The effect of diabetes mellitus on the association between measures of glycaemiccontrol and ICU mortality: a retrospective cohort study

Hyperglycaemia, hypoglycaemia and increased glucose variability in critically illpatients are independently associated with ICU mortality [1‐6]. In the last decade many clinical triallists have attempted to improvemortality rates through intensive insulin therapy. Unfortunately, these trials haveproduced conflicting data, with some of the studies showing lower and others highermortality with strict glucose control, the latter possibly due to an increased incidenceof hypoglycaemia [7‐12]. There is consensus about the importance to avoid hypoglycaemia and many ICUshave therefore increased their lower glucose limit [13]. However, there is no consensus about the optimal target glucose range. In aprevious database cohort study, we found an optimal mean glucose range of 6.7 to 8.4mmol/l in a medical cohort and 7.0 to 9.4 mmol/l in a surgical cohort [14]. We additionally found that glucose concentrations that were low but abovehypoglycaemic levels (between 2.3 and 4.7 mmol/l) were associated with increased ICUmortality [3]. Thus, in addition to the mean glucose concentration, glucose variability andhypoglycaemia, a fourth measure of glycaemic control - low glucose (2.3 to 4.7 mmol/l) -is associated with ICU mortality in the critically ill.

Underlying diabetes mellitus (DM) might affect the abovementioned associations. In arecent review we examined the current literature on glycaemic control and mortality indiabetic ICU patients and we found that, despite patients with DM having an increasedrisk of developing complications when admitted to the ICU, their risk of mortality isnot increased [15]. In addition, ICU patients with DM have lower mortality in the higher meanglucose range compared with those without DM, although varying cutoff values were used [16‐19]. Some studies found the opposite, with higher mortality rates for DM patientsin the low-normal mean glucose range. However, these findings were unadjusted resultsonly [18, 20] and this relation was not significant after adjustment for severity ofdisease [16]. Furthermore, high glucose variability in ICU patients with DM seems to beless harmful than in patients without DM [21, 22] although data are limited. Lastly, hypoglycaemia is associated with mortalityin patients with and without DM [3, 4, 23], while the risk of hypoglycaemia is higher in patients with DM [4, 24]. Altogether, some of the abovementioned findings are inconsistent and none ofthe reviewed studies evaluated all four measures of glycaemic control concomitantly.

The objective of this study was to determine whether the association between measures ofglycaemic control - mean glucose, glucose variability (measured as the mean absoluteglucose (MAG) change), the occurrence of hypoglycaemia (≤ 2.2 mmol/l) or lowglucose (2.3 to 4.7 mmol/l) - and ICU mortality differs between patients without andwith underlying DM in a large cohort of critically ill patients admitted to a generalICU of a teaching hospital in the Netherlands. We also explored the cutoff value fordetrimental low glucose in both populations.

From 11,901 ICU admissions, 10,320 patients were selected for analyses after excluding842 readmissions, 105 patients with a withholding care policy, and 714 patients with< 3 glucose measurements. The remaining cohort consisted of 8,682 (84.2%) patientswho did not have DM at the time of ICU admission (non-DM cohort) and 1,638 (15.8%)patients who had DM at the time of ICU admission (DM cohort). The percentage of medicaland surgical ICU admissions in the entire cohort was 26% and 74%. The non-DM:DM ratiowithin these groups was 9:1 in patients with a medical ICU admission diagnosis and 4:1in patients with a surgical ICU admission diagnosis. Table 1illustrates patient characteristics of the entire study population as well as thedifferences between the non-DM cohort and the DM cohort.

In this retrospective database cohort study evaluating the association of four measuresof glycaemic control and ICU mortality concomitantly, we found striking differencesbetween the non-DM cohort and the DM cohort. In the non-DM cohort, ICU mortality wassignificantly related to all four measures of glycaemic control: mean glucose, glucosevariability, the occurrence of hypoglycaemia (≤ 2.2 mmol/l) and low glucoseconcentrations up to 4.9 mmol/l. In contrast, in the DM cohort, only the occurrence ofhypoglycaemia (≤ 2.2 mmol/l) and low-glucose concentrations up to 3.5 mmol/l weresignificantly associated with ICU mortality, while mean glucose and glucose variabilitywere not. The presence of DM thus seems to affect the association between glucosecontrol and ICU mortality.

Our findings support the results of previous studies that have focused on understandingthe association between the presence of DM at ICU admission, glycaemia, and ICUmortality [7, 8, 16‐19, 31, 32]. In all these studies, a stronger association between hyperglycaemia and ICUmortality was found in patients without DM, in comparison with patients with DM.

Hypoglycaemia has been found to be a risk factor of mortality in patients without andwith DM in the literature [3, 4, 7, 8, 30, 33, 34]. Of note, different cutoff values were used to define hypoglycaemia, rangingfrom ≤ 2.2 mmol/l [4, 35] up to ≤ 4.7 mmol/l [3, 33]. We also found a significant independent association between hypoglycaemia(≤ 2.2 mmol/l) and ICU mortality, in both the non-DM and DM cohorts. However, theassociation between low glucose (2.3 and 4.7 mmol/l) and ICU mortality was onlysignificant in the non-DM cohort, not in the DM cohort. When exploring the cutoff valuefor detrimental low glucose in the present cohort, we found that lowest blood glucoseconcentrations up to 4.9 mmol/l were associated with an increased risk of ICU mortalityin the non-DM cohort, and 3.5 mmol/l in the DM cohort. The cutoff value in the non-DMcohort is in line with our previous study, in which we found that lowest glucose valuesup to 4.7 mmol/l were associated with significant increased ICU mortality [3]. Furthermore, this cutoff value is supported by the finding that the highermortality in the lower half of the U-shaped curve (< 6.8 mmol/l) in the non-DM cohortis mainly determined by the occurrence of glucose values ≤ 4.7 mmol/l and less bythe glucose range between 4.7 and 6.8 mmol/l. The cutoff value for detrimental lowglucose we found in our DM cohort (≤ 3.5 mmol/l) is also in line with theliterature [23, 30]. Both studies found that glucose concentrations ≤ 3.9 mmol/l weresignificantly associated with mortality in a subgroup of DM patients. Altogether, we canconclude that the cutoff value for detrimental low glucose is lower in the DM populationthan in the non-DM population.

The association between glucose variability and ICU mortality in patients without andwith DM was studied previously [22]. In this observational study of 4,084 patients (including 942 DM patients), astrong association of glucose variability - expressed as the coefficient of variation(standard deviation/mean glucose level) - with mortality was found in patients withoutDM, while, in concordance with our study, no association was found in patients with DM [22]. Of note, this measure of glucose variability does not take order and timeinto account.

Several explanations can be considered for the different associations between glycaemiccontrol and ICU mortality in patients without and with pre-existing DM. We previouslysuggested that adaptation to hyperglycaemia might be a key mechanism [15]. Acute hyperglycaemia and inflammation induce oxidative stress, which causesendothelial damage [36]. In patients without DM, cellular adaptation mechanisms will be activated forthe first time in the acute care setting, whereas patients with DM could already haveadapted to these insults during their years with DM and therefore better tolerateepisodes of hyperglycaemia in an acute care setting. In addition, cellular adaptation torecurrent hypoglycaemia is also a well-established phenomenon [37‐39]. Although speculative, adaptation to low glucose will already be present inpatients with DM and might explain why patients with DM can withstand relatively lowglucose values better.

Our results should be viewed in light of the study's strengths and limitations.Strengths of our study include the large number of ICU patients and that glucose valueswere captured automatically, which prevents transcription errors. Furthermore, this isthe first study examining all four markers of glycaemic control in a non-DM cohort and aDM cohort simultaneously. Also, we used a time-based metric for glucose variability andwe explored multiple cutoff values for hypoglycaemia. Potential limitations of the studyare that it is a single-centre study and retrospective in design, and thus ispotentially subject to systematic error and bias. However, all data were prospectivelycollected and independently measured. Moreover, the findings are robust and internallyconsistent.

As in all studies in this field, our definition for a patient's diabetic status may benonrepresentative. Unfortunately, glycosylated haemoglobin testing was not performedbefore ICU admission and we were unable to make a distinction between type 1 and type 2DM patients. In addition, we were not able to distinguish between diabetes patients withgood and poor chronic control, who may become hyperglycaemic due to acute illness.Whether this might affect the optimal glucose target for the DM cohort remainsunknown.

Another limitation was that we were not able to distinguish between spontaneous(illness-related) and treatment-induced hypoglycaemia or variability. However, otherstudies could make this distinction. Finfer and colleagues reported that patients whohad encountered severe or moderate hypoglycaemia while not being treated with insulinwere at an increased mortality risk [23]. But they also demonstrated that, although to a lesser extent,insulin-induced hypoglycaemia was associated with an increased risk for ICU death. Incontrast, Kosiborod and colleagues only reported a high risk for mortality in patientshospitalised with acute myocardial infarction who developed hypoglycaemia spontaneously.Iatrogenic hypoglycaemia after insulin therapy was not associated with higher mortalityrisk [40].

Furthermore, in our cohort, most patients were admitted for cardiothoracic surgery; wecorrected for this potential confounder in our regression analyses and still foundsignificantly increased ICU mortality in the lowest and highest mean glucose quintilesand in the highest glucose variability quartile in the non-DM cohort. Moreover, the highamount of cardiothoracic surgery patients in the studied cohort may also havecontributed to the high administration level of corticosteroids. In our hospital, as inmany European hospitals (but not in most North American cardiac surgical centres),corticosteroid administration during cardiac surgery is part of routine care. Allpatients who were in shock or had sepsis or systemic inflammatory response syndrome alsoreceived corticosteroids. This could possibly limit the external validity of thissingle-centre study.

In our analyses of glucose variability, we did not correct for the frequency of glucosemeasurements during ICU admission. However, we did correct for severity of disease,which in itself is clearly correlated with the frequency of glucose measurements and ICUmortality. Furthermore, the concern that the frequency of blood glucose measurements mayinfluence the relation between the MAG and ICU mortality has been previously discussed [41]. MAG is independent of the number of measurements, as long as blood glucosekeeps changing at a constant rate. The MAG only increases when there is actually moreglucose variability. The possibility to capture variability, if there is any, increaseswhen the number of glucose measurements is increased. However, this can be said for allmeasures of glucose variability and this is not unique for the MAG change.

A limitation of our correction for severity of disease is the use of the APACHE IIscore. Although the validation of the use of APACHE II score to predict mortality incardiac surgery patients is lacking, this adjustment is the best available method [29]. Finally, because of the observational nature of the study, no proof ofcausation can be derived from the abovementioned associations between glycaemic controland ICU mortality.

This retrospective database cohort study shows that the presence of DM affects theassociation between three out of four measures of glycaemic control and ICU mortality.Mean glucose and high glucose variability were associated with ICU mortality in thenon-DM cohort but not in the DM cohort, whereas hypoglycaemia (≤ 2.2 mmol/l) wasassociated with ICU mortality in both. We additionally found a higher cutoff value fordetrimental low glucose in our non-DM cohort (4.9 mmol/l) than the DM cohort (3.5mmol/l). Glucose concentrations ≤ 4.9 mmol/l should therefore be avoided in thenon-DM cohort, while DM patients seem to tolerate a wider glucose range. Further studiesshould examine whether new technologies - that is, continuous glucose monitoringtechnology - could be of use for clinicians to improve glycaemic control.