Silvio E. Inzucchi, MD

Setting

During the study period, YaleNew Haven Hospital (YNHH) was an urban academic medical center in New Haven, Connecticut, with over 950 beds and an average of approximately 32,000 annual adult nonobstetric admissions. YNHH conducted a variety of inpatient glucose control initiatives during the study period. The surgical ICU began an informal medical teamdirected insulin infusion protocol in 2000 to 2001. In 2002, the medical ICU instituted a formal insulin infusion protocol with a target of 100 to 140 mg/dL, which spread to remaining hospital ICUs by the end of 2003. In 2005, YNHH launched a consultative inpatient diabetes management team to assist clinicians in controlling glucose in non‐ICU patients with diabetes. This team covered approximately 10 to 15 patients at a time and consisted of an advanced‐practice nurse practitioner, a supervising endocrinologist and endocrinology fellow, and a nurse educator to provide diabetic teaching. Additionally, in 2005, basal‐boluscorrection insulin order sets became available. The surgical ICU implemented a stringent insulin infusion protocol with target glucose of 80 to 110 mg/dL in 2006, but relaxed it (goal 80150 mg/dL) in 2007. Similarly, in 2006, YNHH made ICU insulin infusion recommendations more stringent in remaining ICUs (goal 90130 mg/dL), but relaxed them in 2010 (goal 120160 mg/dL), based on emerging data from clinical trials and prevailing national guidelines.

Participants and Data Sources

We included all adult, nonobstetric discharges from YNHH between January 1, 2000 and December 31, 2010. Repeat visits by the same patient were linked by medical record number. We obtained data from YNHH administrative billing, laboratory, and point‐of‐care capillary blood glucose databases. The Yale Human Investigation Committee approved our study design and granted a Health Insurance Portability and Accountability Act waiver and a waiver of patient consent.

Variables

Our primary endpoint was in‐hospital mortality. The primary exposure of interest was whether a patient had diabetes mellitus, defined as the presence of International Classification of Diseases, Ninth Revision codes 249.x, 250.x, V4585, V5391, or V6546 in any of the primary or secondary diagnosis codes in the index admission, or in any hospital encounter in the year prior to the index admission.

We assessed 2 effect‐modifying variables: ICU status (as measured by a charge for at least 1 night in the ICU) and service assignment to surgery (including neurosurgery and orthopedics), compared to medicine (including neurology). Independent explanatory variables included time between the start of the study and patient admission (measured as days/365), diabetes status, inpatient glucose control, and long‐term glucose control (as measured by hemoglobin A1c at any time in the 180 days prior to hospital admission in order to have adequate sample size). We assessed inpatient blood glucose control through point‐of‐care blood glucose meters (OneTouch SureStep; LifeScan, Inc., Milipitas, CA) at YNHH. We used 4 validated measures of inpatient glucose control: the proportion of days in each hospitalization in which there was any hypoglycemic episode (blood glucose value <70 mg/dL), the proportion of days in which there was any severely hyperglycemic episode (blood glucose value >299 mg/dL), the proportion of days in which mean blood glucose was considered to be within adequate control (all blood glucose values between 70 and 179 mg/dL), and the standard deviation of mean glucose during hospitalization as a measure of glycemic variability.[28]

Covariates included gender, age at time of admission, length of stay in days, race (defined by hospital registration), payer, Elixhauser comorbidity dummy variables (revised to exclude diabetes and to use only secondary diagnosis codes),[29] and primary discharge diagnosis grouped using Clinical Classifications Software,[30] based on established associations with in‐hospital mortality.

Statistical Analysis

We summarized demographic characteristics numerically and graphically for patients with and without diabetes and compared them using [2] and t tests. We summarized changes in inpatient and outpatient measures of glucose control over time numerically and graphically, and compared across years using the Wilcoxon rank sum test adjusted for multiple hypothesis testing.

We stratified all analyses first by ICU status and then by service assignment (medicine vs surgery). Statistical analyses within each stratum paralleled our previous approach to the full study cohort.[10] Taking each stratum separately (ie, only ICU patients or only medicine patients), we used a difference‐in‐differences approach comparing changes over time in in‐hospital mortality among patients with diabetes compared to those without diabetes. This approach enabled us to determine whether patients with diabetes had a different time trend in risk of in‐hospital mortality than those without diabetes. That is, for each stratum, we constructed multivariate logistic regression models including time in years, diabetes status, and the interaction between time and diabetes status as well as the aforementioned covariates. We calculated odds of death and confidence intervals for each additional year for patients with diabetes by exponentiating the sum of parameter estimates for time and the diabetes‐time interaction term. We evaluated all 2‐way interactions between year or diabetes status and the covariates in a multiple degree of freedom likelihood ratio test. We investigated nonlinearity of the relation between mortality and time by evaluating first and second‐order polynomials.

Because we found a significant decline in mortality risk for patients with versus without diabetes among ICU patients but not among non‐ICU patients, and because service assignment was not found to be an effect modifier, we then limited our sample to ICU patients with diabetes to better understand the role of inpatient and outpatient glucose control in accounting for observed mortality trends. First, we determined the relation between the measures of inpatient glucose control and changes in mortality over time using logistic regression. Then, we repeated this analysis in the subsets of patients who had inpatient glucose data and both inpatient and outpatient glycemic control data, adding inpatient and outpatient measures sequentially. Given the high level of missing outpatient glycemic control data, we compared demographic characteristics for diabetic ICU patients with and without such data using [2] and t tests, and found that patients with data were younger and less likely to be white and had longer mean length of stay, slightly worse performance on several measures of inpatient glucose control, and lower mortality (see Supporting Table 1 in the online version of this article).

To explore the effects of dependence among observations from patients with multiple encounters, we compared parameter estimates derived from a model with all patient encounters (including repeated admissions for the same patient) with those from a model with a randomly sampled single visit per patient, and observed that there was no difference in parameter estimates between the 2 classes of models. For all analyses, we used a type I error of 5% (2 sided) to test for statistical significance using SAS version 9.3 (SAS Institute, Cary, NC) or R software (http://CRAN.R‐project.org).

Mortality Trends Within Strata

Among ICU patients, the overall mortality rate was 9.9%: 10.5% of patients with diabetes and 9.8% of patients without diabetes. Among non‐ICU patients, the overall mortality rate was 0.8%: 0.9% of patients with diabetes and 0.7% of patients without diabetes.

Among medical patients, the overall mortality rate was 2.9%: 3.1% of patients with diabetes and 2.8% of patients without diabetes. Among surgical patients, the overall mortality rate was 1.4%: 1.8% of patients with diabetes and 1.4% of patients without diabetes. Figure 1 shows quarterly in‐hospital mortality for patients with and without diabetes from 2000 to 2010 stratified by ICU status and by service assignment.

Figure 1

Table 2 describes the difference‐in‐differences regression analyses, stratified by ICU status and service assignment. Among ICU patients (Table 2, model 1), each successive year was associated with a 2.6% relative reduction in the adjusted odds of mortality (odds ratio [OR]: 0.974, 95% confidence interval [CI]: 0.963‐0.985) for patients without diabetes compared to a 7.8% relative reduction for those with diabetes (OR: 0.923, 95% CI: 0.906‐0.940). In other words, patients with diabetes compared to patients without diabetes had a significantly greater decline in odds of adjusted mortality of 5.3% per year (OR: 0.947, 95% CI: 0.927‐0.967). As a result, the adjusted odds of mortality among patients with versus without diabetes decreased from 1.352 in 2000 to 0.772 in 2010.

Among non‐ICU patients (Table 2, model 2), each successive year was associated with a 7.5% relative reduction in the adjusted odds of mortality (OR: 0.925, 95% CI: 0.909‐0.940) for patients without diabetes compared to a 9.6% relative reduction for those with diabetes (OR: 0.904, 95% CI: 0.879‐0.929); this greater decline in odds of adjusted mortality of 2.3% per year (OR: 0.977, 95% CI: 0.946‐1.008; P=0.148) was not statistically significant.

We found greater decline in odds of mortality among patients with diabetes than among patients without diabetes over time in both medical patients (3.9% greater decline per year; OR: 0.961, 95% CI: 0.942‐0.980) and surgical patients (4.5% greater decline per year; OR: 0.955, 95% CI: 0.918‐0.994), without a difference between the 2. Detailed results are shown in Table 2, models 3 and 4.

Glycemic Control

Among ICU patients with diabetes (N=14,364), at least 2 inpatient point‐of‐care glucose readings were available for 13,136 (91.5%), with a mean of 4.67 readings per day, whereas hemoglobin A1c data were available for only 5321 patients (37.0%). Both inpatient glucose data and hemoglobin A1c were available for 4989 patients (34.7%). Figure 2 shows trends in inpatient and outpatient glycemic control measures among ICU patients with diabetes over the study period. Mean hemoglobin A1c decreased from 7.7 in 2000 to 7.3 in 2010. Mean hospitalization glucose began at 187.2, reached a nadir of 162.4 in the third quarter (Q3) of 2007, and rose subsequently to 174.4 with loosened glucose control targets. Standard deviation of mean glucose and percentage of patient‐days with a severe hyperglycemic episode followed a similar pattern, though with nadirs in Q4 2007 and Q2 2008, respectively, whereas percentage of patient‐days with a hypoglycemic episode rose from 1.46% in 2000, peaked at 3.00% in Q3 2005, and returned to 2.15% in 2010. All changes in glucose control are significant with P<0.001.

Figure 2

Mortality Trends and Glycemic Control

To determine whether glucose control explained the excess decline in odds of mortality among patients with diabetes in the ICU, we restricted our sample to ICU patients with diabetes and examined the association of diabetes with mortality after including measures of glucose control.

We first verified that the overall adjusted mortality trend among ICU patients with diabetes for whom we had measures of inpatient glucose control was similar to that of the full sample of ICU patients with diabetes. Similar to the full sample, we found that the adjusted excess odds of death significantly declined by a relative 7.3% each successive year (OR: 0.927, 95% CI: 0.907‐0.947; Table 3, model 1). We then included measures of inpatient glucose control in the model and found, as expected, that a higher percentage of days with severe hyperglycemia and with hypoglycemia was associated with an increased odds of death (P<0.001 for both; Table 3, model 2). Nonetheless, after including measures of inpatient glucose control, we found that the rate of change of excess odds of death for patients with diabetes was unchanged (OR: 0.926, 95% CI: 0.905‐0.947).

We then restricted our sample to patients with diabetes with both inpatient and outpatient glycemic control data and found that, in this subpopulation, the adjusted excess odds of death among patients with diabetes relative to those without significantly declined by a relative 4.2% each progressive year (OR: 0.958, 95% CI: 0.918‐0.998; Table 3, model 3). Including measures of inpatient glucose control in the model did not significantly change the rate of change of excess odds of death (OR: 0.956, 95% CI: 0.916‐0.997; Table 3, model 4), nor did including both measures of inpatient and outpatient glycemic control (OR: 0.953, 95% CI: 0.914‐0.994; Table 3, model 5).

Setting

During the study period, YaleNew Haven Hospital (YNHH) was an urban academic medical center in New Haven, Connecticut, with over 950 beds and an average of approximately 32,000 annual adult nonobstetric admissions. YNHH conducted a variety of inpatient glucose control initiatives during the study period. The surgical ICU began an informal medical teamdirected insulin infusion protocol in 2000 to 2001. In 2002, the medical ICU instituted a formal insulin infusion protocol with a target of 100 to 140 mg/dL, which spread to remaining hospital ICUs by the end of 2003. In 2005, YNHH launched a consultative inpatient diabetes management team to assist clinicians in controlling glucose in non‐ICU patients with diabetes. This team covered approximately 10 to 15 patients at a time and consisted of an advanced‐practice nurse practitioner, a supervising endocrinologist and endocrinology fellow, and a nurse educator to provide diabetic teaching. Additionally, in 2005, basal‐boluscorrection insulin order sets became available. The surgical ICU implemented a stringent insulin infusion protocol with target glucose of 80 to 110 mg/dL in 2006, but relaxed it (goal 80150 mg/dL) in 2007. Similarly, in 2006, YNHH made ICU insulin infusion recommendations more stringent in remaining ICUs (goal 90130 mg/dL), but relaxed them in 2010 (goal 120160 mg/dL), based on emerging data from clinical trials and prevailing national guidelines.

Participants and Data Sources

We included all adult, nonobstetric discharges from YNHH between January 1, 2000 and December 31, 2010. Repeat visits by the same patient were linked by medical record number. We obtained data from YNHH administrative billing, laboratory, and point‐of‐care capillary blood glucose databases. The Yale Human Investigation Committee approved our study design and granted a Health Insurance Portability and Accountability Act waiver and a waiver of patient consent.

Variables

Our primary endpoint was in‐hospital mortality. The primary exposure of interest was whether a patient had diabetes mellitus, defined as the presence of International Classification of Diseases, Ninth Revision codes 249.x, 250.x, V4585, V5391, or V6546 in any of the primary or secondary diagnosis codes in the index admission, or in any hospital encounter in the year prior to the index admission.

We assessed 2 effect‐modifying variables: ICU status (as measured by a charge for at least 1 night in the ICU) and service assignment to surgery (including neurosurgery and orthopedics), compared to medicine (including neurology). Independent explanatory variables included time between the start of the study and patient admission (measured as days/365), diabetes status, inpatient glucose control, and long‐term glucose control (as measured by hemoglobin A1c at any time in the 180 days prior to hospital admission in order to have adequate sample size). We assessed inpatient blood glucose control through point‐of‐care blood glucose meters (OneTouch SureStep; LifeScan, Inc., Milipitas, CA) at YNHH. We used 4 validated measures of inpatient glucose control: the proportion of days in each hospitalization in which there was any hypoglycemic episode (blood glucose value <70 mg/dL), the proportion of days in which there was any severely hyperglycemic episode (blood glucose value >299 mg/dL), the proportion of days in which mean blood glucose was considered to be within adequate control (all blood glucose values between 70 and 179 mg/dL), and the standard deviation of mean glucose during hospitalization as a measure of glycemic variability.[28]

Covariates included gender, age at time of admission, length of stay in days, race (defined by hospital registration), payer, Elixhauser comorbidity dummy variables (revised to exclude diabetes and to use only secondary diagnosis codes),[29] and primary discharge diagnosis grouped using Clinical Classifications Software,[30] based on established associations with in‐hospital mortality.

Statistical Analysis

We summarized demographic characteristics numerically and graphically for patients with and without diabetes and compared them using [2] and t tests. We summarized changes in inpatient and outpatient measures of glucose control over time numerically and graphically, and compared across years using the Wilcoxon rank sum test adjusted for multiple hypothesis testing.

We stratified all analyses first by ICU status and then by service assignment (medicine vs surgery). Statistical analyses within each stratum paralleled our previous approach to the full study cohort.[10] Taking each stratum separately (ie, only ICU patients or only medicine patients), we used a difference‐in‐differences approach comparing changes over time in in‐hospital mortality among patients with diabetes compared to those without diabetes. This approach enabled us to determine whether patients with diabetes had a different time trend in risk of in‐hospital mortality than those without diabetes. That is, for each stratum, we constructed multivariate logistic regression models including time in years, diabetes status, and the interaction between time and diabetes status as well as the aforementioned covariates. We calculated odds of death and confidence intervals for each additional year for patients with diabetes by exponentiating the sum of parameter estimates for time and the diabetes‐time interaction term. We evaluated all 2‐way interactions between year or diabetes status and the covariates in a multiple degree of freedom likelihood ratio test. We investigated nonlinearity of the relation between mortality and time by evaluating first and second‐order polynomials.

Because we found a significant decline in mortality risk for patients with versus without diabetes among ICU patients but not among non‐ICU patients, and because service assignment was not found to be an effect modifier, we then limited our sample to ICU patients with diabetes to better understand the role of inpatient and outpatient glucose control in accounting for observed mortality trends. First, we determined the relation between the measures of inpatient glucose control and changes in mortality over time using logistic regression. Then, we repeated this analysis in the subsets of patients who had inpatient glucose data and both inpatient and outpatient glycemic control data, adding inpatient and outpatient measures sequentially. Given the high level of missing outpatient glycemic control data, we compared demographic characteristics for diabetic ICU patients with and without such data using [2] and t tests, and found that patients with data were younger and less likely to be white and had longer mean length of stay, slightly worse performance on several measures of inpatient glucose control, and lower mortality (see Supporting Table 1 in the online version of this article).

To explore the effects of dependence among observations from patients with multiple encounters, we compared parameter estimates derived from a model with all patient encounters (including repeated admissions for the same patient) with those from a model with a randomly sampled single visit per patient, and observed that there was no difference in parameter estimates between the 2 classes of models. For all analyses, we used a type I error of 5% (2 sided) to test for statistical significance using SAS version 9.3 (SAS Institute, Cary, NC) or R software (http://CRAN.R‐project.org).

Mortality Trends Within Strata

Among ICU patients, the overall mortality rate was 9.9%: 10.5% of patients with diabetes and 9.8% of patients without diabetes. Among non‐ICU patients, the overall mortality rate was 0.8%: 0.9% of patients with diabetes and 0.7% of patients without diabetes.

Among medical patients, the overall mortality rate was 2.9%: 3.1% of patients with diabetes and 2.8% of patients without diabetes. Among surgical patients, the overall mortality rate was 1.4%: 1.8% of patients with diabetes and 1.4% of patients without diabetes. Figure 1 shows quarterly in‐hospital mortality for patients with and without diabetes from 2000 to 2010 stratified by ICU status and by service assignment.

Figure 1

Table 2 describes the difference‐in‐differences regression analyses, stratified by ICU status and service assignment. Among ICU patients (Table 2, model 1), each successive year was associated with a 2.6% relative reduction in the adjusted odds of mortality (odds ratio [OR]: 0.974, 95% confidence interval [CI]: 0.963‐0.985) for patients without diabetes compared to a 7.8% relative reduction for those with diabetes (OR: 0.923, 95% CI: 0.906‐0.940). In other words, patients with diabetes compared to patients without diabetes had a significantly greater decline in odds of adjusted mortality of 5.3% per year (OR: 0.947, 95% CI: 0.927‐0.967). As a result, the adjusted odds of mortality among patients with versus without diabetes decreased from 1.352 in 2000 to 0.772 in 2010.

Among non‐ICU patients (Table 2, model 2), each successive year was associated with a 7.5% relative reduction in the adjusted odds of mortality (OR: 0.925, 95% CI: 0.909‐0.940) for patients without diabetes compared to a 9.6% relative reduction for those with diabetes (OR: 0.904, 95% CI: 0.879‐0.929); this greater decline in odds of adjusted mortality of 2.3% per year (OR: 0.977, 95% CI: 0.946‐1.008; P=0.148) was not statistically significant.

We found greater decline in odds of mortality among patients with diabetes than among patients without diabetes over time in both medical patients (3.9% greater decline per year; OR: 0.961, 95% CI: 0.942‐0.980) and surgical patients (4.5% greater decline per year; OR: 0.955, 95% CI: 0.918‐0.994), without a difference between the 2. Detailed results are shown in Table 2, models 3 and 4.

Glycemic Control

Among ICU patients with diabetes (N=14,364), at least 2 inpatient point‐of‐care glucose readings were available for 13,136 (91.5%), with a mean of 4.67 readings per day, whereas hemoglobin A1c data were available for only 5321 patients (37.0%). Both inpatient glucose data and hemoglobin A1c were available for 4989 patients (34.7%). Figure 2 shows trends in inpatient and outpatient glycemic control measures among ICU patients with diabetes over the study period. Mean hemoglobin A1c decreased from 7.7 in 2000 to 7.3 in 2010. Mean hospitalization glucose began at 187.2, reached a nadir of 162.4 in the third quarter (Q3) of 2007, and rose subsequently to 174.4 with loosened glucose control targets. Standard deviation of mean glucose and percentage of patient‐days with a severe hyperglycemic episode followed a similar pattern, though with nadirs in Q4 2007 and Q2 2008, respectively, whereas percentage of patient‐days with a hypoglycemic episode rose from 1.46% in 2000, peaked at 3.00% in Q3 2005, and returned to 2.15% in 2010. All changes in glucose control are significant with P<0.001.

Figure 2

Mortality Trends and Glycemic Control

To determine whether glucose control explained the excess decline in odds of mortality among patients with diabetes in the ICU, we restricted our sample to ICU patients with diabetes and examined the association of diabetes with mortality after including measures of glucose control.

We first verified that the overall adjusted mortality trend among ICU patients with diabetes for whom we had measures of inpatient glucose control was similar to that of the full sample of ICU patients with diabetes. Similar to the full sample, we found that the adjusted excess odds of death significantly declined by a relative 7.3% each successive year (OR: 0.927, 95% CI: 0.907‐0.947; Table 3, model 1). We then included measures of inpatient glucose control in the model and found, as expected, that a higher percentage of days with severe hyperglycemia and with hypoglycemia was associated with an increased odds of death (P<0.001 for both; Table 3, model 2). Nonetheless, after including measures of inpatient glucose control, we found that the rate of change of excess odds of death for patients with diabetes was unchanged (OR: 0.926, 95% CI: 0.905‐0.947).

We then restricted our sample to patients with diabetes with both inpatient and outpatient glycemic control data and found that, in this subpopulation, the adjusted excess odds of death among patients with diabetes relative to those without significantly declined by a relative 4.2% each progressive year (OR: 0.958, 95% CI: 0.918‐0.998; Table 3, model 3). Including measures of inpatient glucose control in the model did not significantly change the rate of change of excess odds of death (OR: 0.956, 95% CI: 0.916‐0.997; Table 3, model 4), nor did including both measures of inpatient and outpatient glycemic control (OR: 0.953, 95% CI: 0.914‐0.994; Table 3, model 5).

Setting

During the study period, YaleNew Haven Hospital (YNHH) was an urban academic medical center in New Haven, Connecticut, with over 950 beds and an average of approximately 32,000 annual adult nonobstetric admissions. YNHH conducted a variety of inpatient glucose control initiatives during the study period. The surgical ICU began an informal medical teamdirected insulin infusion protocol in 2000 to 2001. In 2002, the medical ICU instituted a formal insulin infusion protocol with a target of 100 to 140 mg/dL, which spread to remaining hospital ICUs by the end of 2003. In 2005, YNHH launched a consultative inpatient diabetes management team to assist clinicians in controlling glucose in non‐ICU patients with diabetes. This team covered approximately 10 to 15 patients at a time and consisted of an advanced‐practice nurse practitioner, a supervising endocrinologist and endocrinology fellow, and a nurse educator to provide diabetic teaching. Additionally, in 2005, basal‐boluscorrection insulin order sets became available. The surgical ICU implemented a stringent insulin infusion protocol with target glucose of 80 to 110 mg/dL in 2006, but relaxed it (goal 80150 mg/dL) in 2007. Similarly, in 2006, YNHH made ICU insulin infusion recommendations more stringent in remaining ICUs (goal 90130 mg/dL), but relaxed them in 2010 (goal 120160 mg/dL), based on emerging data from clinical trials and prevailing national guidelines.

Participants and Data Sources

We included all adult, nonobstetric discharges from YNHH between January 1, 2000 and December 31, 2010. Repeat visits by the same patient were linked by medical record number. We obtained data from YNHH administrative billing, laboratory, and point‐of‐care capillary blood glucose databases. The Yale Human Investigation Committee approved our study design and granted a Health Insurance Portability and Accountability Act waiver and a waiver of patient consent.

Variables

Our primary endpoint was in‐hospital mortality. The primary exposure of interest was whether a patient had diabetes mellitus, defined as the presence of International Classification of Diseases, Ninth Revision codes 249.x, 250.x, V4585, V5391, or V6546 in any of the primary or secondary diagnosis codes in the index admission, or in any hospital encounter in the year prior to the index admission.

We assessed 2 effect‐modifying variables: ICU status (as measured by a charge for at least 1 night in the ICU) and service assignment to surgery (including neurosurgery and orthopedics), compared to medicine (including neurology). Independent explanatory variables included time between the start of the study and patient admission (measured as days/365), diabetes status, inpatient glucose control, and long‐term glucose control (as measured by hemoglobin A1c at any time in the 180 days prior to hospital admission in order to have adequate sample size). We assessed inpatient blood glucose control through point‐of‐care blood glucose meters (OneTouch SureStep; LifeScan, Inc., Milipitas, CA) at YNHH. We used 4 validated measures of inpatient glucose control: the proportion of days in each hospitalization in which there was any hypoglycemic episode (blood glucose value <70 mg/dL), the proportion of days in which there was any severely hyperglycemic episode (blood glucose value >299 mg/dL), the proportion of days in which mean blood glucose was considered to be within adequate control (all blood glucose values between 70 and 179 mg/dL), and the standard deviation of mean glucose during hospitalization as a measure of glycemic variability.[28]

Covariates included gender, age at time of admission, length of stay in days, race (defined by hospital registration), payer, Elixhauser comorbidity dummy variables (revised to exclude diabetes and to use only secondary diagnosis codes),[29] and primary discharge diagnosis grouped using Clinical Classifications Software,[30] based on established associations with in‐hospital mortality.

Statistical Analysis

We summarized demographic characteristics numerically and graphically for patients with and without diabetes and compared them using [2] and t tests. We summarized changes in inpatient and outpatient measures of glucose control over time numerically and graphically, and compared across years using the Wilcoxon rank sum test adjusted for multiple hypothesis testing.

We stratified all analyses first by ICU status and then by service assignment (medicine vs surgery). Statistical analyses within each stratum paralleled our previous approach to the full study cohort.[10] Taking each stratum separately (ie, only ICU patients or only medicine patients), we used a difference‐in‐differences approach comparing changes over time in in‐hospital mortality among patients with diabetes compared to those without diabetes. This approach enabled us to determine whether patients with diabetes had a different time trend in risk of in‐hospital mortality than those without diabetes. That is, for each stratum, we constructed multivariate logistic regression models including time in years, diabetes status, and the interaction between time and diabetes status as well as the aforementioned covariates. We calculated odds of death and confidence intervals for each additional year for patients with diabetes by exponentiating the sum of parameter estimates for time and the diabetes‐time interaction term. We evaluated all 2‐way interactions between year or diabetes status and the covariates in a multiple degree of freedom likelihood ratio test. We investigated nonlinearity of the relation between mortality and time by evaluating first and second‐order polynomials.

Because we found a significant decline in mortality risk for patients with versus without diabetes among ICU patients but not among non‐ICU patients, and because service assignment was not found to be an effect modifier, we then limited our sample to ICU patients with diabetes to better understand the role of inpatient and outpatient glucose control in accounting for observed mortality trends. First, we determined the relation between the measures of inpatient glucose control and changes in mortality over time using logistic regression. Then, we repeated this analysis in the subsets of patients who had inpatient glucose data and both inpatient and outpatient glycemic control data, adding inpatient and outpatient measures sequentially. Given the high level of missing outpatient glycemic control data, we compared demographic characteristics for diabetic ICU patients with and without such data using [2] and t tests, and found that patients with data were younger and less likely to be white and had longer mean length of stay, slightly worse performance on several measures of inpatient glucose control, and lower mortality (see Supporting Table 1 in the online version of this article).

To explore the effects of dependence among observations from patients with multiple encounters, we compared parameter estimates derived from a model with all patient encounters (including repeated admissions for the same patient) with those from a model with a randomly sampled single visit per patient, and observed that there was no difference in parameter estimates between the 2 classes of models. For all analyses, we used a type I error of 5% (2 sided) to test for statistical significance using SAS version 9.3 (SAS Institute, Cary, NC) or R software (http://CRAN.R‐project.org).

Mortality Trends Within Strata

Among ICU patients, the overall mortality rate was 9.9%: 10.5% of patients with diabetes and 9.8% of patients without diabetes. Among non‐ICU patients, the overall mortality rate was 0.8%: 0.9% of patients with diabetes and 0.7% of patients without diabetes.

Among medical patients, the overall mortality rate was 2.9%: 3.1% of patients with diabetes and 2.8% of patients without diabetes. Among surgical patients, the overall mortality rate was 1.4%: 1.8% of patients with diabetes and 1.4% of patients without diabetes. Figure 1 shows quarterly in‐hospital mortality for patients with and without diabetes from 2000 to 2010 stratified by ICU status and by service assignment.

Figure 1

Table 2 describes the difference‐in‐differences regression analyses, stratified by ICU status and service assignment. Among ICU patients (Table 2, model 1), each successive year was associated with a 2.6% relative reduction in the adjusted odds of mortality (odds ratio [OR]: 0.974, 95% confidence interval [CI]: 0.963‐0.985) for patients without diabetes compared to a 7.8% relative reduction for those with diabetes (OR: 0.923, 95% CI: 0.906‐0.940). In other words, patients with diabetes compared to patients without diabetes had a significantly greater decline in odds of adjusted mortality of 5.3% per year (OR: 0.947, 95% CI: 0.927‐0.967). As a result, the adjusted odds of mortality among patients with versus without diabetes decreased from 1.352 in 2000 to 0.772 in 2010.

Among non‐ICU patients (Table 2, model 2), each successive year was associated with a 7.5% relative reduction in the adjusted odds of mortality (OR: 0.925, 95% CI: 0.909‐0.940) for patients without diabetes compared to a 9.6% relative reduction for those with diabetes (OR: 0.904, 95% CI: 0.879‐0.929); this greater decline in odds of adjusted mortality of 2.3% per year (OR: 0.977, 95% CI: 0.946‐1.008; P=0.148) was not statistically significant.

We found greater decline in odds of mortality among patients with diabetes than among patients without diabetes over time in both medical patients (3.9% greater decline per year; OR: 0.961, 95% CI: 0.942‐0.980) and surgical patients (4.5% greater decline per year; OR: 0.955, 95% CI: 0.918‐0.994), without a difference between the 2. Detailed results are shown in Table 2, models 3 and 4.

Glycemic Control

Among ICU patients with diabetes (N=14,364), at least 2 inpatient point‐of‐care glucose readings were available for 13,136 (91.5%), with a mean of 4.67 readings per day, whereas hemoglobin A1c data were available for only 5321 patients (37.0%). Both inpatient glucose data and hemoglobin A1c were available for 4989 patients (34.7%). Figure 2 shows trends in inpatient and outpatient glycemic control measures among ICU patients with diabetes over the study period. Mean hemoglobin A1c decreased from 7.7 in 2000 to 7.3 in 2010. Mean hospitalization glucose began at 187.2, reached a nadir of 162.4 in the third quarter (Q3) of 2007, and rose subsequently to 174.4 with loosened glucose control targets. Standard deviation of mean glucose and percentage of patient‐days with a severe hyperglycemic episode followed a similar pattern, though with nadirs in Q4 2007 and Q2 2008, respectively, whereas percentage of patient‐days with a hypoglycemic episode rose from 1.46% in 2000, peaked at 3.00% in Q3 2005, and returned to 2.15% in 2010. All changes in glucose control are significant with P<0.001.

Figure 2

Mortality Trends and Glycemic Control

To determine whether glucose control explained the excess decline in odds of mortality among patients with diabetes in the ICU, we restricted our sample to ICU patients with diabetes and examined the association of diabetes with mortality after including measures of glucose control.

We first verified that the overall adjusted mortality trend among ICU patients with diabetes for whom we had measures of inpatient glucose control was similar to that of the full sample of ICU patients with diabetes. Similar to the full sample, we found that the adjusted excess odds of death significantly declined by a relative 7.3% each successive year (OR: 0.927, 95% CI: 0.907‐0.947; Table 3, model 1). We then included measures of inpatient glucose control in the model and found, as expected, that a higher percentage of days with severe hyperglycemia and with hypoglycemia was associated with an increased odds of death (P<0.001 for both; Table 3, model 2). Nonetheless, after including measures of inpatient glucose control, we found that the rate of change of excess odds of death for patients with diabetes was unchanged (OR: 0.926, 95% CI: 0.905‐0.947).

We then restricted our sample to patients with diabetes with both inpatient and outpatient glycemic control data and found that, in this subpopulation, the adjusted excess odds of death among patients with diabetes relative to those without significantly declined by a relative 4.2% each progressive year (OR: 0.958, 95% CI: 0.918‐0.998; Table 3, model 3). Including measures of inpatient glucose control in the model did not significantly change the rate of change of excess odds of death (OR: 0.956, 95% CI: 0.916‐0.997; Table 3, model 4), nor did including both measures of inpatient and outpatient glycemic control (OR: 0.953, 95% CI: 0.914‐0.994; Table 3, model 5).

Defining the Target Patient Population

The first decision to be made is which patients to include in your analysis. Choices include the following:

• 1

  Patients with a discharge diagnosis of diabetes: this group has face validity and intuitive appeal, is easy to identify retrospectively, and may capture some untested/untreated diabetics, but will miss patients with otherwise undiagnosed diabetes and stress hyperglycemia. It is also subject to the variable accuracy of billing codes.

• 2

  Patients with a certain number of point‐of‐care (POC) glucose measurements: this group is also easy to identify, easy to measure, and will include patients with hyperglycemia without a previous diagnosis of diabetes, but will miss patients with untested/untreated hyperglycemia. Also, if glucose levels are checked on normoglycemic, nondiabetic patients, these values may dilute the overall assessment of glycemic control.

• 3

  Patients treated with insulin in the hospital: this is a good choice if the purpose is mainly drug safety and avoidance of hypoglycemia, but by definition excludes most untreated patients.

• 4

  Patients with 2 or more BG values (laboratory and/or POC) over a certain threshold (eg, >180 mg/dL). This will likely capture more patients with inpatient hyperglycemia, whether or not detected by the medical team, but is subject to wide variations in the frequency and timing of laboratory glucose testing, including whether or not the values are pre‐prandial (note that even preprandial POC glucose measurements are not always in fact fasting values).

Other considerations include the following:

• 1

  Are there natural patient subgroups that should be measured and analyzed separately because of different guidelines? For example, there probably should be separate/emndependent inclusion criteria and analyses for critical care and noncritical care units because their glycemic targets and management considerations differ.

• 2

  Which patients should be excluded? For example, if targeting subcutaneous insulin use in general hospitalized patients, one might eliminate those patients who are admitted specifically as the result of a diabetes emergency (eg, diabetic ketoacidosis [DKA] and hyperglycemic hyperosmolar state [HHS]), as their marked and prolonged hyperglycemia will skew BG data. Pregnant women should generally be excluded from broad‐based analyses or considered as a discrete category because they have very different targets for BG therapy. Patients with short lengths of stay may be less likely to benefit from tight glucose control and may also be considered for post hoc exclusion. One might also exclude patients with very few evaluable glucose readings (eg, fewer than 5) to ensure that measurement is meaningful for a given patient, keeping in mind that this may also exclude patients with undetected hyperglycemia, as mentioned above. Finally, patients receiving palliative care should also be considered for exclusion if feasible.

Recommendation: Do not limit analyses to only those patients with a diagnosis of diabetes or only those on insulin, which will lead to biased results.

• For noncritical care patients, we recommend a combined approach: adult patients with a diagnosis of diabetes (e. g. using diagnosis‐related group [DRG] codes 294 or 295 or International Classification of Diseases 9th edition [ICD9] codes 250.xx) or with hyperglycemia (eg, 2 or more random laboratory and/or point of care (POC) BG values >180 mg/dL or 2 or more fasting BG values >130 mg/dL), excluding patients with DKA or HHS or who are pregnant.

• For critical care units, we recommend either all patients, or patients with at least mild hyperglycemia (eg, 2 random glucose levels >140 mg/dL). Critical care patients with DKA, HHS, and pregnancy should be evaluated separately if possible.

Which Glucose Values to Include and Exclude

To answer this question, we first need to decide which method to use for BG measurement. There are several ways to measure BG, including the type of sample collected (capillary [fingerstick], arterial, and venous) and the technique used (central laboratory analyzing plasma, central laboratory analyzing whole blood [eg, from an arterial blood gas sample], glucose meter [usually calibrated to plasma], etc.). The POC (eg, capillary, glucose meter) glucose measurements alone are often preferred in the non‐ICU setting because laboratory plasma values generally provide little additional information and typically lower the mean glucose by including redundant fasting values.1 In critical care units, several different methods are often used together, and each merits inclusion. The inherent differences in calibration between the methods do not generally require separate analyses, especially given the frequency of testing in the ICU setting.

The next question is which values to include in analyses. In some situations, it may be most useful to focus on a certain period of hospitalization, such as the day of a procedure and the next 2 days in assessing the impact of the quality of perioperative care, or the first 14 days of a noncritical care stay to keep outliers for length of stay (LOS) from skewing the data. In the non‐ICU setting, it may be reasonable to exclude the first day of hospitalization, as early BG control is impacted by multiple variables beyond direct control of the clinician (eg, glucose control prior to admission, severity of presenting illness) and may not realistically reflect your interventions. (Keep in mind, however, that it may be useful to adjust for the admission glucose value in multivariable models given its importance to clinical outcomes and its strong relationship to subsequent inpatient glucose control.) However, in critical care units, it is reasonable to include the first day's readings in analyses given the high frequency of glucose measurements in this setting and the expectation that glucose control should be achieved within a few hours of starting an intravenous insulin infusion.

If feasible to do so with your institution's data capture methods, you may wish to select only the regularly scheduled (before each meal [qAC] and at bedtime [qHS], or every 6 hours [q6h]) glucose readings for inclusion in the summary data of glycemic control in the non‐ICU setting, thereby reducing bias caused by repeated measurements around extremes of glycemic excursions. An alternative in the non‐ICU setting is to censor glucose readings within 60 minutes of a previous reading.

Recommendation:

• In the non‐ICU setting, we recommend first looking at all POC glucose values and if possible repeating the analyses excluding hospital day 1 and hospital day 15 and beyond, and also excluding glucose values measured within 60 minutes of a previous value.

• In critical care units, we recommend evaluating all glucose readings used to guide care.

Units of Analysis

There are several different units of analysis, each with its own advantages and disadvantages:

• 1

  Glucose value: this is the simplest measure and the one with the most statistical power. All glucose values for all patients of interest comprise the denominator. A report might say, for example, that 1% of the 1000 glucose values were <70 mg/dL during a certain period or that the mean of all glucose values collected for the month from patients in noncritical care areas was 160 mg/dL. The potential disadvantages of this approach are that these analyses are less clinically relevant than patient‐level analyses and that patients with many glucose readings and long hospitalizations may skew the data.

• 2

  Patient (or the Patient Stay, [ie, the entire hospitalization]): all patients who are monitored make up the denominator. The numerator may be the percentage of patients with any hypoglycemia during their hospital stay or the percentage of patients achieving a certain mean glucose during their hospitalization, for example. This is inherently more clinically meaningful than using glucose value as a unit of analysis. A major disadvantage is not controlling for LOS effects. For example, a hospitalized patient with a long LOS is much more likely to be characterized as having at least 1 hypoglycemic value than is a patient with a shorter LOS. Another shortcoming is that this approach does not correct for uneven distribution of testing. A patient's mean glucose might be calculated on the basis of 8 glucose values on the first day of hospitalization, 4 on the second day, and 1 on the third day. Despite all these shortcomings, reporting by patient remains a popular and valid method of presenting glycemic control results, particularly when complemented by other views and refined to control for the number of readings per day.

• 3

  Monitored Patient‐Day: The denominator in this setting is the total number of days a patient glucose level is monitored. The benefits of this method have been described and advocated in the literature.1 As with patient‐level analyses, this measure will be more rigorous and meaningful if the BG measures to be evaluated have been standardized. Typical reports might include percentage of monitored days with any hypoglycemia, or percentage of monitored days with all glucose values in the desired range. This unit of analysis may be considered more difficult to generate and to interpret. On the other hand, it is clinically relevant, less biased by LOS effects, and may be considered the most actionable metric by clinicians. This method provides a good balance when presented with data organized by patient.

The following example uses all 3 units of measurement, in this case to determine the rate of hypoglycemia, demonstrating the different but complementary information that each method provides:

• In 1 month, 3900 POC glucose measurements were obtained from 286 patients representing 986 monitored patient‐days. With hypoglycemia defined as POC BG 60 mg/dL, the results showed the following:

• 50 of 3900 measurements (1.4%) were hypoglycemic 22 of 286 patients (7.7%) had 1 hypoglycemic episodes

• 40 of 986 monitored days (4.4%) had 1 hypoglycemic episodes.

The metric based on the number of glucose readings could be considered the least clinically relevant because it is unclear how many patients were affected; moreover, it may be based on variable testing patterns among patients, and could be influenced disproportionately by 1 patient with frequent hypoglycemia, many glucose readings, and/or a long LOS. One could argue that the patient‐stay metric is artificially elevated because a single hypoglycemic episode characterizes the entire stay as hypoglycemic. On the other hand, at least it acknowledges the number of patients affected by hypoglycemia. The patient‐day unit of analysis likely provides the most balanced view, one that is clinically relevant and measured over a standard period of time, and less biased by LOS and frequency of testing.

One way to express patient‐day glycemic control that deserves special mention is the patient‐day weighted mean. A mean glucose is calculated for each patient‐day, and then the mean is calculated across all patient‐days. The advantage of this approach is that it corrects for variation in the number of glucose readings each day; all hospital days are weighted equally.

Recommendation:

• In noncritical care units, we recommend a combination of patient‐day and patient‐stay measures.

• In critical care units, it is acceptable to also use glucose reading as the unit of measurement given more frequent and uniform data collection, but it should be complemented by more meaningful patient‐day and patient‐stay measures.

Measures of Control

In addition to deciding the unit(s) of analysis, another issue concerns which measures of control to use. These could include rates of hypoglycemia and hyperglycemia, percentage of glucose readings within various ranges (eg, <70, 70180, >180 mg/dL), mean glucose value, percentage of patient‐days during which the mean glucose is within various ranges, or the in control rate (ie, when all glucose values are within a certain range).

As with the various units of analysis, each of these measures of control has various advantages and disadvantages. For example, mean glucose is easy to report and understand, but masks extreme values. Percentage of glucose values within a certain range (eg, per patient, averaged across patients) presents a more complete picture but is a little harder to understand and will vary depending on the frequency of glucose monitoring. As mentioned above, this latter problem can be corrected in part by including only certain glucose values. Percent of glucose values within range may also be less sensitive to change than mean glucose (eg, a glucose that is lowered from 300 mg/dL to 200 mg/dL is still out of range). We recommend choosing a few, but not all, measures of control in order to get a complete picture of glycemic control. Over time one can then refine the measures being used to meet the needs of the glycemic control team and provide data that will drive the performance improvement process.

In critical care and perioperative settings, interest in glycemic control is often more intense around the time of a particular event such as major surgery or after admission to the ICU. Some measures commonly used in performing such analyses are:

• 1

  All values outside a target range within a designated crucial period. For example, the University Healthcare Consortium and other organizations use a simple metric to gauge perioperative glycemic control. They collect the fasting glucose on postoperative days 1 and 2 and then calculate the percentage of postoperative days with any fasting glucose >200 mg/dL. Of course, this is a very liberal target, but it can always be lowered in a stepwise fashion once it is regularly being reached.

• 2

  Three‐day blood glucose average. The Portland group uses the mean glucose of each patient for the period that includes the day of coronary artery bypass graft (CABG) surgery and the following 2 days. The 3‐day BG average (3‐BG) correlates very well with patient outcomes and can serve as a well‐defined target.2 It is likely that use of the 3‐BG would work well in other perioperative/trauma settings and could work in the medical ICU as well, with admission to the ICU as the starting point for calculation of the 3‐BG.

Definitions of Hyperglycemia and Hypoglycemia

Glucometrics outcomes will obviously depend on the thresholds established for hyperglycemia and hypoglycemia. Many centers define hypoglycemia as 60 mg/dL, whereas the ADA definition, based on physiologic changes that may take place, defines hypoglycemia (at least in the outpatient setting) as 70 mg/dL. Hypoglycemia may be further stratified by severity, with any glucose 40 mg/dL, for instance, defined as severe hypoglycemia.

Similarly, the definition of hyperglycemia (and therefore good control) must also be defined. Based on definitions developed by the ADA and AACE, the state of the medical literature, and current understanding of the pathophysiology of hyperglycemia, thresholds for critical care units include 110 mg/dL, 130 mg/dL, and 140 mg/dL, and options in noncritical care units include 130 mg/dL, 140 mg/dL, and 180 mg/dL. Because these thresholds implicitly assume adverse effects when glucose levels are above them, these levels are subject to revision as data become available confirming the benefits and safety of targeted glycemic control in various settings and patient populations.

Introducing optimal BG targets in a stepped fashion over time should also be considered. Furnary et al.2 have done this in the Portland Project, which tracks glycemic control in cardiac surgery patients receiving intravenous insulin therapy. The initial BG target for this project was <200 mg/dL; it was subsequently lowered stepwise over several years to 150 mg/dL, then to 120 mg/dL, and most recently to 110 mg/dL. This approach allows the safe introduction of targeted glycemic control and promotes acceptance of the concept by physicians and the allied nursing and medical staff.

Recommendations:

• In noncritical care units, it is reasonable to use 40 mg/dL for severe hypoglycemia, 70 mg/dL for hypoglycemia, 130 mg/dL for fasting hyperglycemia, 180 mg/dL for random or postprandial hyperglycemia, and 300 mg/dL for severe hyperglycemia, keeping in mind that these thresholds are arbitrary. In critical care units, values from 110 mg/dL to 140 mg/dL might be better thresholds for hyperglycemia, but it may take time to safely and effectively move an organization toward these lower targets.

Other Considerations Relative to Glucometrics

ICU and Perioperative Settings

For ICU and perioperative settings, the major process measure will likely be use of the insulin infusion order set. Designation of BG levels that trigger insulin infusion in these settings should be agreed upon in advance. The number of patients who meet the predefined glycemic criteria would make up the denominator, and the number of patients on the insulin infusion order set would make up the numerator.

NonCritical Care Units

On noncritical care units, measuring the percentage of subcutaneous insulin regimens that contain a basal insulin is a useful way to monitor the impact of an intervention. A more detailed analysis could examine the percentage of patients on simultaneous basal and nutritional insulin (if applicable). An important measure of clinical inertia is to track the percentage of patients who had changes in their insulin regimens on days after hypoglycemic or hyperglycemic excursions. Another important measure is the frequency with which the standardized order set is being used, analogous to the measure of insulin infusion use in the ICU. A final process measure, indirectly related to insulin use, is the frequency of use of oral diabetes agents, especially by patients for whom their use is contraindicated (eg, patients with congestive heart failure who are on thiazolidinediones and patients with renal insufficiency or receiving intravenous contrast continued on metformin).

Defining the Target Patient Population

The first decision to be made is which patients to include in your analysis. Choices include the following:

• 1

  Patients with a discharge diagnosis of diabetes: this group has face validity and intuitive appeal, is easy to identify retrospectively, and may capture some untested/untreated diabetics, but will miss patients with otherwise undiagnosed diabetes and stress hyperglycemia. It is also subject to the variable accuracy of billing codes.

• 2

  Patients with a certain number of point‐of‐care (POC) glucose measurements: this group is also easy to identify, easy to measure, and will include patients with hyperglycemia without a previous diagnosis of diabetes, but will miss patients with untested/untreated hyperglycemia. Also, if glucose levels are checked on normoglycemic, nondiabetic patients, these values may dilute the overall assessment of glycemic control.

• 3

  Patients treated with insulin in the hospital: this is a good choice if the purpose is mainly drug safety and avoidance of hypoglycemia, but by definition excludes most untreated patients.

• 4

  Patients with 2 or more BG values (laboratory and/or POC) over a certain threshold (eg, >180 mg/dL). This will likely capture more patients with inpatient hyperglycemia, whether or not detected by the medical team, but is subject to wide variations in the frequency and timing of laboratory glucose testing, including whether or not the values are pre‐prandial (note that even preprandial POC glucose measurements are not always in fact fasting values).

Other considerations include the following:

• 1

  Are there natural patient subgroups that should be measured and analyzed separately because of different guidelines? For example, there probably should be separate/emndependent inclusion criteria and analyses for critical care and noncritical care units because their glycemic targets and management considerations differ.

• 2

  Which patients should be excluded? For example, if targeting subcutaneous insulin use in general hospitalized patients, one might eliminate those patients who are admitted specifically as the result of a diabetes emergency (eg, diabetic ketoacidosis [DKA] and hyperglycemic hyperosmolar state [HHS]), as their marked and prolonged hyperglycemia will skew BG data. Pregnant women should generally be excluded from broad‐based analyses or considered as a discrete category because they have very different targets for BG therapy. Patients with short lengths of stay may be less likely to benefit from tight glucose control and may also be considered for post hoc exclusion. One might also exclude patients with very few evaluable glucose readings (eg, fewer than 5) to ensure that measurement is meaningful for a given patient, keeping in mind that this may also exclude patients with undetected hyperglycemia, as mentioned above. Finally, patients receiving palliative care should also be considered for exclusion if feasible.

Recommendation: Do not limit analyses to only those patients with a diagnosis of diabetes or only those on insulin, which will lead to biased results.

• For noncritical care patients, we recommend a combined approach: adult patients with a diagnosis of diabetes (e. g. using diagnosis‐related group [DRG] codes 294 or 295 or International Classification of Diseases 9th edition [ICD9] codes 250.xx) or with hyperglycemia (eg, 2 or more random laboratory and/or point of care (POC) BG values >180 mg/dL or 2 or more fasting BG values >130 mg/dL), excluding patients with DKA or HHS or who are pregnant.

• For critical care units, we recommend either all patients, or patients with at least mild hyperglycemia (eg, 2 random glucose levels >140 mg/dL). Critical care patients with DKA, HHS, and pregnancy should be evaluated separately if possible.

Which Glucose Values to Include and Exclude

To answer this question, we first need to decide which method to use for BG measurement. There are several ways to measure BG, including the type of sample collected (capillary [fingerstick], arterial, and venous) and the technique used (central laboratory analyzing plasma, central laboratory analyzing whole blood [eg, from an arterial blood gas sample], glucose meter [usually calibrated to plasma], etc.). The POC (eg, capillary, glucose meter) glucose measurements alone are often preferred in the non‐ICU setting because laboratory plasma values generally provide little additional information and typically lower the mean glucose by including redundant fasting values.1 In critical care units, several different methods are often used together, and each merits inclusion. The inherent differences in calibration between the methods do not generally require separate analyses, especially given the frequency of testing in the ICU setting.

The next question is which values to include in analyses. In some situations, it may be most useful to focus on a certain period of hospitalization, such as the day of a procedure and the next 2 days in assessing the impact of the quality of perioperative care, or the first 14 days of a noncritical care stay to keep outliers for length of stay (LOS) from skewing the data. In the non‐ICU setting, it may be reasonable to exclude the first day of hospitalization, as early BG control is impacted by multiple variables beyond direct control of the clinician (eg, glucose control prior to admission, severity of presenting illness) and may not realistically reflect your interventions. (Keep in mind, however, that it may be useful to adjust for the admission glucose value in multivariable models given its importance to clinical outcomes and its strong relationship to subsequent inpatient glucose control.) However, in critical care units, it is reasonable to include the first day's readings in analyses given the high frequency of glucose measurements in this setting and the expectation that glucose control should be achieved within a few hours of starting an intravenous insulin infusion.

If feasible to do so with your institution's data capture methods, you may wish to select only the regularly scheduled (before each meal [qAC] and at bedtime [qHS], or every 6 hours [q6h]) glucose readings for inclusion in the summary data of glycemic control in the non‐ICU setting, thereby reducing bias caused by repeated measurements around extremes of glycemic excursions. An alternative in the non‐ICU setting is to censor glucose readings within 60 minutes of a previous reading.

Recommendation:

• In the non‐ICU setting, we recommend first looking at all POC glucose values and if possible repeating the analyses excluding hospital day 1 and hospital day 15 and beyond, and also excluding glucose values measured within 60 minutes of a previous value.

• In critical care units, we recommend evaluating all glucose readings used to guide care.

Units of Analysis

There are several different units of analysis, each with its own advantages and disadvantages:

• 1

  Glucose value: this is the simplest measure and the one with the most statistical power. All glucose values for all patients of interest comprise the denominator. A report might say, for example, that 1% of the 1000 glucose values were <70 mg/dL during a certain period or that the mean of all glucose values collected for the month from patients in noncritical care areas was 160 mg/dL. The potential disadvantages of this approach are that these analyses are less clinically relevant than patient‐level analyses and that patients with many glucose readings and long hospitalizations may skew the data.

• 2

  Patient (or the Patient Stay, [ie, the entire hospitalization]): all patients who are monitored make up the denominator. The numerator may be the percentage of patients with any hypoglycemia during their hospital stay or the percentage of patients achieving a certain mean glucose during their hospitalization, for example. This is inherently more clinically meaningful than using glucose value as a unit of analysis. A major disadvantage is not controlling for LOS effects. For example, a hospitalized patient with a long LOS is much more likely to be characterized as having at least 1 hypoglycemic value than is a patient with a shorter LOS. Another shortcoming is that this approach does not correct for uneven distribution of testing. A patient's mean glucose might be calculated on the basis of 8 glucose values on the first day of hospitalization, 4 on the second day, and 1 on the third day. Despite all these shortcomings, reporting by patient remains a popular and valid method of presenting glycemic control results, particularly when complemented by other views and refined to control for the number of readings per day.

• 3

  Monitored Patient‐Day: The denominator in this setting is the total number of days a patient glucose level is monitored. The benefits of this method have been described and advocated in the literature.1 As with patient‐level analyses, this measure will be more rigorous and meaningful if the BG measures to be evaluated have been standardized. Typical reports might include percentage of monitored days with any hypoglycemia, or percentage of monitored days with all glucose values in the desired range. This unit of analysis may be considered more difficult to generate and to interpret. On the other hand, it is clinically relevant, less biased by LOS effects, and may be considered the most actionable metric by clinicians. This method provides a good balance when presented with data organized by patient.

The following example uses all 3 units of measurement, in this case to determine the rate of hypoglycemia, demonstrating the different but complementary information that each method provides:

• In 1 month, 3900 POC glucose measurements were obtained from 286 patients representing 986 monitored patient‐days. With hypoglycemia defined as POC BG 60 mg/dL, the results showed the following:

• 50 of 3900 measurements (1.4%) were hypoglycemic 22 of 286 patients (7.7%) had 1 hypoglycemic episodes

• 40 of 986 monitored days (4.4%) had 1 hypoglycemic episodes.

The metric based on the number of glucose readings could be considered the least clinically relevant because it is unclear how many patients were affected; moreover, it may be based on variable testing patterns among patients, and could be influenced disproportionately by 1 patient with frequent hypoglycemia, many glucose readings, and/or a long LOS. One could argue that the patient‐stay metric is artificially elevated because a single hypoglycemic episode characterizes the entire stay as hypoglycemic. On the other hand, at least it acknowledges the number of patients affected by hypoglycemia. The patient‐day unit of analysis likely provides the most balanced view, one that is clinically relevant and measured over a standard period of time, and less biased by LOS and frequency of testing.

One way to express patient‐day glycemic control that deserves special mention is the patient‐day weighted mean. A mean glucose is calculated for each patient‐day, and then the mean is calculated across all patient‐days. The advantage of this approach is that it corrects for variation in the number of glucose readings each day; all hospital days are weighted equally.

Recommendation:

• In noncritical care units, we recommend a combination of patient‐day and patient‐stay measures.

• In critical care units, it is acceptable to also use glucose reading as the unit of measurement given more frequent and uniform data collection, but it should be complemented by more meaningful patient‐day and patient‐stay measures.

Measures of Control

In addition to deciding the unit(s) of analysis, another issue concerns which measures of control to use. These could include rates of hypoglycemia and hyperglycemia, percentage of glucose readings within various ranges (eg, <70, 70180, >180 mg/dL), mean glucose value, percentage of patient‐days during which the mean glucose is within various ranges, or the in control rate (ie, when all glucose values are within a certain range).

As with the various units of analysis, each of these measures of control has various advantages and disadvantages. For example, mean glucose is easy to report and understand, but masks extreme values. Percentage of glucose values within a certain range (eg, per patient, averaged across patients) presents a more complete picture but is a little harder to understand and will vary depending on the frequency of glucose monitoring. As mentioned above, this latter problem can be corrected in part by including only certain glucose values. Percent of glucose values within range may also be less sensitive to change than mean glucose (eg, a glucose that is lowered from 300 mg/dL to 200 mg/dL is still out of range). We recommend choosing a few, but not all, measures of control in order to get a complete picture of glycemic control. Over time one can then refine the measures being used to meet the needs of the glycemic control team and provide data that will drive the performance improvement process.

In critical care and perioperative settings, interest in glycemic control is often more intense around the time of a particular event such as major surgery or after admission to the ICU. Some measures commonly used in performing such analyses are:

• 1

  All values outside a target range within a designated crucial period. For example, the University Healthcare Consortium and other organizations use a simple metric to gauge perioperative glycemic control. They collect the fasting glucose on postoperative days 1 and 2 and then calculate the percentage of postoperative days with any fasting glucose >200 mg/dL. Of course, this is a very liberal target, but it can always be lowered in a stepwise fashion once it is regularly being reached.

• 2

  Three‐day blood glucose average. The Portland group uses the mean glucose of each patient for the period that includes the day of coronary artery bypass graft (CABG) surgery and the following 2 days. The 3‐day BG average (3‐BG) correlates very well with patient outcomes and can serve as a well‐defined target.2 It is likely that use of the 3‐BG would work well in other perioperative/trauma settings and could work in the medical ICU as well, with admission to the ICU as the starting point for calculation of the 3‐BG.

Definitions of Hyperglycemia and Hypoglycemia

Glucometrics outcomes will obviously depend on the thresholds established for hyperglycemia and hypoglycemia. Many centers define hypoglycemia as 60 mg/dL, whereas the ADA definition, based on physiologic changes that may take place, defines hypoglycemia (at least in the outpatient setting) as 70 mg/dL. Hypoglycemia may be further stratified by severity, with any glucose 40 mg/dL, for instance, defined as severe hypoglycemia.

Similarly, the definition of hyperglycemia (and therefore good control) must also be defined. Based on definitions developed by the ADA and AACE, the state of the medical literature, and current understanding of the pathophysiology of hyperglycemia, thresholds for critical care units include 110 mg/dL, 130 mg/dL, and 140 mg/dL, and options in noncritical care units include 130 mg/dL, 140 mg/dL, and 180 mg/dL. Because these thresholds implicitly assume adverse effects when glucose levels are above them, these levels are subject to revision as data become available confirming the benefits and safety of targeted glycemic control in various settings and patient populations.

Introducing optimal BG targets in a stepped fashion over time should also be considered. Furnary et al.2 have done this in the Portland Project, which tracks glycemic control in cardiac surgery patients receiving intravenous insulin therapy. The initial BG target for this project was <200 mg/dL; it was subsequently lowered stepwise over several years to 150 mg/dL, then to 120 mg/dL, and most recently to 110 mg/dL. This approach allows the safe introduction of targeted glycemic control and promotes acceptance of the concept by physicians and the allied nursing and medical staff.

Recommendations:

• In noncritical care units, it is reasonable to use 40 mg/dL for severe hypoglycemia, 70 mg/dL for hypoglycemia, 130 mg/dL for fasting hyperglycemia, 180 mg/dL for random or postprandial hyperglycemia, and 300 mg/dL for severe hyperglycemia, keeping in mind that these thresholds are arbitrary. In critical care units, values from 110 mg/dL to 140 mg/dL might be better thresholds for hyperglycemia, but it may take time to safely and effectively move an organization toward these lower targets.

Other Considerations Relative to Glucometrics

ICU and Perioperative Settings

For ICU and perioperative settings, the major process measure will likely be use of the insulin infusion order set. Designation of BG levels that trigger insulin infusion in these settings should be agreed upon in advance. The number of patients who meet the predefined glycemic criteria would make up the denominator, and the number of patients on the insulin infusion order set would make up the numerator.

NonCritical Care Units

On noncritical care units, measuring the percentage of subcutaneous insulin regimens that contain a basal insulin is a useful way to monitor the impact of an intervention. A more detailed analysis could examine the percentage of patients on simultaneous basal and nutritional insulin (if applicable). An important measure of clinical inertia is to track the percentage of patients who had changes in their insulin regimens on days after hypoglycemic or hyperglycemic excursions. Another important measure is the frequency with which the standardized order set is being used, analogous to the measure of insulin infusion use in the ICU. A final process measure, indirectly related to insulin use, is the frequency of use of oral diabetes agents, especially by patients for whom their use is contraindicated (eg, patients with congestive heart failure who are on thiazolidinediones and patients with renal insufficiency or receiving intravenous contrast continued on metformin).

Defining the Target Patient Population

The first decision to be made is which patients to include in your analysis. Choices include the following:

• 1

  Patients with a discharge diagnosis of diabetes: this group has face validity and intuitive appeal, is easy to identify retrospectively, and may capture some untested/untreated diabetics, but will miss patients with otherwise undiagnosed diabetes and stress hyperglycemia. It is also subject to the variable accuracy of billing codes.

• 2

  Patients with a certain number of point‐of‐care (POC) glucose measurements: this group is also easy to identify, easy to measure, and will include patients with hyperglycemia without a previous diagnosis of diabetes, but will miss patients with untested/untreated hyperglycemia. Also, if glucose levels are checked on normoglycemic, nondiabetic patients, these values may dilute the overall assessment of glycemic control.

• 3

  Patients treated with insulin in the hospital: this is a good choice if the purpose is mainly drug safety and avoidance of hypoglycemia, but by definition excludes most untreated patients.

• 4

  Patients with 2 or more BG values (laboratory and/or POC) over a certain threshold (eg, >180 mg/dL). This will likely capture more patients with inpatient hyperglycemia, whether or not detected by the medical team, but is subject to wide variations in the frequency and timing of laboratory glucose testing, including whether or not the values are pre‐prandial (note that even preprandial POC glucose measurements are not always in fact fasting values).

Other considerations include the following:

• 1

  Are there natural patient subgroups that should be measured and analyzed separately because of different guidelines? For example, there probably should be separate/emndependent inclusion criteria and analyses for critical care and noncritical care units because their glycemic targets and management considerations differ.

• 2

  Which patients should be excluded? For example, if targeting subcutaneous insulin use in general hospitalized patients, one might eliminate those patients who are admitted specifically as the result of a diabetes emergency (eg, diabetic ketoacidosis [DKA] and hyperglycemic hyperosmolar state [HHS]), as their marked and prolonged hyperglycemia will skew BG data. Pregnant women should generally be excluded from broad‐based analyses or considered as a discrete category because they have very different targets for BG therapy. Patients with short lengths of stay may be less likely to benefit from tight glucose control and may also be considered for post hoc exclusion. One might also exclude patients with very few evaluable glucose readings (eg, fewer than 5) to ensure that measurement is meaningful for a given patient, keeping in mind that this may also exclude patients with undetected hyperglycemia, as mentioned above. Finally, patients receiving palliative care should also be considered for exclusion if feasible.

Recommendation: Do not limit analyses to only those patients with a diagnosis of diabetes or only those on insulin, which will lead to biased results.

• For noncritical care patients, we recommend a combined approach: adult patients with a diagnosis of diabetes (e. g. using diagnosis‐related group [DRG] codes 294 or 295 or International Classification of Diseases 9th edition [ICD9] codes 250.xx) or with hyperglycemia (eg, 2 or more random laboratory and/or point of care (POC) BG values >180 mg/dL or 2 or more fasting BG values >130 mg/dL), excluding patients with DKA or HHS or who are pregnant.

• For critical care units, we recommend either all patients, or patients with at least mild hyperglycemia (eg, 2 random glucose levels >140 mg/dL). Critical care patients with DKA, HHS, and pregnancy should be evaluated separately if possible.

Which Glucose Values to Include and Exclude

To answer this question, we first need to decide which method to use for BG measurement. There are several ways to measure BG, including the type of sample collected (capillary [fingerstick], arterial, and venous) and the technique used (central laboratory analyzing plasma, central laboratory analyzing whole blood [eg, from an arterial blood gas sample], glucose meter [usually calibrated to plasma], etc.). The POC (eg, capillary, glucose meter) glucose measurements alone are often preferred in the non‐ICU setting because laboratory plasma values generally provide little additional information and typically lower the mean glucose by including redundant fasting values.1 In critical care units, several different methods are often used together, and each merits inclusion. The inherent differences in calibration between the methods do not generally require separate analyses, especially given the frequency of testing in the ICU setting.

The next question is which values to include in analyses. In some situations, it may be most useful to focus on a certain period of hospitalization, such as the day of a procedure and the next 2 days in assessing the impact of the quality of perioperative care, or the first 14 days of a noncritical care stay to keep outliers for length of stay (LOS) from skewing the data. In the non‐ICU setting, it may be reasonable to exclude the first day of hospitalization, as early BG control is impacted by multiple variables beyond direct control of the clinician (eg, glucose control prior to admission, severity of presenting illness) and may not realistically reflect your interventions. (Keep in mind, however, that it may be useful to adjust for the admission glucose value in multivariable models given its importance to clinical outcomes and its strong relationship to subsequent inpatient glucose control.) However, in critical care units, it is reasonable to include the first day's readings in analyses given the high frequency of glucose measurements in this setting and the expectation that glucose control should be achieved within a few hours of starting an intravenous insulin infusion.

If feasible to do so with your institution's data capture methods, you may wish to select only the regularly scheduled (before each meal [qAC] and at bedtime [qHS], or every 6 hours [q6h]) glucose readings for inclusion in the summary data of glycemic control in the non‐ICU setting, thereby reducing bias caused by repeated measurements around extremes of glycemic excursions. An alternative in the non‐ICU setting is to censor glucose readings within 60 minutes of a previous reading.

Recommendation:

• In the non‐ICU setting, we recommend first looking at all POC glucose values and if possible repeating the analyses excluding hospital day 1 and hospital day 15 and beyond, and also excluding glucose values measured within 60 minutes of a previous value.

• In critical care units, we recommend evaluating all glucose readings used to guide care.

Units of Analysis

There are several different units of analysis, each with its own advantages and disadvantages:

• 1

  Glucose value: this is the simplest measure and the one with the most statistical power. All glucose values for all patients of interest comprise the denominator. A report might say, for example, that 1% of the 1000 glucose values were <70 mg/dL during a certain period or that the mean of all glucose values collected for the month from patients in noncritical care areas was 160 mg/dL. The potential disadvantages of this approach are that these analyses are less clinically relevant than patient‐level analyses and that patients with many glucose readings and long hospitalizations may skew the data.

• 2

  Patient (or the Patient Stay, [ie, the entire hospitalization]): all patients who are monitored make up the denominator. The numerator may be the percentage of patients with any hypoglycemia during their hospital stay or the percentage of patients achieving a certain mean glucose during their hospitalization, for example. This is inherently more clinically meaningful than using glucose value as a unit of analysis. A major disadvantage is not controlling for LOS effects. For example, a hospitalized patient with a long LOS is much more likely to be characterized as having at least 1 hypoglycemic value than is a patient with a shorter LOS. Another shortcoming is that this approach does not correct for uneven distribution of testing. A patient's mean glucose might be calculated on the basis of 8 glucose values on the first day of hospitalization, 4 on the second day, and 1 on the third day. Despite all these shortcomings, reporting by patient remains a popular and valid method of presenting glycemic control results, particularly when complemented by other views and refined to control for the number of readings per day.

• 3

  Monitored Patient‐Day: The denominator in this setting is the total number of days a patient glucose level is monitored. The benefits of this method have been described and advocated in the literature.1 As with patient‐level analyses, this measure will be more rigorous and meaningful if the BG measures to be evaluated have been standardized. Typical reports might include percentage of monitored days with any hypoglycemia, or percentage of monitored days with all glucose values in the desired range. This unit of analysis may be considered more difficult to generate and to interpret. On the other hand, it is clinically relevant, less biased by LOS effects, and may be considered the most actionable metric by clinicians. This method provides a good balance when presented with data organized by patient.

The following example uses all 3 units of measurement, in this case to determine the rate of hypoglycemia, demonstrating the different but complementary information that each method provides:

• In 1 month, 3900 POC glucose measurements were obtained from 286 patients representing 986 monitored patient‐days. With hypoglycemia defined as POC BG 60 mg/dL, the results showed the following:

• 50 of 3900 measurements (1.4%) were hypoglycemic 22 of 286 patients (7.7%) had 1 hypoglycemic episodes

• 40 of 986 monitored days (4.4%) had 1 hypoglycemic episodes.

The metric based on the number of glucose readings could be considered the least clinically relevant because it is unclear how many patients were affected; moreover, it may be based on variable testing patterns among patients, and could be influenced disproportionately by 1 patient with frequent hypoglycemia, many glucose readings, and/or a long LOS. One could argue that the patient‐stay metric is artificially elevated because a single hypoglycemic episode characterizes the entire stay as hypoglycemic. On the other hand, at least it acknowledges the number of patients affected by hypoglycemia. The patient‐day unit of analysis likely provides the most balanced view, one that is clinically relevant and measured over a standard period of time, and less biased by LOS and frequency of testing.

One way to express patient‐day glycemic control that deserves special mention is the patient‐day weighted mean. A mean glucose is calculated for each patient‐day, and then the mean is calculated across all patient‐days. The advantage of this approach is that it corrects for variation in the number of glucose readings each day; all hospital days are weighted equally.

Recommendation:

• In noncritical care units, we recommend a combination of patient‐day and patient‐stay measures.

• In critical care units, it is acceptable to also use glucose reading as the unit of measurement given more frequent and uniform data collection, but it should be complemented by more meaningful patient‐day and patient‐stay measures.

Measures of Control

In addition to deciding the unit(s) of analysis, another issue concerns which measures of control to use. These could include rates of hypoglycemia and hyperglycemia, percentage of glucose readings within various ranges (eg, <70, 70180, >180 mg/dL), mean glucose value, percentage of patient‐days during which the mean glucose is within various ranges, or the in control rate (ie, when all glucose values are within a certain range).

As with the various units of analysis, each of these measures of control has various advantages and disadvantages. For example, mean glucose is easy to report and understand, but masks extreme values. Percentage of glucose values within a certain range (eg, per patient, averaged across patients) presents a more complete picture but is a little harder to understand and will vary depending on the frequency of glucose monitoring. As mentioned above, this latter problem can be corrected in part by including only certain glucose values. Percent of glucose values within range may also be less sensitive to change than mean glucose (eg, a glucose that is lowered from 300 mg/dL to 200 mg/dL is still out of range). We recommend choosing a few, but not all, measures of control in order to get a complete picture of glycemic control. Over time one can then refine the measures being used to meet the needs of the glycemic control team and provide data that will drive the performance improvement process.

In critical care and perioperative settings, interest in glycemic control is often more intense around the time of a particular event such as major surgery or after admission to the ICU. Some measures commonly used in performing such analyses are:

• 1

  All values outside a target range within a designated crucial period. For example, the University Healthcare Consortium and other organizations use a simple metric to gauge perioperative glycemic control. They collect the fasting glucose on postoperative days 1 and 2 and then calculate the percentage of postoperative days with any fasting glucose >200 mg/dL. Of course, this is a very liberal target, but it can always be lowered in a stepwise fashion once it is regularly being reached.

• 2

  Three‐day blood glucose average. The Portland group uses the mean glucose of each patient for the period that includes the day of coronary artery bypass graft (CABG) surgery and the following 2 days. The 3‐day BG average (3‐BG) correlates very well with patient outcomes and can serve as a well‐defined target.2 It is likely that use of the 3‐BG would work well in other perioperative/trauma settings and could work in the medical ICU as well, with admission to the ICU as the starting point for calculation of the 3‐BG.

Definitions of Hyperglycemia and Hypoglycemia

Glucometrics outcomes will obviously depend on the thresholds established for hyperglycemia and hypoglycemia. Many centers define hypoglycemia as 60 mg/dL, whereas the ADA definition, based on physiologic changes that may take place, defines hypoglycemia (at least in the outpatient setting) as 70 mg/dL. Hypoglycemia may be further stratified by severity, with any glucose 40 mg/dL, for instance, defined as severe hypoglycemia.

Similarly, the definition of hyperglycemia (and therefore good control) must also be defined. Based on definitions developed by the ADA and AACE, the state of the medical literature, and current understanding of the pathophysiology of hyperglycemia, thresholds for critical care units include 110 mg/dL, 130 mg/dL, and 140 mg/dL, and options in noncritical care units include 130 mg/dL, 140 mg/dL, and 180 mg/dL. Because these thresholds implicitly assume adverse effects when glucose levels are above them, these levels are subject to revision as data become available confirming the benefits and safety of targeted glycemic control in various settings and patient populations.

Introducing optimal BG targets in a stepped fashion over time should also be considered. Furnary et al.2 have done this in the Portland Project, which tracks glycemic control in cardiac surgery patients receiving intravenous insulin therapy. The initial BG target for this project was <200 mg/dL; it was subsequently lowered stepwise over several years to 150 mg/dL, then to 120 mg/dL, and most recently to 110 mg/dL. This approach allows the safe introduction of targeted glycemic control and promotes acceptance of the concept by physicians and the allied nursing and medical staff.

Recommendations:

• In noncritical care units, it is reasonable to use 40 mg/dL for severe hypoglycemia, 70 mg/dL for hypoglycemia, 130 mg/dL for fasting hyperglycemia, 180 mg/dL for random or postprandial hyperglycemia, and 300 mg/dL for severe hyperglycemia, keeping in mind that these thresholds are arbitrary. In critical care units, values from 110 mg/dL to 140 mg/dL might be better thresholds for hyperglycemia, but it may take time to safely and effectively move an organization toward these lower targets.

Other Considerations Relative to Glucometrics

ICU and Perioperative Settings

For ICU and perioperative settings, the major process measure will likely be use of the insulin infusion order set. Designation of BG levels that trigger insulin infusion in these settings should be agreed upon in advance. The number of patients who meet the predefined glycemic criteria would make up the denominator, and the number of patients on the insulin infusion order set would make up the numerator.

NonCritical Care Units

On noncritical care units, measuring the percentage of subcutaneous insulin regimens that contain a basal insulin is a useful way to monitor the impact of an intervention. A more detailed analysis could examine the percentage of patients on simultaneous basal and nutritional insulin (if applicable). An important measure of clinical inertia is to track the percentage of patients who had changes in their insulin regimens on days after hypoglycemic or hyperglycemic excursions. Another important measure is the frequency with which the standardized order set is being used, analogous to the measure of insulin infusion use in the ICU. A final process measure, indirectly related to insulin use, is the frequency of use of oral diabetes agents, especially by patients for whom their use is contraindicated (eg, patients with congestive heart failure who are on thiazolidinediones and patients with renal insufficiency or receiving intravenous contrast continued on metformin).

Standardize the Monitoring of Blood Glucose Values and Glucosylated Hemoglobin

Guidance for the coordination of glucose testing, nutrition delivery, and insulin administration, should be integrated into your protocols, and order sets. For noncritical care areas, the minimal frequency for blood glucose monitoring for patients who are eating is before meals and at bedtime. For the patient designated nothing by mouth (NPO) or the patient on continuous tube feeding, the type of nutritional/correction insulin used should drive the minimum frequency (every 4‐6 hours if rapid acting analog insulins [RAA‐I] are used, and every 6 hours if regular insulin is used). Directions for administering scheduled RAA‐I immediately before or immediately after nutrition delivery should be incorporated into protocols, order sets, and medication administration records. Unfortunately, having this guidance in the order sets and protocols does not automatically translate into its being carried out in the real world. Wide variability in the coordination of glucose monitoring, nutritional delivery, and insulin administration is common, so monitoring the process to make sure the protocol is followed is important.

Obtaining a glucosylated hemoglobin (HbA1c) level is important in gauging how well the patient's outpatient regimen is maintaining glycemic control, distinguishing stress hyperglycemia from established diabetes, and guiding the inpatient approach to glycemic control. ADA guidelines2, 3 endorse obtaining HbA1c levels of inpatients if these levels are not already available from the month prior to admission.

Establish a Target Range for Blood Glucose in NonCritical Care Areas

It is important to adopt a glycemic target that is institution‐wide, for critical care areas and noncritical care areas alike. Your glycemic target need not be identical to the ADA/AACE glycemic targets, but should be similar to them.

Examples of institutional glycemic targets for noncritical care areas:

• Preprandial target 90‐130 mg/dL, maximum random glucose <180 mg/dL (ADA/AACE consensus target)

• 90‐150 mg/dL (a target used in some hospitals)

• Preprandial target 90‐130 mg/dL for most patients, 100‐150 mg/dL if there are hypoglycemia risk factors, and <180 mg/dL if comfort‐care or end‐of‐life care (a more refined target, allowing for customization based on patient characteristics).

Your multidisciplinary glycemic control steering committee should pick the glycemic target it can most successfully implement and disseminate. It is fine to start with a conservative target and then ratchet down the goals as the environment becomes more accepting of the concept of tighter control of blood glucose in the hospital.

Although the choice of glycemic target is somewhat arbitrary, establishing an institutional glycemic target is critical to motivate clinical action. Your committee should design interventions, for instances when a patient's glycemic target is consistently not being met, including an assignment of responsibility.

Prompt Clinicians to Consider Discontinuing Oral Agents

Oral antihyperglycemic agents, in general, are difficult to quickly titrate to effect, and have side effects that limit their use in the hospital. In contrast, insulin acts rapidly and can be used in virtually all patients and clinical situations, making it the treatment of choice for treatment of hyperglycemia in the hospital.3, 11, 12 In certain circumstances, it may be entirely appropriate to continue a well‐controlled patient on his or her prior outpatient oral regimen. It is often also reasonable to resume oral agents in some patients when preparing for hospital discharge.

Incorporate Nutritional Management

Because diet is so integral to the management of diabetes and hyperglycemia, diet orders should be embedded in all diabetes or insulin‐related order sets. Diets with the same amount of carbohydrate with each meal should be the default rule for patients with diabetes. Nutritionist consultation should be considered and easy to access for patients with malnutrition, obesity, and other common conditions of the inpatient with diabetes.

Access Diabetes Education and Appropriate Consultation

Diabetes education should be offered to all hyperglycemic patients with normal mental status, complete with written materials, a listing of community resources, and survival skills. Consultation with physicians in internal medicine or endocrinology for difficult‐to‐control cases, or for cases in which the primary physician of record is not familiar with (or not adherent to) principles of inpatient glycemic management, should be very easy to obtain, or perhaps mandated, depending on your institution‐specific environment.

Prescribe Physiologic (Basal‐Nutritional‐Correction Dose) Insulin Regimens

Physiologic insulin use is the backbone of the recommended best practice for diabetes and hyperglycemia management in the hospital. The principles of such regimens are summarized elsewhere in this supplement.12 These principles will not be reiterated in detail here, but the major concepts that should be integrated into the protocols and order sets will be highlighted.

Special Situations

Most of the preferred regimens for different situations are outlined in Table 2 in a straightforward manner, and can be depicted in your protocols and order sets in the same way. Some conditions have enough complexity, however, that you will have difficulty placing all of the details into your one‐page protocol and order set. Details should be placed on your more detailed protocol, and educational programs should include the topics outlined below. Although insulin infusion is often the option that would provide the most reliable and expedient control of hyperglycemia in these special situations, it is an option not available in many noncritical care settings. Therefore, the discussion is limited to subcutaneous insulin control regimens.

Regiment the Management of Hypoglycemia

Hypoglycemia is defined by the ADA as a blood glucose of 70 mg/dL or less, based on the physiologic changes that can occur at this glucose level, even in subjectively asymptomatic patients.3 Protocols for management of hypoglycemia should be linked to your diabetes/hyperglycemia protocols. There are many hypoglycemia protocols available for review in the SHM Glycemic Control Resource Room and Glycemic Control Implementation Guide.10 Some common themes for effective implementation stand out. First, the protocols need to walk the balance between simplicity of use, and the need to provide instructions that will provide guidance in a variety of patient situations. Second, the protocols need to be nurse driven, so that nurses can initiate treatment without waiting for a physician order. Third, education and instruction regarding recognition of risk factors, and avoidance of hypoglycemia are needed to support a successful protocol. Importantly, any hypoglycemic event should lead to a reconsideration of the current anti‐hyperglycemic regimen so that future events can be prevented.

Plan for Discharge and Provide Guidance for the Transition

Your institution should have policies and procedures outlining all the steps needed to complete the important transition out of the hospital. At a minimum, this planning should include adequate education (including a learner assessment), appropriate follow‐up, referral to community resources, and a discharge glycemic control regimen that is tailored to the educational, financial, and motivational profile of a patient. The more your inpatient insulin management is driven by protocol, the more likely it is the patient will be on an inpatient treatment plan that differs from their outpatient regimen; therefore, it is even more important to plan this transition carefully and reliably.

Communicating the accurate hyperglycemia related diagnosis and related problems to the primary care provider is important for good care, perhaps even more so for patients who had hyperglycemia while hospitalized without a prior diagnosis of diabetes. Some centers place a prompt for hyperglycemia related diagnosis in the order set and/or discharge paperwork, to remind the clinician to convey the diagnosis to the primary provider, and to encourage more complete documentation. Improved documentation can also improve the business case for glycemic control, along with other strategies outlined elsewhere in this supplement.13

Transitions in care (including transitions out of the hospital and off of infusion insulin) are discussed in more detail14, 15 elsewhere in this supplement. The principles outlined in these references should be incorporated into your institutional protocol. Briefly, not all patients require or are capable of intensive basal‐bolus regimens upon discharge. The HbA1c can be very valuable in arriving at the optimal outpatient regimen.14 The capacities and preferences of the patient and the context of his or her outpatient care environment (including the preferences of the primary care provider) must be taken into consideration as an outpatient management program is planned.

Standardize the Monitoring of Blood Glucose Values and Glucosylated Hemoglobin

Guidance for the coordination of glucose testing, nutrition delivery, and insulin administration, should be integrated into your protocols, and order sets. For noncritical care areas, the minimal frequency for blood glucose monitoring for patients who are eating is before meals and at bedtime. For the patient designated nothing by mouth (NPO) or the patient on continuous tube feeding, the type of nutritional/correction insulin used should drive the minimum frequency (every 4‐6 hours if rapid acting analog insulins [RAA‐I] are used, and every 6 hours if regular insulin is used). Directions for administering scheduled RAA‐I immediately before or immediately after nutrition delivery should be incorporated into protocols, order sets, and medication administration records. Unfortunately, having this guidance in the order sets and protocols does not automatically translate into its being carried out in the real world. Wide variability in the coordination of glucose monitoring, nutritional delivery, and insulin administration is common, so monitoring the process to make sure the protocol is followed is important.

Obtaining a glucosylated hemoglobin (HbA1c) level is important in gauging how well the patient's outpatient regimen is maintaining glycemic control, distinguishing stress hyperglycemia from established diabetes, and guiding the inpatient approach to glycemic control. ADA guidelines2, 3 endorse obtaining HbA1c levels of inpatients if these levels are not already available from the month prior to admission.

Establish a Target Range for Blood Glucose in NonCritical Care Areas

It is important to adopt a glycemic target that is institution‐wide, for critical care areas and noncritical care areas alike. Your glycemic target need not be identical to the ADA/AACE glycemic targets, but should be similar to them.

Examples of institutional glycemic targets for noncritical care areas:

• Preprandial target 90‐130 mg/dL, maximum random glucose <180 mg/dL (ADA/AACE consensus target)

• 90‐150 mg/dL (a target used in some hospitals)

• Preprandial target 90‐130 mg/dL for most patients, 100‐150 mg/dL if there are hypoglycemia risk factors, and <180 mg/dL if comfort‐care or end‐of‐life care (a more refined target, allowing for customization based on patient characteristics).

Your multidisciplinary glycemic control steering committee should pick the glycemic target it can most successfully implement and disseminate. It is fine to start with a conservative target and then ratchet down the goals as the environment becomes more accepting of the concept of tighter control of blood glucose in the hospital.

Although the choice of glycemic target is somewhat arbitrary, establishing an institutional glycemic target is critical to motivate clinical action. Your committee should design interventions, for instances when a patient's glycemic target is consistently not being met, including an assignment of responsibility.

Prompt Clinicians to Consider Discontinuing Oral Agents

Oral antihyperglycemic agents, in general, are difficult to quickly titrate to effect, and have side effects that limit their use in the hospital. In contrast, insulin acts rapidly and can be used in virtually all patients and clinical situations, making it the treatment of choice for treatment of hyperglycemia in the hospital.3, 11, 12 In certain circumstances, it may be entirely appropriate to continue a well‐controlled patient on his or her prior outpatient oral regimen. It is often also reasonable to resume oral agents in some patients when preparing for hospital discharge.

Incorporate Nutritional Management

Because diet is so integral to the management of diabetes and hyperglycemia, diet orders should be embedded in all diabetes or insulin‐related order sets. Diets with the same amount of carbohydrate with each meal should be the default rule for patients with diabetes. Nutritionist consultation should be considered and easy to access for patients with malnutrition, obesity, and other common conditions of the inpatient with diabetes.

Access Diabetes Education and Appropriate Consultation

Diabetes education should be offered to all hyperglycemic patients with normal mental status, complete with written materials, a listing of community resources, and survival skills. Consultation with physicians in internal medicine or endocrinology for difficult‐to‐control cases, or for cases in which the primary physician of record is not familiar with (or not adherent to) principles of inpatient glycemic management, should be very easy to obtain, or perhaps mandated, depending on your institution‐specific environment.

Prescribe Physiologic (Basal‐Nutritional‐Correction Dose) Insulin Regimens

Physiologic insulin use is the backbone of the recommended best practice for diabetes and hyperglycemia management in the hospital. The principles of such regimens are summarized elsewhere in this supplement.12 These principles will not be reiterated in detail here, but the major concepts that should be integrated into the protocols and order sets will be highlighted.

Special Situations

Most of the preferred regimens for different situations are outlined in Table 2 in a straightforward manner, and can be depicted in your protocols and order sets in the same way. Some conditions have enough complexity, however, that you will have difficulty placing all of the details into your one‐page protocol and order set. Details should be placed on your more detailed protocol, and educational programs should include the topics outlined below. Although insulin infusion is often the option that would provide the most reliable and expedient control of hyperglycemia in these special situations, it is an option not available in many noncritical care settings. Therefore, the discussion is limited to subcutaneous insulin control regimens.

Regiment the Management of Hypoglycemia

Hypoglycemia is defined by the ADA as a blood glucose of 70 mg/dL or less, based on the physiologic changes that can occur at this glucose level, even in subjectively asymptomatic patients.3 Protocols for management of hypoglycemia should be linked to your diabetes/hyperglycemia protocols. There are many hypoglycemia protocols available for review in the SHM Glycemic Control Resource Room and Glycemic Control Implementation Guide.10 Some common themes for effective implementation stand out. First, the protocols need to walk the balance between simplicity of use, and the need to provide instructions that will provide guidance in a variety of patient situations. Second, the protocols need to be nurse driven, so that nurses can initiate treatment without waiting for a physician order. Third, education and instruction regarding recognition of risk factors, and avoidance of hypoglycemia are needed to support a successful protocol. Importantly, any hypoglycemic event should lead to a reconsideration of the current anti‐hyperglycemic regimen so that future events can be prevented.

Plan for Discharge and Provide Guidance for the Transition

Your institution should have policies and procedures outlining all the steps needed to complete the important transition out of the hospital. At a minimum, this planning should include adequate education (including a learner assessment), appropriate follow‐up, referral to community resources, and a discharge glycemic control regimen that is tailored to the educational, financial, and motivational profile of a patient. The more your inpatient insulin management is driven by protocol, the more likely it is the patient will be on an inpatient treatment plan that differs from their outpatient regimen; therefore, it is even more important to plan this transition carefully and reliably.

Communicating the accurate hyperglycemia related diagnosis and related problems to the primary care provider is important for good care, perhaps even more so for patients who had hyperglycemia while hospitalized without a prior diagnosis of diabetes. Some centers place a prompt for hyperglycemia related diagnosis in the order set and/or discharge paperwork, to remind the clinician to convey the diagnosis to the primary provider, and to encourage more complete documentation. Improved documentation can also improve the business case for glycemic control, along with other strategies outlined elsewhere in this supplement.13

Transitions in care (including transitions out of the hospital and off of infusion insulin) are discussed in more detail14, 15 elsewhere in this supplement. The principles outlined in these references should be incorporated into your institutional protocol. Briefly, not all patients require or are capable of intensive basal‐bolus regimens upon discharge. The HbA1c can be very valuable in arriving at the optimal outpatient regimen.14 The capacities and preferences of the patient and the context of his or her outpatient care environment (including the preferences of the primary care provider) must be taken into consideration as an outpatient management program is planned.

Standardize the Monitoring of Blood Glucose Values and Glucosylated Hemoglobin

Guidance for the coordination of glucose testing, nutrition delivery, and insulin administration, should be integrated into your protocols, and order sets. For noncritical care areas, the minimal frequency for blood glucose monitoring for patients who are eating is before meals and at bedtime. For the patient designated nothing by mouth (NPO) or the patient on continuous tube feeding, the type of nutritional/correction insulin used should drive the minimum frequency (every 4‐6 hours if rapid acting analog insulins [RAA‐I] are used, and every 6 hours if regular insulin is used). Directions for administering scheduled RAA‐I immediately before or immediately after nutrition delivery should be incorporated into protocols, order sets, and medication administration records. Unfortunately, having this guidance in the order sets and protocols does not automatically translate into its being carried out in the real world. Wide variability in the coordination of glucose monitoring, nutritional delivery, and insulin administration is common, so monitoring the process to make sure the protocol is followed is important.

Obtaining a glucosylated hemoglobin (HbA1c) level is important in gauging how well the patient's outpatient regimen is maintaining glycemic control, distinguishing stress hyperglycemia from established diabetes, and guiding the inpatient approach to glycemic control. ADA guidelines2, 3 endorse obtaining HbA1c levels of inpatients if these levels are not already available from the month prior to admission.

Establish a Target Range for Blood Glucose in NonCritical Care Areas

It is important to adopt a glycemic target that is institution‐wide, for critical care areas and noncritical care areas alike. Your glycemic target need not be identical to the ADA/AACE glycemic targets, but should be similar to them.

Examples of institutional glycemic targets for noncritical care areas:

• Preprandial target 90‐130 mg/dL, maximum random glucose <180 mg/dL (ADA/AACE consensus target)

• 90‐150 mg/dL (a target used in some hospitals)

• Preprandial target 90‐130 mg/dL for most patients, 100‐150 mg/dL if there are hypoglycemia risk factors, and <180 mg/dL if comfort‐care or end‐of‐life care (a more refined target, allowing for customization based on patient characteristics).

Your multidisciplinary glycemic control steering committee should pick the glycemic target it can most successfully implement and disseminate. It is fine to start with a conservative target and then ratchet down the goals as the environment becomes more accepting of the concept of tighter control of blood glucose in the hospital.

Although the choice of glycemic target is somewhat arbitrary, establishing an institutional glycemic target is critical to motivate clinical action. Your committee should design interventions, for instances when a patient's glycemic target is consistently not being met, including an assignment of responsibility.

Prompt Clinicians to Consider Discontinuing Oral Agents

Oral antihyperglycemic agents, in general, are difficult to quickly titrate to effect, and have side effects that limit their use in the hospital. In contrast, insulin acts rapidly and can be used in virtually all patients and clinical situations, making it the treatment of choice for treatment of hyperglycemia in the hospital.3, 11, 12 In certain circumstances, it may be entirely appropriate to continue a well‐controlled patient on his or her prior outpatient oral regimen. It is often also reasonable to resume oral agents in some patients when preparing for hospital discharge.

Incorporate Nutritional Management

Because diet is so integral to the management of diabetes and hyperglycemia, diet orders should be embedded in all diabetes or insulin‐related order sets. Diets with the same amount of carbohydrate with each meal should be the default rule for patients with diabetes. Nutritionist consultation should be considered and easy to access for patients with malnutrition, obesity, and other common conditions of the inpatient with diabetes.

Access Diabetes Education and Appropriate Consultation

Diabetes education should be offered to all hyperglycemic patients with normal mental status, complete with written materials, a listing of community resources, and survival skills. Consultation with physicians in internal medicine or endocrinology for difficult‐to‐control cases, or for cases in which the primary physician of record is not familiar with (or not adherent to) principles of inpatient glycemic management, should be very easy to obtain, or perhaps mandated, depending on your institution‐specific environment.

Prescribe Physiologic (Basal‐Nutritional‐Correction Dose) Insulin Regimens

Physiologic insulin use is the backbone of the recommended best practice for diabetes and hyperglycemia management in the hospital. The principles of such regimens are summarized elsewhere in this supplement.12 These principles will not be reiterated in detail here, but the major concepts that should be integrated into the protocols and order sets will be highlighted.

Special Situations

Most of the preferred regimens for different situations are outlined in Table 2 in a straightforward manner, and can be depicted in your protocols and order sets in the same way. Some conditions have enough complexity, however, that you will have difficulty placing all of the details into your one‐page protocol and order set. Details should be placed on your more detailed protocol, and educational programs should include the topics outlined below. Although insulin infusion is often the option that would provide the most reliable and expedient control of hyperglycemia in these special situations, it is an option not available in many noncritical care settings. Therefore, the discussion is limited to subcutaneous insulin control regimens.

Regiment the Management of Hypoglycemia

Hypoglycemia is defined by the ADA as a blood glucose of 70 mg/dL or less, based on the physiologic changes that can occur at this glucose level, even in subjectively asymptomatic patients.3 Protocols for management of hypoglycemia should be linked to your diabetes/hyperglycemia protocols. There are many hypoglycemia protocols available for review in the SHM Glycemic Control Resource Room and Glycemic Control Implementation Guide.10 Some common themes for effective implementation stand out. First, the protocols need to walk the balance between simplicity of use, and the need to provide instructions that will provide guidance in a variety of patient situations. Second, the protocols need to be nurse driven, so that nurses can initiate treatment without waiting for a physician order. Third, education and instruction regarding recognition of risk factors, and avoidance of hypoglycemia are needed to support a successful protocol. Importantly, any hypoglycemic event should lead to a reconsideration of the current anti‐hyperglycemic regimen so that future events can be prevented.

Plan for Discharge and Provide Guidance for the Transition

Your institution should have policies and procedures outlining all the steps needed to complete the important transition out of the hospital. At a minimum, this planning should include adequate education (including a learner assessment), appropriate follow‐up, referral to community resources, and a discharge glycemic control regimen that is tailored to the educational, financial, and motivational profile of a patient. The more your inpatient insulin management is driven by protocol, the more likely it is the patient will be on an inpatient treatment plan that differs from their outpatient regimen; therefore, it is even more important to plan this transition carefully and reliably.

Communicating the accurate hyperglycemia related diagnosis and related problems to the primary care provider is important for good care, perhaps even more so for patients who had hyperglycemia while hospitalized without a prior diagnosis of diabetes. Some centers place a prompt for hyperglycemia related diagnosis in the order set and/or discharge paperwork, to remind the clinician to convey the diagnosis to the primary provider, and to encourage more complete documentation. Improved documentation can also improve the business case for glycemic control, along with other strategies outlined elsewhere in this supplement.13

Transitions in care (including transitions out of the hospital and off of infusion insulin) are discussed in more detail14, 15 elsewhere in this supplement. The principles outlined in these references should be incorporated into your institutional protocol. Briefly, not all patients require or are capable of intensive basal‐bolus regimens upon discharge. The HbA1c can be very valuable in arriving at the optimal outpatient regimen.14 The capacities and preferences of the patient and the context of his or her outpatient care environment (including the preferences of the primary care provider) must be taken into consideration as an outpatient management program is planned.