in this logistic regression model (P < 0.001). The characteristics of our populations, also subdivided in groups with low, 'safe range', and high glucose values, are displayed in Tables 1 and 2.

Overall, 9.9% and 1.8% of the medical and surgical patients, respectively, sustained at least one hypoglycemic episode, defined as a glucose value of not more than 2.2 mmol/L, during ICU admission. Seventeen point five percent of all deaths during ICU admission concerned patients who had experienced severe hypoglycemia (both groups). Twenty-eight percent of the patients who were in the lowest mean glucose strata and who died in the ICU experienced hypoglycemia, and 72% did not. The incidence of severe and mild (≤4.7 mmol/L) hypoglycemia in the different mean glucose strata is reported in Figure 3. When we adjusted the logistic regression model for occurrence of mild hypoglycemia with a cutoff value of 4.7 mmol/L, which is also independently associated with mortality [17], the OR (95% confidence interval) for ICU mortality in the lowest glucose stratum remained significant (medical: 2.6 [1.6 to 4.4], P < 0.001; surgical: 4.9 [1.1 to 22.1], P = 0.04).

In the medical cohort, glucose variability, both when expressed as the median of individual SDs and MAG changes [6], linearly increased with increasing glucose strata (SD median [IQR] 1.6 [1.2 to 1.9] to 3.8 [2.7 to 5.4] mmol/L, P for trend < 0.001; MAG 0.5 [0.3 to 0.8] to 1.4 [0.9 to 2.0] mmol/L per hour, P for trend 0.007). However, in the surgical cohort, no consistent trend in glucose variability across the glucose strata was seen (SD median [IQR] 1.8 [1.3 to 2.3] mmol/L; MAG 0.6 [0.4 to 0.8] mmol/L per hour). Adjusting the logistic regression model for variability did not change the above-described relationship between mean glucose and mortality (data not shown).