Effect of Diabetes and Admission Blood Glucose Levels on Short-term Outcomes in Patients with Critical illnesses: A Real-world Study Based on Propensity Score Analyses

Background: Effect of diabetes and optimal range of blood glucose level on the short-term prognosis in patients with critical illnesses are currently ambiguous. We aimed to determine whether diabetes affects short-term prognosis and optimal range of blood glucose level for critically ill patients. Methods: We performed a retrospective analysis of data on 46,476 critically ill patients from the critical care database (version 1.4), collected prospectively between 2002 and 2012. Association of diabetes with 28-day mortality was assessed by inverse probability weighting based on the propensity score. Smoothing splines and threshold effect analysis were applied to explore the optimal glucose range. Results: Of the 33,680 patients enrolled in the study, 8,701 (25.83%) had diabetes. In the main analysis, the 28-day mortality was reduced by 29% (hazard ratio (HR)=0.71, 95% condence interval (CI) 0.67-0.76) in patients with diabetes compared to those without diabetes. The E-value of 2.17 indicated robustness to unmeasured confounders. The effect of the association between with diabetes and 28-day mortality was generally in line for all subgroup variables, signicant interactions were observed for glucose at ICU admission, admission type, and insulin use (Interaction P <0.05). A V-shaped relationship was observed between glucose and 28-day mortality in patients without diabetes, with the lowest 28-day mortality corresponding to the glucose level was 101.75 mg/dl (95% CI 94.64-105.80 mg/dl), and hypoglycemia or hyperglycemia should be avoided, especially in patients admitted to the surgical intensive care unit (SICU), cardiac surgery recovery unit (CSRU), and coronary care unit (CCU); for patients with diabetes, no optimal threshold for glucose was found, and elevated blood glucose does not appear to be associated with a poor prognosis, and perhaps may be benecial except for CCU and CSRU.


Background
Diabetes is rapidly emerging as a pandemic worldwide. Of the overall U.S. population, approximately 34.2 million people had developed diabetes in 2018, of which 34.1 million were adults (> 18 years), accounting for 13.0% of all U.S. adults [1]. In critical ill patients admitted to the intensive care units (ICUs), diabetes is a relatively frequent diagnosis. Diabetes occasionally leads to ICU admissions but is usually a part of a comorbid condition. As diabetes is a multifaceted disease that can lead to immune system impairments and metabolic dysregulations, treatment of patients is complicated, potentially increasing the severity of the primary disease [2,3]. On reviewing studies over the past 20 years, the percentage of critically ill patients with diabetes as a comorbid condition was found to range from 13.2% to 30.4% [4][5][6][7][8][9][10][11][12][13][14].
Furthermore, these studies were inconclusive about the relationship between diabetes and the prognosis in patients with acute illness. Some studies suggested that in critically ill patients, diabetes increased the risk of mortality as well as the incidence of infection [8][9][10]. Other studies suggested that diabetes may not negatively affect the clinical outcomes of critically ill patients and may even be protective, which seems to con ict with clinical judgment [11][12][13].
In addition, there is no consensus on the scope of optimal glycemic control to date. In the past two decades, numerous clinical trials have attempted to explore the optimal range of glycemic control to reduce mortality in critically ill patients; however, these clinical trials have yielded con icting results [4][5][6][7][14][15][16][17][18][19]. From the rst clinical trial in 2001 until now, there has only been a consensus that the safest range for glycemic control in critically ill patients should not be higher than 180 mg/dl; however, the safest range for glycemic control is still unknown [4]. In addition, the quality of glycemic control in critically ill patients has been suggested to be closely related to prognosis [20]. Indeed, poor glycemic control may result in glycemic variability, either hypoglycemia or hyperglycemia, which could be associated with negative outcomes in patients with critical illness. Determining the optimal range of glycemic control from the results of clinical trials with large sample sizes is di cult, and previous clinical trials have been both hard-won and praiseworthy. Considering the susceptibility to variations in blood glucose levels in critically ill patients after ICU admission, we tested blood glucose levels at admission to retrospectively investigate whether there is an optimal blood glucose level range for critically ill patients and, if so, to determine this range.
The objectives of our study were to determine 1) whether diabetes affects short-term prognosis and 2) the optimal range of blood glucose level in critically ill patients.

Patient data
We carried out a retrospective analysis of data on 46,476 critically ill patients from the Critical Care Multiparameter Database III (MIMIC-III) version 1.4, collected prospectively between 2002 and 2012 [21]. The Beth Israel Deaconess Medical Center and the institutional review boards of MIT-a liated institutions granted permission to access the database (Record ID: 33460949). The requirement for patient consent was waived because of the anonymity of data.
All adult (≥18 years) patients were recruited for the study, but patients with a follow-up of >1 day were included in the analysis. We only analyzed patients who were admitted to the ICU for the rst time. Blood glucose level was determined on the basis of the rst plasma glucose level obtained in the patients admitted to the ICUs, which was measured by hospital laboratory staff and uploaded onto the electronic medical records system. Together with conventional variables such as demographic characteristics (age, sex, admission type, etc.), therapeutic interventions, and clinical outcomes, we simultaneously extracted data on the simpli ed acute physiology score II (SAPS II) [22], sepsis, and speci ed comorbidities.
Therapeutic interventions included the requirement for mechanical ventilation (MV) or renal replacement therapy (RRT) on the rst day and insulin for the entire ICU period. Data were extracted using Structured Query Language (SQL), and the code to support MIMIC-III documentation and websites is publicly available (https://github.com/MIT-LCP/mimic-code) [23,24].

Outcomes
The primary outcomes were 28-day mortality in critically ill patients with diabetes and the optimal blood glucose level range.

Statistical analysis
Data were expressed as mean ± standard deviation or median (interquartile range, IQR) for continuous variables, with comparison of group characteristics using the Kruskal-Wallis test (or Fisher's exact test); categorical variables were expressed as numbers and percentages, which were compared using the Chisquare test. We constructed propensity score methods utilizing the following three propensity score models: covariate-adjusted propensity scores, propensity score matching (PSM), and inverse probability weighting based on the propensity score. A logistic regression model was used to compute the propensity scores, which consisted of the following baseline covariates: age, sex, sepsis, admission type, use of insulin, type of ICU at admission, SAPS II, MV on the rst day, blood glucose level at ICU admission, RRT on the rst day, cardiac arrhythmias, congestive heart failure, peripheral vascular disease, valvular disease, hypertension, other neurological diseases, chronic pulmonary disease, renal failure, liver disease, acquired immune de ciency syndrome (AIDS), lymphoma, metastatic cancer, solid tumor, obesity, uid and electrolyte disorders, alcohol abuse, drug abuse, and depression [25]. Propensity score matching was performed using a 1:1 matching protocol without replacement (greedy-matching algorithm), with a caliper width equal to 0.05 of the standard deviation of the logit of the propensity score. For the inverseprobability-weighted analysis, we used the predicted probability of a propensity score model to obtain stable inverse probability weights [26]. Standardized differences between covariates that were less than 10.0% represented a relatively balanced condition [27]. The adjustment variables in the multivariate analysis model were consistent with those in the propensity score models. The results of the main analysis were used along with inverse probability weighting. We also calculated E-values to explore the potential effect of unmeasured confounders of diabetes and 28-day mortality [28,29]. The E-value allowed for quantifying the required magnitude of an unmeasured confounding factor that might counteract the observed correlation between diabetes and 28-day mortality. The Kaplan-Meier method (log-rank test) was used to describe survival differences of the PSM cohort.
Subsequently, to determine whether there was a U-shaped association between 28-day mortality and blood glucose levels on admission, we performed a smoothing splines using a Cox model to t the 28day mortality. In ection points were measured by the two-piece-wise regression model constructed in the threshold effects analysis, with the differences compared via the log-likelihood ratio test and the 95% con dence interval (CI) for the in ection point computed using the bootstrap method [30]. Finally, we performed strati ed analyses and interaction tests to explore the concordance of the associations for outcomes within different subgroup variables, which included demographic characteristics (e.g., age, sex, etc.), therapeutic regimens (e.g., use of insulin or not), different glucose concentrations, types of ICUs, sepsis, and SAPS II. EmpowerStats (www.empowerstats.com) and R (http://www.R-project.org, version 3.4.3) were applied to all data analyses. A P value <0.05 was considered statistically signi cant.

Characteristics of the original cohort
In this study, 33,680 patients with a mean age of 63.65 years were enrolled, among whom 8,701 had diabetes, accounting for about 25.83% of patients in the original cohort [ Table 1 and Figure 1]. Before propensity score matching, there were several differences in the baseline variables between the groups of patients with and without diabetes. With propensity score matching in a 1:1 ratio, 7,261 patients with diabetes were matched with 7,261 patients without diabetes. The model had a C-statistic of 0.8214. After matching, all variables had standardized differences less than 10.0%, which suggested that the betweengroup difference was relatively small [ Table 2].  Clinical outcomes in the original cohort and PSM cohort Altogether, in the original cohort, there were no obvious differences in the 28-day mortality, in-hospital mortality, or ICU mortality between the two groups (all P>0.05) [ Table 3]. After matching, patients with diabetes had a signi cantly lower 28-day, ICU, and in-hospital mortality than did those without diabetes (all P<0.001); Kaplan-Meier curves indicated that patients with diabetes had a considerably better survival advantage than did matched patients without diabetes (log-rank test: P <0.0001) [ Figure 2]. Associations between with diabetes and clinical outcomes As shown in Table 4, a multivariable model with inverse probability weighting based on the propensity score showed that the 28-day mortality rate was reduced by 29% (hazard ratio (HR)=0.71, 95% CI 0.67-0.76) in the group with diabetes compared with the group without diabetes. The estimated E-value was 2.17 (upper con dence limit 1.96), which means that if there were no immeasurable confounders associated with diabetes as well as 28-day mortality, with relative risks of ≥2.17 for both, then the results we obtained were robust. Propensity score matching was performed with the use of a 1:1 matching protocol without replacement (greedy-matching algorithm), with a caliper width equal to 0.05 of the standard deviation of the logit of the propensity score.
Inverse probability weighting was used with the same covariates according to the propensity score.

Strati ed analyses and interaction tests
Altogether, the effect of the association between diabetes and 28-day mortality was generally in line for all subgroup variables [

Smoothing splines and threshold effect analysis
We performed smoothing splines of blood glucose concentrations to 28-day mortality in patients with and without diabetes, indicating a V-shaped relationship between blood glucose concentrations and 28day mortality in patients without diabetes, with either a relatively low or high blood glucose concentration being negatively detrimental to clinical outcomes. However, in patients with diabetes, the effect of glucose concentration on 28-day mortality was weaker than in those without diabetes [ Figure 3]. Subsequently, in the quantitative analysis of the threshold effect, the in ection point was determined to be 101.75 mg/dl (95% CI 94.64-105.80 mg/dl) for 28-day mortality in patients without diabetes. After correction for potential confounders, the threshold effect of glucose on 28-day mortality was markedly signi cant, namely, the HR for glucose <101.75 was 0.98, and the HR for glucose ≥101.75 was 1.01 (Pvalue for log-likelihood ratio test, <0.001) [ Table 5]. As for the 28-day mortality in patients with diabetes, threshold effects analysis revealed that no signi cant differences were observed in either the one-line linear regression model or the two-piece-wise linear regression model (P-value for log-likelihood ratio test, 0.132) [ Table 5]. Additionally, we tted smoothing splines for the relationship between glucose and 28day mortality in diabetic and non-diabetic patients in different ICUs and further analyzed the effects of different glucose level groups on 28-day mortality in diabetic and non-diabetic patients in different ICUs [ Table 7, and Figure 4 and 5]. As shown in Table 7 was not associated with a poor prognosis, and is perhaps bene cial for the patients. Abbreviations: HR, hazard ratio; CI, con dence interval.

Discussion
In this study, we purposefully selected a well-represented study population of critically ill patients admitted to ICUs from a large critical healthcare database and investigated the clinical outcomes of these patients with and without diabetes, in addition to exploring the impact of blood glucose level at admission on the presence or absence of diabetes. We found that 1) diabetes was not a detrimental factor for critically ill patients in the ICUs, which would reduce the risk of 28-day mortality by about 29%, 2) a V-shaped relationship was observed between blood glucose level and 28-day mortality in patients without diabetes, and hypoglycemia or hyperglycemia should be avoided, especially in patients admitted to the SICU, CSRU, and CCU; for patients with diabetes, no optimal threshold for glucose has been identi ed, and an elevated blood glucose level does not appear to be associated with a poor prognosis and is perhaps bene cial for certain ICU patients, and 3) particular attention should be paid to hypoglycemic events in critically ill patients without diabetes in the SICU and hyperglycemic events in all critically ill patients in the CCU and CSRU regardless of the presence of diabetes, which warrants the attention of clinicians.
Currently, the mechanisms underlying what appears to be a predominantly neutrally or protective link between diabetes and mortality in critically ill patients continue to be elusive. If diabetes is associated with decreased mortality in patients with critical illness, just a single mechanism may not be involved. Biologically, part of the potential mechanism may be that glucose plays a critical role in the function of activated immune cells and that glucose is a key contributor to energy production and maintenance of immune cell functions as well as the synthesis of immunomodulators [31][32][33]. Furthermore, diabetic patients develop a tolerance to hyperglycemia because of chronically elevated blood glucose concentrations, making the harmful hyperglycemia transform into an "energy factory" [34], and for the harmful effects to persist, higher blood glucose concentrations would be required [35,36]. Namely, the smoothing splines revealed graphically that blood glucose level has a relatively small effect on patients with diabetes, and in subsequent further analyses, elevated blood glucose was not statistically associated with an increase in 28-day mortality. Our results also revealed an association between the use of insulin and diabetes on 28-day mortality, that is, there was a signi cant reduction in 28-day mortality in patients with diabetes who used insulin, whereas no such difference between patients with and without diabetes who did not use insulin was observed. Further analysis of the with or without insulin used cohort revealed that only 1732 patients (11.85%) had diabetes in the non-insulin cohort, which may have weakened the decreasing effect of 28-day mortality in the diabetes cohort and resulted in the negative ndings. However, this does not completely refute the association between diabetes and favorable outcomes, from an HR of 0.86 in multivariable analysis to an HR of 0.76 in strati ed analysis, suggesting that the insulin use further reduced 28-day mortality in diabetic patients. The underlying mechanisms have been partially explained in previous studies with dysfunctional autophagy in critically ill patients, which plays a key role in both host defense and cell survival [37,38]; insulin not only plays a role in glucose regulation but also inhibits the autophagic catabolic process [39]. Still, the occurrence of hypoglycemia resulting from the use of insulin should not be ignored. Although the incidence of hypoglycemia was high in the two Leuven studies [4,5], the condition of patients who experienced hypoglycemia did not worsen when compared to that in who did not experience hypoglycemia. The use of intensive insulin therapy to lower the blood glucose level to normal values requires careful monitoring of blood glucose, as classical neurological symptoms can be offset by sedation or underlying mental status disorders.
In terms of glucose concentrations, glycemic control in critically ill patients is still an area of considerable concern. The initial recognition of the potentially detrimental effects of hyperglycemia prompted a sequence of studies that targeted intensive insulin treatment strategies with the goal of tight glycemic control. With the accumulation of knowledge, however, there have been mixed results regarding interventions for intensive insulin therapy, i.e., the excessive pursuit of tight glycemic control in critically ill patients is exactly a counterproductive step [4-7, 14-19, 40-44]. On the basis of the data from the two Leuven studies, it was considered practical to achieve blood glucose levels of 80-110 mg/dl (rather than 180-200 mg/dl) [4,5]. The NICE-SUGAR study concluded that a blood glucose target of 180 mg/dl or less was less likely to result in mortality than a target of 81-108 mg/dl [7]. Krinsley et al. adopted different glycemic control strategies on the basis of diabetes status and hemoglobin A1c (HbA1c) levels in critically ill patients and found that a blood glucose level of 80-140 mg/dl was safe and effective in patients without diabetes and in those with diabetes but with a low HbA1c level; however, for patients with diabetes and HbA1c levels greater than 7%, the glycemic target remains ambiguous [19]. These ndings mean that the moderate glycemic control strategy has been widely established in critically ill patients. Additionally, the quality of glycemic control can have an impact on clinical outcomes [20,45]. Glycemic variability has been found to be an independent risk factor for adverse outcomes in critically ill patients [15,[46][47][48]. Uyttendaele et al. found that the quality of glycemic control in critically ill patients is related prognostically rather than because of the metabolic status [20]. However, the study conducted by Krinsley et al. suggested that diabetes status modulates glycemic control and mortality, as shown by the fact that diabetic patients may bene t from a higher glycemic target range compared to critically ill patients without diabetes [15,48]. In our opinion, diabetes affects organisms in many aspects, not only resulting in abnormal blood glucose levels but also in the in ammatory response to the disease itself and to factors such as trauma, which can disrupt insulin sensitivity. However, the quality of glycemic control also affects blood glucose levels, with hypoglycemia, hyperglycemia, and glucose variability having deleterious effects on clinical outcomes. Currently, an accurate answer regarding optimal blood glucose concentrations remains elusive. Moreover, the complexity and variability of conducting large samples of clinical trials is well known. Our ndings were consistent with the "personalized" treatment strategy for patients with and without diabetes. In our study, for patients without diabetes, the glucose concentration corresponding to the lowest 28-day mortality was 101.75 mg/dl (95% CI 94.64-105.80 mg/dl), whereas for critically ill patients with diabetes, hyperglycemia did not signi cantly increase the 28-day mortality, and they even bene ted from a higher blood glucose level (up to 200 mg/dl), with the exception of patients admitted to the CCU and CSRU. Our study primarily established an optimal threshold for the glycemic range by retrospective analysis of a large sample, which may potentially inform the practice of glycemic control and treatment strategies in critically ill patients. Nevertheless, it should be noted that we used blood glucose concentrations at ICU admission, and we acknowledged that this might not be fully extrapolated to the optimal range of glycemic control.
There are limitations of our study as well. First, we attempted to obtain information on the plasma glucose levels at ICU admission to eliminate the in uence of interventions on glucose levels, but we were unable to de nitively state whether the interventions that the patients received before ICU admission, such as intravenous uid administration and steroid hormone injection, affected glucose levels. Second, as no data on HbA1c levels are available yet, we cannot exclude the possibility of new-onset diabetes.
Indeed, in previous studies on measuring HbA1c in patients without diabetes, it was found that 5.5%, 6.8%, and 9.3% of critically ill patients had higher than normal HbA1c levels [49][50][51], con rming that certain patients may have had undetected diabetes before ICU admission. In addition, it should be emphasized that our study did not differentiate the patients' diabetes type (e.g., type 1 or type 2). Third, we were unable to obtain information on the duration, severity, and complications of diabetes, as well as medication prescribed and therefore could not measure the impact of these factors on the outcomes. We used different models, including inverse-probability-weighted analysis, to investigate the independent role of diabetes and the clinical outcomes, but as with all retrospective studies, it was possible that residual confounders may exist. However, these clinical and electronic data were prospectively collected and independently measured, which makes them not easily amenable to manipulation. In addition, with the calculation of E-value to quantify the potential impact of unmeasured confounders, we found that unmeasured confounders are not likely to contribute to the overall effect. Fourth, we should be cautious in interpreting these results, as the results of a correlation analysis should not be mistaken for proof of causality. Finally, as the single-center study design results in reduced external validity, the aspects of glycemic control strategies and mortality reduction differing between critically ill patients with and without diabetes warrant prospective studies that can address the aforementioned limitations.

Conclusions
From this retrospective review of the prospectively collected data, the non-detrimental effect of diabetes on the short-term prognosis of critically ill patients was further con rmed, which would reduce 28-day mortality by approximately 29%. Furthermore, for non-diabetic patients, the glucose level corresponding to the lowest 28-day mortality was 101.75 mg/dl (95% CI 94.64-105.80 mg/dl), and hypoglycemic or hyperglycemic events should be avoided as much as possible, especially in patients admitted to the SICU, CSRU, and CCU. Moreover, for diabetic patients other than those admitted in the CCU and CSRU, elevated blood glucose levels do not appear to be associated with a poor prognosis and could bene t the patients as well. Finally, clinicians should be particularly attentive to hypoglycemic events in critically ill patients without diabetes in the SICU and to hyperglycemic events in critically ill patients in the CCU and CSRU, regardless of the presence of diabetes.