Hyperglycaemic index as a tool to assess glucose control: a retrospective study

Acute hyperglycaemia is a prognostic factor for mortality in critically ill patients either in the presence or in the absence of diabetes mellitus [1‐3]. The benefit of strict glucose control in the intensive care unit (ICU) was demonstrated by the Leuven study [4, 5]. Remarkable reduction in morbidity and mortality was achieved in patients who were treated according to a protocol that aimed to achieve normoglycaemia. Thus, the logical aims of glucose control are to eliminate hyperglycaemia as rapidly as possible and to maintain normoglycaemia from then onward, while avoiding hypoglycaemia [6, 7]. Although this represents a clear goal for algorithms, there is no clear way to assess the performance of such algorithms [8].

In ICU patients we do not possess a measure such as glycosylated haemoglobin A_1c, which has proven to be an important predictor of long-term complications and to be useful for evaluating the quality of glucose control [9‐11]. Therefore, glucose itself must be measured during the ICU stay in order to determine whether hyperglycaemia is present. In studies of acutely ill patients, regular indices of glucose regulation that have been used are admission glucose, maximum glucose, mean morning glucose and mean glucose [5, 12‐15]. All of these indices have specific drawbacks. Admission glucose, maximum glucose and mean morning glucose are all based on either a single measurement or a subset of measurements, and therefore they are not indicative of overall hyperglycaemia. A single mean glucose that uses all measurements can be strongly biased by unequal time distribution between measurements, as commonly occurs in practice [16‐18]. Calculating time-averaged glucose compensates for an unequal time distribution of glucose measurements. However, hypoglycaemic episodes may still lower such an index, thus falsely suggesting normoglycaemia when in reality hyperglycaemia is present.

We hypothesized that an index that takes into account the unequal time distribution of glucose sampling and which is not falsely lowered by low glucose values would be a better index of glucose regulation. We defined the hyperglycaemic index (HGI; Fig. 1) as the area under the glucose curve above the normal range divided by the length of stay.

We evaluated the association of HGI and conventional glucose indices of regulation with mortality in a large group of ICU patients with a prolonged ICU stay.

In a retrospective analysis we included all patients older than 15 years of age admitted to the surgical ICU of our tertiary teaching hospital from 1990 to the end of 2001. Because glucose control appears to be particularly important in patients with prolonged stay in the ICU, we studied only those patients who stayed for 4 days or longer in the ICU [4, 5]. Age, sex, admission type and the Acute Physiology and Chronic Health Evaluation (APACHE) II score were obtained from case records and electronic databases of all admitted patients to our hospital. Blood glucose values were obtained from the central laboratory database.

During the 12-year period of the study, 6885 patients were admitted to the ICU. A total of 1779 patients (26%) stayed for a period of at least 4 days and were included in the present study. The mean age was 55 years (standard deviation 19 years) and 65% were male. Table 1 lists the demographical data and glucose-related measures for survivors and nonsurvivors. APACHE II scores were available for the years 1992–1999; for all other parameters there were no missing data. Abdominal surgery and trauma were the most frequent reasons for ICU admission.

A total of 65,528 glucose measurements were performed in the 1779 included patients, with a median number of glucose measurements of 21 (IQR 11–42). In fewer than 1% of the patients not a single glucose measurement was taken. The median mean glucose concentration of all patients was 7.0 mmol/l (IQR 6.1–8.6 mmol/l), median morning glucose was 6.7 mmol/l (IQR 5.9–8.1 mmol/l), median admission glucose was 7.3 mmol/l (IQR 5.8–9.7 mmol/l), median maximum glucose was 8.7 mmol/l (IQR 6.9–11.6 mmol/l) and median HGI was 1.0 mmol/l (IQR 0.4–2.4 mmol/l). Severe hypoglycaemia (glucose < 2.7 mmol/l) occurred in 177 (6.6%) patients. The median duration of such hypoglycaemic episodes was 1.5 hours (IQR 0.6–3.4 hours).

Survivors and nonsurvivors both stayed in the ICU for a median of 10 days (IQR 6–20 days for survivors and 6–17 days for nonsurvivors). A total of 295 patients (17%) died within 30 days after ICU admission.

In the univariate analysis, the median mean glucose level was 7.7 mmol/l in nonsurvivors and 6.8 mmol/l in survivors (P < 0.001). Median HGI was 1.8 mmol/l in nonsurvivors, which was twice that in survivors (P < 0.001). The ROC curves for all glucose-derived parameters are shown in Fig. 2. HGI had the highest area under the curve (0.64). Fig. 3 shows the relations between HGI quartiles and mortality. Mortality in the lowest HGI quartile was 8.6% as compared with 25.1% in the highest HGI quartile (P < 0.001).

Fig. 4 shows the area under the ROC curve for HGI when cutoff values other than 6.0 mmol/l are used.

In multivariate analysis with APACHE II score, sex and age, HGI remained the only statistically significant glucose index in the binary logistic model (P < 0.001). P values for mean glucose, morning glucose, admission glucose and maximum glucose were 0.08, 0.17, 0.43 and 0.49, respectively. With regard to mortality, the results of regression analysis did not differ between the cohort of patients whose APACHE II scores were available and the cohort of patients whose APACHE II scores were not.

Of all measures of hyperglycaemia evaluated, HGI correlated best with 30-day mortality in this population of critically ill patients. This supports our hypothesis that HGI is a useful index for quantifying glucose control. Therefore, assuming that normoglycaemia is the aim, the goal of a glucose–insulin algorithm is clear. The algorithm should obtain an HGI as close to zero as possible.

Both in the univariate analysis and in the multivariate binary logistic regression analysis – in which severity of illness, age and sex were included – HGI emerged as the best indicator of hyperglycaemia. We assume this reflects the fact that HGI takes better account of the variation in glucose concentrations over time, and avoids the possibility that alternating high and low values will average out to yield a normal value.

The principle of calculating the area under the glucose curve is not new. Brown and Dodek [8] determined the area under the curve above a glucose threshold of 11.5 mmol/l. The area under the curve was used to assess the speed of initial normalization of glucose with an insulin algorithm. Other recent studies have also used glucose thresholds above 10.0 mmol/l to separate good control from poor control [20, 21]. However, following the publication of the Leuven study findings [4, 5] such thresholds are now considered high, because that study demonstrated improved outcome if normoglycaemia (4.4–6.1 mmol/l) was pursued. Because the vast majority of glucose concentrations in the patients studied here were below 10 mmol/l, crucial information would have been lost if a cutoff of 10.0 mmol/l has been used, as reflected by a decreased area under the ROC curve at a cutoff value for calculation of HGI of 10.0 mmol/l (Fig. 4). The cutoff of 6.0 mmol/l was based on the upper limit of the group of patients with strict regulation in the Leuven study [4, 5]. Recently, Finney and colleagues [19] found that glucose regulation below 8.3 mmol/l was not related to a better outcome. This is in accord with our observations; Fig. 4 shows that the optimal cutoff for calculation of HGI lies between 6.0 and 8.0 mmol/l.

Some limitations of the present study should be mentioned. The study is a retrospective, single ICU study that covers a period when strict glucose control was not a major issue. Mortality at 30 days was used as an outcome measure to identify the best glucose index. HGI was the best measure of glucose control, but with a ROC area of 0.64 HGI alone cannot serve as a useful predictor of mortality. The relative contributions of endogenous glucose production, exogenous glucose supply and insulin to HGI and other some other measures could not be identified because our study did not include glucose infusion or (par)enteral feeding, and neither did it include intensive treatment with insulin.

It should be stressed that HGI was designed to quantify hyperglycaemia and not hypoglycaemia. Prevention of hypoglycaemia is a critical requirement of any algorithm for glucose control [5‐7]. However, unlike hyperglycaemia, hypoglycaemia is a phenomenon that tends to be relatively short-lived, as our results show, and could be quantified using more straightforward measures such as the lowest glucose concentration.

An elevated admission glucose level is associated with a worse outcome; this has been found by many investigators in various patient categories, and was also found in the present study [1‐3, 12, 13, 22‐30]. In our study, however, the area under the ROC curves was smaller for conventional measures of glucose control than it was for HGI.

Like other indices of glucose control, HGI is related to outcome. In contrast to admission glucose, however, HGI is also amenable to therapy. HGI involves additional computation (Fig. 1) as compared with more straightforward indices but it does not require more information. Calculating HGI should be feasible in ICUs that possess a patient database management system that can provide automated input for the HGI calculation. The fact that HGI expresses glucose regulation as a single value has methodological advantages. The performance of glucose–insulin algorithms could be compared with HGI, and therefore it is important to measure glucose regularly. A major advantage of HGI is that periods of very frequent sampling (e.g. during hyperglycaemia or hypoglycaemia) are compensated for because HGI is based on an average over time.

HGI must be reassessed in the era of tighter glucose control. Moreover, the value of HGI needs confirmation in other ICUs. Because HGI has not been used by other investigators, it would be of interest to determine how HGI compares with other glucose indices in observational or intervention studies. Existing glucose patient databases could be reanalyzed to determine HGI. The use of glucose measures to predict outcome independently of other parameters such as age and severity scores is interesting but lacks power, as is shown by the area under the ROC. In general, HGI may be more useful for relating hyperglycaemia to organ failure scores such as the Sequential Organ Failure Assessment score [31] or parameters of systemic inflammation.

Continuous measurements of blood glucose will allow us to calculate and compare HGI and the value of other glucose measures to a degree that is not possible with intermittent measurements [32‐35]. Currently available glucose sensors are promising but have not yet proven to be sufficiently reliable in critically ill patients and do not allow continuous measurements over prolonged periods [32‐35].

In conclusion, HGI quantifies the impact of hyperglycaemia in critically ill patients better than other glucose indices. HGI may thus be a useful measure of glucose control.

On request an annotated computer program with source code that calculates HGI, as well as regular glucose indices, will be provided. The program is written in the multiplatform language Java, and should run on every major platform.