Setting

The University of Pennsylvania Health System (UPHS) includes 3 hospitals, with a combined capacity of over 1500 beds and 70,000 annual admissions. All hospitals currently utilize Sunrise Clinical Manager version 5.5 (Allscripts, Chicago, IL) as their EHR. The study sample included all adult admissions to any of the 3 UPHS hospitals during the study period. Admissions to short procedure, rehabilitation, and hospice units were excluded. The study received expedited approval and a HIPAA waiver from the University of Pennsylvania institutional review board.

Development of Predictive Model

The UPHS Center for Evidence‐based Practice[15, 16] performed a systematic review to identify factors associated with hospital readmission within 30 days of discharge. We then examined the data available from our hospital EHR at the time of admission for those factors identified in the review. Using different threshold values and look‐back periods, we developed and tested 30 candidate prediction models using these variables alone and in combination (Table 1). Prediction models were evaluated using 24 months of historical data between August 1, 2009 and August 1, 2011.

Implementation

An automated readmission risk flag was then integrated into the EHR. Patients classified as being at high risk for readmission with the automated prediction model were flagged in the EHR on admission (Figure 1A). The flag can be double‐clicked to display a separate screen with information relevant to discharge planning including inpatient and emergency department (ED) visits in the prior 12 months, as well as information about the primary team, length of stay, and admitting problem associated with those admissions (Figure 1B). The prediction model was integrated into our EHR using Arden Syntax for Medical Logic Modules.[17] The readmission risk screen involved presenting the provider with a new screen and was thus developed in Microsoft .NET using C# and Windows Forms (Microsoft Corp., Redmond, WA).

Figure 1

The flag was visible on the patient lists of all providers who utilized the EHR. This included but was not limited to nurses, social workers, unit pharmacists, and physicians. At the time of implementation, educational events regarding the readmission risk flag were provided in forums targeting administrators, pharmacists, social workers, and housestaff. Information about the flag and recommendations for use were distributed through emails and broadcast screensaver messages disseminated throughout the inpatient units of the health system. Providers were asked to pay special attention to discharge planning for patients triggering the readmission risk flag, including medication reconciliation by pharmacists for these patients prior to discharge, and arrangement of available home services by social workers.

The risk flag was 1 of 4 classes of interventions developed and endorsed by the health system in its efforts to reduce readmissions. Besides risk stratification, the other classes were: interdisciplinary rounding, patient education, and discharge communication. None of the interventions alone were expected to decrease readmissions, but as all 4 classes of interventions were implemented and performed routinely, the expectation was that they would work in concert to reduce readmissions.

Analysis

The primary outcome was all‐cause hospital readmissions in the healthcare system within 30 days of discharge. Although this outcome is commonly used both in the literature and as a quality metric, significant debate persists as to the appropriateness of this metric.[18] Many of the factors driving 30‐day readmissions may be dependent on factors outside of the discharging hospital's control and it has been argued that nearer‐term, nonelective readmission rates may provide a more meaningful quality metric.[18] Seven‐day unplanned readmissions were thus used as a secondary outcome measure for this study.

Sensitivity, specificity, predictive value, C statistic, F score (the harmonic mean of positive predictive value and sensitivity),[19] and screen‐positive rate were calculated for each of the 30 prediction models evaluated using the historical data. The prediction model with the best balance of F score and screen‐positive rate was selected as the prediction model to be integrated into the EHR. Prospective validation of the selected prediction model was performed using the 12‐month period following implementation of the risk flag (September 2011September 2012).

To assess the impact of the automated prediction model on monthly readmission rate, we used the 24‐month period immediately before and the 12‐month period immediately after implementation of the readmission risk flag. Segmented regression analysis was performed testing for changes in level and slope of readmission rates between preimplementation and postimplementation time periods. This quasiexperimental interrupted time series methodology[20] allows us to control for secular trends in readmission rates and to assess the preimplementation trend (secular trend), the difference in rates immediately before and after the implementation (immediate effect), and the postimplementation change over time (sustained effect). We used Cochrane‐Orcutt estimation[21] to correct for serial autocorrelation.

All analyses were performed using Stata 12.1 software (Stata Corp, College Station, TX).

Predictors of Readmission

Our systematic review of the literature identified several patient and healthcare utilization patterns predictive of 30‐day readmission risk. Utilization factors included length of stay, number of prior admissions, previous 30‐day readmissions, and previous ED visits. Patient characteristics included number of comorbidities, living alone, and payor. Evidence was inconsistent regarding threshold values for these variables.

Many variables readily available in our EHR were either found by the systematic review not to be reliably predictive of 30‐day readmission (including age and gender) or were not readily or reliably available on admission (including length of stay and payor). At the time of implementation, our EHR did not include vital sign or nursing assessment variables, so these were not considered for inclusion in our model.

Of the available variables, 3 were consistently accurate and available in the EHR at the time of patient admission: prior hospital admission, emergency department visit, and 30‐day readmission within UPHS. We then developed 30 candidate prediction models using a combination of these variables, including 1 and 2 prior admissions, ED visits, and 30‐day readmissions in the 6 and 12 months preceding the index visit (Table 1).

Development and Validation

We used 24 months of retrospective data, which included 120,396 discharges with 17,337 thirty‐day readmissions (14.4% 30‐day all‐cause readmission rate) to test the candidate prediction models. A single risk factor, 2 inpatient admissions in the past 12 months, was found to have the best balance of sensitivity (40%), positive predictive value (31%), and proportion of patients flagged (18%) (Table 1).

Prospective validation of the prediction model was performed using the 12‐month period directly following readmission risk flag implementation. During this period, the 30‐day all‐cause readmission rate was 15.1%. Sensitivity (39%), positive predictive value (30%), and proportion of patients flagged (18%) were consistent with the values derived from the retrospective data, supporting the reproducibility and predictive stability of the chosen risk prediction model (Table 1). The C statistic of the model was also consistent between the retrospective and prospective datasets (0.62 and 0.61, respectively).

Readmission Rates

The mean 30‐day all‐cause readmission rate for the 24‐month period prior to the intervention was 14.4%, whereas the mean for the 12‐month period after the implementation was 15.1%. Thirty‐day all‐cause and 7‐day unplanned monthly readmission rates do not appear to have been impacted by the intervention (Figure 2). There was no evidence for either an immediate or sustained effect (Table 2).

Figure 2

Limitations

There are several limitations of this intervention. First, the prediction model was developed using 30‐day all‐cause readmissions, rather than attempting to identify potentially preventable readmissions. Thirty‐day readmission rates may not be a good proxy for preventable readmissions,[18] and as a consequence, the ability to predict 30‐day readmissions may not ensure that a prediction model is able to predict preventable readmissions. Nonetheless, 30‐day readmission rates remain the most commonly used quality metric.

Second, the impact of the risk flag on provider behavior is uncertain. We did not formally assess how the readmission risk flag was used by healthcare team members. Informal assessment has, however, revealed that the readmission risk flag is gradually being adopted by different members of the care team including unit‐based pharmacists who are using the flag to prioritize the delivery of medication education, social workers who are using the flag to prompt providers to consider higher level services for patients at high risk of readmission, and patient navigators who are using the flag to prioritize follow‐up phone calls. As a result, we hope that the flag will ultimately improve the processes of care for high‐risk patients.

Third, we did not capture readmissions to hospitals outside of our healthcare system and have therefore underestimated the readmission rate in our population. However, our assessment of the effect of the risk flag on readmissions focused on relative readmission rates over time, and the use of the interrupted time series methodology should protect against secular changes in outside hospital readmission rates that were not associated with the intervention.

Fourth, it is possible that the prediction model implemented could be significantly improved by including additional variables or data available during the hospital stay. However, simple classification models using a single variable have repeatedly been shown to have the ability to compete favorably with state‐of‐the‐art multivariable classification models.[28]

Fifth, our study was limited to a single academic health system, and our experience may not be generalizable to smaller healthcare systems with limited EHR systems. However, the simplicity of our prediction model and the integration into a commercial EHR may improve the generalizability of our experience to other healthcare settings. Additionally, partly due to recent policy initiatives, the adoption of integrated EHR systems by hospitals is expected to continue at a rapid rate and become the standard of care within the near future.[29]

Setting

The University of Pennsylvania Health System (UPHS) includes 3 hospitals, with a combined capacity of over 1500 beds and 70,000 annual admissions. All hospitals currently utilize Sunrise Clinical Manager version 5.5 (Allscripts, Chicago, IL) as their EHR. The study sample included all adult admissions to any of the 3 UPHS hospitals during the study period. Admissions to short procedure, rehabilitation, and hospice units were excluded. The study received expedited approval and a HIPAA waiver from the University of Pennsylvania institutional review board.

Development of Predictive Model

The UPHS Center for Evidence‐based Practice[15, 16] performed a systematic review to identify factors associated with hospital readmission within 30 days of discharge. We then examined the data available from our hospital EHR at the time of admission for those factors identified in the review. Using different threshold values and look‐back periods, we developed and tested 30 candidate prediction models using these variables alone and in combination (Table 1). Prediction models were evaluated using 24 months of historical data between August 1, 2009 and August 1, 2011.

Implementation

An automated readmission risk flag was then integrated into the EHR. Patients classified as being at high risk for readmission with the automated prediction model were flagged in the EHR on admission (Figure 1A). The flag can be double‐clicked to display a separate screen with information relevant to discharge planning including inpatient and emergency department (ED) visits in the prior 12 months, as well as information about the primary team, length of stay, and admitting problem associated with those admissions (Figure 1B). The prediction model was integrated into our EHR using Arden Syntax for Medical Logic Modules.[17] The readmission risk screen involved presenting the provider with a new screen and was thus developed in Microsoft .NET using C# and Windows Forms (Microsoft Corp., Redmond, WA).

Figure 1

The flag was visible on the patient lists of all providers who utilized the EHR. This included but was not limited to nurses, social workers, unit pharmacists, and physicians. At the time of implementation, educational events regarding the readmission risk flag were provided in forums targeting administrators, pharmacists, social workers, and housestaff. Information about the flag and recommendations for use were distributed through emails and broadcast screensaver messages disseminated throughout the inpatient units of the health system. Providers were asked to pay special attention to discharge planning for patients triggering the readmission risk flag, including medication reconciliation by pharmacists for these patients prior to discharge, and arrangement of available home services by social workers.

The risk flag was 1 of 4 classes of interventions developed and endorsed by the health system in its efforts to reduce readmissions. Besides risk stratification, the other classes were: interdisciplinary rounding, patient education, and discharge communication. None of the interventions alone were expected to decrease readmissions, but as all 4 classes of interventions were implemented and performed routinely, the expectation was that they would work in concert to reduce readmissions.

Analysis

The primary outcome was all‐cause hospital readmissions in the healthcare system within 30 days of discharge. Although this outcome is commonly used both in the literature and as a quality metric, significant debate persists as to the appropriateness of this metric.[18] Many of the factors driving 30‐day readmissions may be dependent on factors outside of the discharging hospital's control and it has been argued that nearer‐term, nonelective readmission rates may provide a more meaningful quality metric.[18] Seven‐day unplanned readmissions were thus used as a secondary outcome measure for this study.

Sensitivity, specificity, predictive value, C statistic, F score (the harmonic mean of positive predictive value and sensitivity),[19] and screen‐positive rate were calculated for each of the 30 prediction models evaluated using the historical data. The prediction model with the best balance of F score and screen‐positive rate was selected as the prediction model to be integrated into the EHR. Prospective validation of the selected prediction model was performed using the 12‐month period following implementation of the risk flag (September 2011September 2012).

To assess the impact of the automated prediction model on monthly readmission rate, we used the 24‐month period immediately before and the 12‐month period immediately after implementation of the readmission risk flag. Segmented regression analysis was performed testing for changes in level and slope of readmission rates between preimplementation and postimplementation time periods. This quasiexperimental interrupted time series methodology[20] allows us to control for secular trends in readmission rates and to assess the preimplementation trend (secular trend), the difference in rates immediately before and after the implementation (immediate effect), and the postimplementation change over time (sustained effect). We used Cochrane‐Orcutt estimation[21] to correct for serial autocorrelation.

All analyses were performed using Stata 12.1 software (Stata Corp, College Station, TX).

Predictors of Readmission

Our systematic review of the literature identified several patient and healthcare utilization patterns predictive of 30‐day readmission risk. Utilization factors included length of stay, number of prior admissions, previous 30‐day readmissions, and previous ED visits. Patient characteristics included number of comorbidities, living alone, and payor. Evidence was inconsistent regarding threshold values for these variables.

Many variables readily available in our EHR were either found by the systematic review not to be reliably predictive of 30‐day readmission (including age and gender) or were not readily or reliably available on admission (including length of stay and payor). At the time of implementation, our EHR did not include vital sign or nursing assessment variables, so these were not considered for inclusion in our model.

Of the available variables, 3 were consistently accurate and available in the EHR at the time of patient admission: prior hospital admission, emergency department visit, and 30‐day readmission within UPHS. We then developed 30 candidate prediction models using a combination of these variables, including 1 and 2 prior admissions, ED visits, and 30‐day readmissions in the 6 and 12 months preceding the index visit (Table 1).

Development and Validation

We used 24 months of retrospective data, which included 120,396 discharges with 17,337 thirty‐day readmissions (14.4% 30‐day all‐cause readmission rate) to test the candidate prediction models. A single risk factor, 2 inpatient admissions in the past 12 months, was found to have the best balance of sensitivity (40%), positive predictive value (31%), and proportion of patients flagged (18%) (Table 1).

Prospective validation of the prediction model was performed using the 12‐month period directly following readmission risk flag implementation. During this period, the 30‐day all‐cause readmission rate was 15.1%. Sensitivity (39%), positive predictive value (30%), and proportion of patients flagged (18%) were consistent with the values derived from the retrospective data, supporting the reproducibility and predictive stability of the chosen risk prediction model (Table 1). The C statistic of the model was also consistent between the retrospective and prospective datasets (0.62 and 0.61, respectively).

Readmission Rates

The mean 30‐day all‐cause readmission rate for the 24‐month period prior to the intervention was 14.4%, whereas the mean for the 12‐month period after the implementation was 15.1%. Thirty‐day all‐cause and 7‐day unplanned monthly readmission rates do not appear to have been impacted by the intervention (Figure 2). There was no evidence for either an immediate or sustained effect (Table 2).

Figure 2

Limitations

There are several limitations of this intervention. First, the prediction model was developed using 30‐day all‐cause readmissions, rather than attempting to identify potentially preventable readmissions. Thirty‐day readmission rates may not be a good proxy for preventable readmissions,[18] and as a consequence, the ability to predict 30‐day readmissions may not ensure that a prediction model is able to predict preventable readmissions. Nonetheless, 30‐day readmission rates remain the most commonly used quality metric.

Second, the impact of the risk flag on provider behavior is uncertain. We did not formally assess how the readmission risk flag was used by healthcare team members. Informal assessment has, however, revealed that the readmission risk flag is gradually being adopted by different members of the care team including unit‐based pharmacists who are using the flag to prioritize the delivery of medication education, social workers who are using the flag to prompt providers to consider higher level services for patients at high risk of readmission, and patient navigators who are using the flag to prioritize follow‐up phone calls. As a result, we hope that the flag will ultimately improve the processes of care for high‐risk patients.

Third, we did not capture readmissions to hospitals outside of our healthcare system and have therefore underestimated the readmission rate in our population. However, our assessment of the effect of the risk flag on readmissions focused on relative readmission rates over time, and the use of the interrupted time series methodology should protect against secular changes in outside hospital readmission rates that were not associated with the intervention.

Fourth, it is possible that the prediction model implemented could be significantly improved by including additional variables or data available during the hospital stay. However, simple classification models using a single variable have repeatedly been shown to have the ability to compete favorably with state‐of‐the‐art multivariable classification models.[28]

Fifth, our study was limited to a single academic health system, and our experience may not be generalizable to smaller healthcare systems with limited EHR systems. However, the simplicity of our prediction model and the integration into a commercial EHR may improve the generalizability of our experience to other healthcare settings. Additionally, partly due to recent policy initiatives, the adoption of integrated EHR systems by hospitals is expected to continue at a rapid rate and become the standard of care within the near future.[29]

Setting

The University of Pennsylvania Health System (UPHS) includes 3 hospitals, with a combined capacity of over 1500 beds and 70,000 annual admissions. All hospitals currently utilize Sunrise Clinical Manager version 5.5 (Allscripts, Chicago, IL) as their EHR. The study sample included all adult admissions to any of the 3 UPHS hospitals during the study period. Admissions to short procedure, rehabilitation, and hospice units were excluded. The study received expedited approval and a HIPAA waiver from the University of Pennsylvania institutional review board.

Development of Predictive Model

The UPHS Center for Evidence‐based Practice[15, 16] performed a systematic review to identify factors associated with hospital readmission within 30 days of discharge. We then examined the data available from our hospital EHR at the time of admission for those factors identified in the review. Using different threshold values and look‐back periods, we developed and tested 30 candidate prediction models using these variables alone and in combination (Table 1). Prediction models were evaluated using 24 months of historical data between August 1, 2009 and August 1, 2011.

Implementation

An automated readmission risk flag was then integrated into the EHR. Patients classified as being at high risk for readmission with the automated prediction model were flagged in the EHR on admission (Figure 1A). The flag can be double‐clicked to display a separate screen with information relevant to discharge planning including inpatient and emergency department (ED) visits in the prior 12 months, as well as information about the primary team, length of stay, and admitting problem associated with those admissions (Figure 1B). The prediction model was integrated into our EHR using Arden Syntax for Medical Logic Modules.[17] The readmission risk screen involved presenting the provider with a new screen and was thus developed in Microsoft .NET using C# and Windows Forms (Microsoft Corp., Redmond, WA).

Figure 1

The flag was visible on the patient lists of all providers who utilized the EHR. This included but was not limited to nurses, social workers, unit pharmacists, and physicians. At the time of implementation, educational events regarding the readmission risk flag were provided in forums targeting administrators, pharmacists, social workers, and housestaff. Information about the flag and recommendations for use were distributed through emails and broadcast screensaver messages disseminated throughout the inpatient units of the health system. Providers were asked to pay special attention to discharge planning for patients triggering the readmission risk flag, including medication reconciliation by pharmacists for these patients prior to discharge, and arrangement of available home services by social workers.

The risk flag was 1 of 4 classes of interventions developed and endorsed by the health system in its efforts to reduce readmissions. Besides risk stratification, the other classes were: interdisciplinary rounding, patient education, and discharge communication. None of the interventions alone were expected to decrease readmissions, but as all 4 classes of interventions were implemented and performed routinely, the expectation was that they would work in concert to reduce readmissions.

Analysis

The primary outcome was all‐cause hospital readmissions in the healthcare system within 30 days of discharge. Although this outcome is commonly used both in the literature and as a quality metric, significant debate persists as to the appropriateness of this metric.[18] Many of the factors driving 30‐day readmissions may be dependent on factors outside of the discharging hospital's control and it has been argued that nearer‐term, nonelective readmission rates may provide a more meaningful quality metric.[18] Seven‐day unplanned readmissions were thus used as a secondary outcome measure for this study.

Sensitivity, specificity, predictive value, C statistic, F score (the harmonic mean of positive predictive value and sensitivity),[19] and screen‐positive rate were calculated for each of the 30 prediction models evaluated using the historical data. The prediction model with the best balance of F score and screen‐positive rate was selected as the prediction model to be integrated into the EHR. Prospective validation of the selected prediction model was performed using the 12‐month period following implementation of the risk flag (September 2011September 2012).

To assess the impact of the automated prediction model on monthly readmission rate, we used the 24‐month period immediately before and the 12‐month period immediately after implementation of the readmission risk flag. Segmented regression analysis was performed testing for changes in level and slope of readmission rates between preimplementation and postimplementation time periods. This quasiexperimental interrupted time series methodology[20] allows us to control for secular trends in readmission rates and to assess the preimplementation trend (secular trend), the difference in rates immediately before and after the implementation (immediate effect), and the postimplementation change over time (sustained effect). We used Cochrane‐Orcutt estimation[21] to correct for serial autocorrelation.

All analyses were performed using Stata 12.1 software (Stata Corp, College Station, TX).

Predictors of Readmission

Our systematic review of the literature identified several patient and healthcare utilization patterns predictive of 30‐day readmission risk. Utilization factors included length of stay, number of prior admissions, previous 30‐day readmissions, and previous ED visits. Patient characteristics included number of comorbidities, living alone, and payor. Evidence was inconsistent regarding threshold values for these variables.

Many variables readily available in our EHR were either found by the systematic review not to be reliably predictive of 30‐day readmission (including age and gender) or were not readily or reliably available on admission (including length of stay and payor). At the time of implementation, our EHR did not include vital sign or nursing assessment variables, so these were not considered for inclusion in our model.

Of the available variables, 3 were consistently accurate and available in the EHR at the time of patient admission: prior hospital admission, emergency department visit, and 30‐day readmission within UPHS. We then developed 30 candidate prediction models using a combination of these variables, including 1 and 2 prior admissions, ED visits, and 30‐day readmissions in the 6 and 12 months preceding the index visit (Table 1).

Development and Validation

We used 24 months of retrospective data, which included 120,396 discharges with 17,337 thirty‐day readmissions (14.4% 30‐day all‐cause readmission rate) to test the candidate prediction models. A single risk factor, 2 inpatient admissions in the past 12 months, was found to have the best balance of sensitivity (40%), positive predictive value (31%), and proportion of patients flagged (18%) (Table 1).

Prospective validation of the prediction model was performed using the 12‐month period directly following readmission risk flag implementation. During this period, the 30‐day all‐cause readmission rate was 15.1%. Sensitivity (39%), positive predictive value (30%), and proportion of patients flagged (18%) were consistent with the values derived from the retrospective data, supporting the reproducibility and predictive stability of the chosen risk prediction model (Table 1). The C statistic of the model was also consistent between the retrospective and prospective datasets (0.62 and 0.61, respectively).

Readmission Rates

The mean 30‐day all‐cause readmission rate for the 24‐month period prior to the intervention was 14.4%, whereas the mean for the 12‐month period after the implementation was 15.1%. Thirty‐day all‐cause and 7‐day unplanned monthly readmission rates do not appear to have been impacted by the intervention (Figure 2). There was no evidence for either an immediate or sustained effect (Table 2).

Figure 2

Limitations

There are several limitations of this intervention. First, the prediction model was developed using 30‐day all‐cause readmissions, rather than attempting to identify potentially preventable readmissions. Thirty‐day readmission rates may not be a good proxy for preventable readmissions,[18] and as a consequence, the ability to predict 30‐day readmissions may not ensure that a prediction model is able to predict preventable readmissions. Nonetheless, 30‐day readmission rates remain the most commonly used quality metric.

Second, the impact of the risk flag on provider behavior is uncertain. We did not formally assess how the readmission risk flag was used by healthcare team members. Informal assessment has, however, revealed that the readmission risk flag is gradually being adopted by different members of the care team including unit‐based pharmacists who are using the flag to prioritize the delivery of medication education, social workers who are using the flag to prompt providers to consider higher level services for patients at high risk of readmission, and patient navigators who are using the flag to prioritize follow‐up phone calls. As a result, we hope that the flag will ultimately improve the processes of care for high‐risk patients.

Third, we did not capture readmissions to hospitals outside of our healthcare system and have therefore underestimated the readmission rate in our population. However, our assessment of the effect of the risk flag on readmissions focused on relative readmission rates over time, and the use of the interrupted time series methodology should protect against secular changes in outside hospital readmission rates that were not associated with the intervention.

Fourth, it is possible that the prediction model implemented could be significantly improved by including additional variables or data available during the hospital stay. However, simple classification models using a single variable have repeatedly been shown to have the ability to compete favorably with state‐of‐the‐art multivariable classification models.[28]

Fifth, our study was limited to a single academic health system, and our experience may not be generalizable to smaller healthcare systems with limited EHR systems. However, the simplicity of our prediction model and the integration into a commercial EHR may improve the generalizability of our experience to other healthcare settings. Additionally, partly due to recent policy initiatives, the adoption of integrated EHR systems by hospitals is expected to continue at a rapid rate and become the standard of care within the near future.[29]

The Patient Needing Transition to Multiple Daily Insulin Injections

A 29‐year‐old male with T1DM, on CSII, was admitted for fever and chills. His latest glycated hemoglobin (HbA_1c) level was 6.8%. His glucose levels started rising, and he wished to be taken off the insulin pump. He was started by the primary team on multiple daily insulin injections (MDII) with insulin glargine and insulin lispro. His glucose levels continued to rise, so an intravenous (IV) insulin infusion was started. Endocrinology was then consulted. The patient's condition was concerning for the potential development of DKA, so he was kept on IV insulin. When he was ready for transitioning to subcutaneous insulin, the pump had been taken home by a family member, and the patient could not recall his CSII basal rates but knew his total basal insulin dose, carbohydrate ratio, and sensitivity factor. Endocrinology assisted in transitioning him from the insulin infusion to MDII based on these recalled doses. When the insulin pump was available, the pump settings were interrogated, and he was transitioned back to it.

Key points:

• Having key hospital personnel trained on CSII, including interrogating the pump's settings, facilitates the transitioning of these patients from one hospital unit, or level of care, to another.
• Accessing the pump's settings involves pushing several buttons on the insulin pump. Because key hospital personnel will encounter patients on different insulin pumps, it may be helpful to keep menu maps handy as a quick reference. A menu map will show at a glance where certain information can be found, such as the basal insulin rate or the sensitivity factor (see examples in Figures 1 and 2).[12, 13]
• Knowing the HbA_1c will help determine if pump use has been effective.

Figure 1

Figure 2

The Patient With Technical Problems

An 84‐year‐old gentleman with T2DM was admitted for heart failure and aortic valve replacement. His HbA_1c was 6.2%, and he had had several outpatient hypoglycemic events. While on CSII in the hospital, his point of care testing (POCT) glucose readings ranged from 105 to 260 mg/dL. On the afternoon of the third hospital day, POCT readings stayed above 220 mg/dL and rose to 348 mg/dL on the fourth hospital morning, despite multiple blousing and changing the insulin, insertion site, reservoir cartridge, and pump tubing. There was no evidence of infection and no medication change that would have impacted glucose levels. Review of his procedures revealed that he had undergone computed tomography (CT) on the morning of hospital day 3 and wore his pump while being scanned. The pump company was notified.

Key points:

• Patients, and medical and nursing staff, should be reminded to remove insulin pumps for CT scans, magnetic resonance imaging, x‐rays, or other tests with high electromagnetic fields.
• If there is a suspicion of pump malfunction from such a procedure, notify the pump company.

The Patient Who Can Benefit From Inpatient Education

A 70‐year‐old female with T1DM was admitted for heart failure. The patient had been using CSII for 20 years. Her latest HbA_1c was 6.9%. She had 1 hypoglycemic event every 1 to 2 weeks. In the hospital, she experienced 2 hypoglycemic events within 3 days, both around bedtime. It was discovered that the patient was giving a bolus of insulin for elevated glucose levels based on the hospital POCT, and when the meal arrived (3060 minutes later), she again delivered a bolus based on her own glucometer reading plus insulin based on the carbohydrates in her meal. The patient was then instructed to request the POCT when her meal tray arrived, and she was taught how to use the pump's built‐in calculator. Glucose excursions improved.

Key points:

• Patients on CSII, though able to exercise autonomy in managing their insulin doses, may also need assistance in dosing insulin properly.
• Although pump education is ideally done on an outpatient basis, hospital‐based providers may encounter patients who need reinforcement of their training while hospitalized. Hospital personnel trained on insulin pumps (such as physicians, nurse practitioners, physician assistants, and certified diabetes educators) can help augment the patient's knowledge while in the hospital. In the absence of such key personnel, patient safety has to be addressed with re‐evaluation of the need to discontinue the pump and switch to multiple doses of subcutaneous insulin.

Initial Patient Assessment

On admission, patients are asked whether they use an insulin pump. This is included in the nursing assessment form. If they do, the physician is notified.

The insulin pump might be missed unless specifically asked for because (1) the insulin pump may be thought of more as a device rather than a medication, and (2) the insulin pump may be worn in less obvious areas, not only on the abdomen where providers are more apt to detect it.

Hospital Policy and Insulin Orders

Written hospital policies on how to safely manage patients presenting with an insulin pump will delineate patients who can safely be allowed to continue on the pump, and the responsibilities that come along with this. Our institution has such a policy. Experts from both the legal and biomedical engineering departments were consulted when the policy was crafted. Patients must be fully alert, able, and willing to self‐manage the pump. General contraindications to pump use in the hospital, such as altered mental status or DKA, are listed in Table 1. In addition, patients in the intensive care units are best managed on an IV insulin infusion during their critical illness, in keeping with several society guidelines.[14] Controlling severe hyperglycemia and DKA with multiple boluses through the insulin pump can potentially lead to stacking of insulin with subsequent hypoglycemia.

In our institution, a computerized insulin pump order set has to be activated. Apart from insulin, POCT, and hypoglycemia‐management orders, this order set contains documents aimed at balancing patient autonomy with delivery of appropriate and safe medical care that the bedside nurse goes over with the patient (Table 2). By policy at our institution, insulin should be dispensed from the hospital's pharmacy (except for that already in the pump), so the order set is linked to the pharmacy and a 3‐mL insulin vial is delivered to the hospital floor and stored in the patient‐specific medication bin. The order set triggers an Endocrinology consult so that the patient can be assessed by key trained personnel.

Patient Diagnosis

It is important to try to distinguish T1DM vs T2DM, as patients with T1DM are prone to ketoacidosis when the pump is disconnected.

Patient Assessment by the Endocrinology Consult Service

Hospitalized patients on the pump have varying degrees of pump knowledge and skill sets. We have encountered highly trained patients who meticulously count their carbohydrates and double‐check the insulin doses calculated by the built‐in pump calculator, and those who have knowledge gaps because their physicians, and not they themselves, change pump settings at the clinic visits.

Therefore, the Endocrinology consult team members (comprising physicians, nurse practitioners, and certified diabetes educators) go through the following items to be able to order the insulin correctly, assess whether patients are still able and willing to continue on their pump despite their illness, formulate alternative insulin regimens as needed, or help empower patients who may have forgotten some aspects of pump management:

• Insulin pump manufacturer/model.
• Insulin used in the pump.
• Often fast‐acting insulin (lispro, aspart, or glulisine).
• Some patients use regular insulin.
• A few patients use U500 insulin (5 more potent than other insulins).
• Insulin doses/pump settings.

Patients are assessed for:

• Hypoglycemia awareness.
• Previous glucose control.
• Bolus calculation (either using the built‐in calculator, computing this mentally, or using a different calculator).
• Ability to deliver a bolus (including vision and dexterity challenges).
• Ability to change the basal rate, or set a temporary rate, and suspend insulin delivery.

Discussion on Options for Inpatient Management

After assessment, education is provided as needed. If there are concerns on the part of the patient, the primary team, or the Endocrinology team about safe continuation of CSII during the hospitalization, then alternative insulin regimens are discussed. Patients who cannot access their basal rates and cannot adjust the doses are not able to self‐manage; they should be taken off the pump and treated with multiple subcutaneous insulin doses. Conversion to MDII is discussed under Interruption of Continuous Subcutaneous Insulin Infusion for Short and Prolonged Periods.

Provision of Pump Information for Hospital Healthcare Providers

Users of CSII, even if perfectly competent using their pumps in the ambulatory setting, may need assistance in the hospital for various reasons. They may not know what to do for surgical or radiologic procedures (discussed below) and may not be familiar with hospital policies involving CSII. Hospital providers trained on insulin pumps may need a refresher on locating a particular pump setting.

The provider can call the toll‐free number on the back of the pump for assistance (Table 3). Insulin‐pump companies also have menu maps to aid in finding information on pump settings (samples shown in Figures 1 and 2).[12, 13] Documentation of the patient's pump settings will assist in CSII dose changes during the acute illness or assist in switching to MDII if needed. The following information need to be collected:

Point‐of‐Care Testing for Glucose Monitoring

Our policy specifies that the hospital glucose meter is the meter of record upon which dose adjustments are based. Point‐of‐care testing is performed by our patient care nursing assistants or bedside nurses. The timing is typically before meals, at bedtime, between 2 and 3 AM, and with allowance for other times that patients are used to checking when they were at home, such as after meals. Frequent POCT has to be especially borne in mind for patients with hypoglycemia unawareness. Some patients are used to checking with their own home glucose meters in between these times, and we do work with them with the understanding that dose‐change decisions are based on the hospital glucose meter readings.

Dose Adjustments

Continuous subcutaneous insulin infusion dose adjustments for hypoglycemia and hyperglycemia are usually done in 10% to 20% decrements/emncrements. Our Endocrinology team discusses these with the patients and ensures that the new settings are entered into the pump and into the order set.

Basal Dose

In most cases, these situations stretch for greater than 24 hours, such as surgery and the anticipated recovery from anesthesia. We favor long‐acting insulin for basal needs, given 2 hours before discontinuing the pump. The total basal insulin dose per day can be given as the starting long‐acting insulin dose and then adjusted as needed. The total daily basal insulin dose can be retrieved from the insulin pump. In one study on T1DM patients using insulin lispro on the pump, total daily basal dose was given as insulin glargine without adverse effects.[16] If there is concern for hypoglycemia, then the dose can be reduced by 10% to 20%. Care should be made to ascertain that the basal insulin delivered via the pump is appropriate.

There are several ways to estimate this:

• If the daily total basal and the total prandial insulin requirements approximate a 50:50 ratio, then the basal rate via the pump is appropriate.
• If the daily total basal rate via the pump is similar to a weight‐based estimate of the basal dose (often 0.150.2 units/kg/day in patients with T1DM, 0.20.3 units/kg/day for T2DM, and higher in both cases with longer duration of DM or greater insulin resistance), then the basal rate via the pump is appropriate.

Bolus Dose

Patients can still continue to calculate their sensitivity factor and carbohydrate ratio and request for the equivalent dose of insulin. In the ideal situation, bedside nurses would be taught how to calculate this ratio and dose rapid‐acting insulin accordingly should the patient need to come off the insulin pump. Because of the logistic difficulties of making this uniform in our institution, we have worked instead on providing patients with information on the carbohydrate content of their meal tray. If the pump is discontinued, the patient would continue to calculate their prandial insulin based on their carbohydrates ratio and indicate to the nurse how much he/she would need. Our subcutaneous insulin orders for MDII allow for us to put a range of insulin doses based on the patient's typical insulin needs for mealtime.

Pump Removal for Certain Hospital Procedures

Patients may not remember that the pump has to be removed before entering high‐radiation areas. The pump owner manuals tell patients not to use the pump when going for magnetic resonance imaging, CT scans, or x‐rays, or near equipment with high electromagnetic fields.

Interrogation of the Pump

If there is concern about pump malfunction, patients should be switched to MDII. The pump company can be contacted for pump interrogation and provision of a temporary pump (Table 3).

Pump Disconnection

In our institution, the Radiology department has signage instructing patients to inform the technician if they are wearing an insulin pump. The pump is handed off to a family member or stored until the procedure is over. Another option is to leave it with the bedside nurse or the floor nurse manager for safekeeping. This is less ideal, because the wait for the radiologic procedure might take longer than expected and the patient is left without any insulin on board.

Interruption of Continuous Subcutaneous Insulin Infusion for Surgical Procedures

In our institution, the anesthesiology department has worked with the Endocrinology, Surgery and Medicine departments regarding patients with insulin pumps. Discontinuation of CSII is recommended for surgical procedures longer than 1 hour; patients are asked to continue on the insulin pump until they are taken to the preoperative suite, at which point they are placed on IV insulin infusion. Ideally, there should be an overlap of 1530 minutes. Providing an alternative continuous source of insulin during pump interruption is important, especially for patients with T1DM.[17]

Pump Resumption

Once the patient is ready to resume the pump, any subcutaneous insulin that was delivered and might still be active has to be accounted for and subtracted from the basal pump dose so that hypoglycemia is avoided. An alternative would be to wait until the last basal subcutaneous insulin dose is expected to be cleared before restarting CSII.

The Patient Needing Transition to Multiple Daily Insulin Injections

A 29‐year‐old male with T1DM, on CSII, was admitted for fever and chills. His latest glycated hemoglobin (HbA_1c) level was 6.8%. His glucose levels started rising, and he wished to be taken off the insulin pump. He was started by the primary team on multiple daily insulin injections (MDII) with insulin glargine and insulin lispro. His glucose levels continued to rise, so an intravenous (IV) insulin infusion was started. Endocrinology was then consulted. The patient's condition was concerning for the potential development of DKA, so he was kept on IV insulin. When he was ready for transitioning to subcutaneous insulin, the pump had been taken home by a family member, and the patient could not recall his CSII basal rates but knew his total basal insulin dose, carbohydrate ratio, and sensitivity factor. Endocrinology assisted in transitioning him from the insulin infusion to MDII based on these recalled doses. When the insulin pump was available, the pump settings were interrogated, and he was transitioned back to it.

Key points:

• Having key hospital personnel trained on CSII, including interrogating the pump's settings, facilitates the transitioning of these patients from one hospital unit, or level of care, to another.
• Accessing the pump's settings involves pushing several buttons on the insulin pump. Because key hospital personnel will encounter patients on different insulin pumps, it may be helpful to keep menu maps handy as a quick reference. A menu map will show at a glance where certain information can be found, such as the basal insulin rate or the sensitivity factor (see examples in Figures 1 and 2).[12, 13]
• Knowing the HbA_1c will help determine if pump use has been effective.

Figure 1

Figure 2

The Patient With Technical Problems

An 84‐year‐old gentleman with T2DM was admitted for heart failure and aortic valve replacement. His HbA_1c was 6.2%, and he had had several outpatient hypoglycemic events. While on CSII in the hospital, his point of care testing (POCT) glucose readings ranged from 105 to 260 mg/dL. On the afternoon of the third hospital day, POCT readings stayed above 220 mg/dL and rose to 348 mg/dL on the fourth hospital morning, despite multiple blousing and changing the insulin, insertion site, reservoir cartridge, and pump tubing. There was no evidence of infection and no medication change that would have impacted glucose levels. Review of his procedures revealed that he had undergone computed tomography (CT) on the morning of hospital day 3 and wore his pump while being scanned. The pump company was notified.

Key points:

• Patients, and medical and nursing staff, should be reminded to remove insulin pumps for CT scans, magnetic resonance imaging, x‐rays, or other tests with high electromagnetic fields.
• If there is a suspicion of pump malfunction from such a procedure, notify the pump company.

The Patient Who Can Benefit From Inpatient Education

A 70‐year‐old female with T1DM was admitted for heart failure. The patient had been using CSII for 20 years. Her latest HbA_1c was 6.9%. She had 1 hypoglycemic event every 1 to 2 weeks. In the hospital, she experienced 2 hypoglycemic events within 3 days, both around bedtime. It was discovered that the patient was giving a bolus of insulin for elevated glucose levels based on the hospital POCT, and when the meal arrived (3060 minutes later), she again delivered a bolus based on her own glucometer reading plus insulin based on the carbohydrates in her meal. The patient was then instructed to request the POCT when her meal tray arrived, and she was taught how to use the pump's built‐in calculator. Glucose excursions improved.

Key points:

• Patients on CSII, though able to exercise autonomy in managing their insulin doses, may also need assistance in dosing insulin properly.
• Although pump education is ideally done on an outpatient basis, hospital‐based providers may encounter patients who need reinforcement of their training while hospitalized. Hospital personnel trained on insulin pumps (such as physicians, nurse practitioners, physician assistants, and certified diabetes educators) can help augment the patient's knowledge while in the hospital. In the absence of such key personnel, patient safety has to be addressed with re‐evaluation of the need to discontinue the pump and switch to multiple doses of subcutaneous insulin.

Initial Patient Assessment

On admission, patients are asked whether they use an insulin pump. This is included in the nursing assessment form. If they do, the physician is notified.

The insulin pump might be missed unless specifically asked for because (1) the insulin pump may be thought of more as a device rather than a medication, and (2) the insulin pump may be worn in less obvious areas, not only on the abdomen where providers are more apt to detect it.

Hospital Policy and Insulin Orders

Written hospital policies on how to safely manage patients presenting with an insulin pump will delineate patients who can safely be allowed to continue on the pump, and the responsibilities that come along with this. Our institution has such a policy. Experts from both the legal and biomedical engineering departments were consulted when the policy was crafted. Patients must be fully alert, able, and willing to self‐manage the pump. General contraindications to pump use in the hospital, such as altered mental status or DKA, are listed in Table 1. In addition, patients in the intensive care units are best managed on an IV insulin infusion during their critical illness, in keeping with several society guidelines.[14] Controlling severe hyperglycemia and DKA with multiple boluses through the insulin pump can potentially lead to stacking of insulin with subsequent hypoglycemia.

In our institution, a computerized insulin pump order set has to be activated. Apart from insulin, POCT, and hypoglycemia‐management orders, this order set contains documents aimed at balancing patient autonomy with delivery of appropriate and safe medical care that the bedside nurse goes over with the patient (Table 2). By policy at our institution, insulin should be dispensed from the hospital's pharmacy (except for that already in the pump), so the order set is linked to the pharmacy and a 3‐mL insulin vial is delivered to the hospital floor and stored in the patient‐specific medication bin. The order set triggers an Endocrinology consult so that the patient can be assessed by key trained personnel.

Patient Diagnosis

It is important to try to distinguish T1DM vs T2DM, as patients with T1DM are prone to ketoacidosis when the pump is disconnected.

Patient Assessment by the Endocrinology Consult Service

Hospitalized patients on the pump have varying degrees of pump knowledge and skill sets. We have encountered highly trained patients who meticulously count their carbohydrates and double‐check the insulin doses calculated by the built‐in pump calculator, and those who have knowledge gaps because their physicians, and not they themselves, change pump settings at the clinic visits.

Therefore, the Endocrinology consult team members (comprising physicians, nurse practitioners, and certified diabetes educators) go through the following items to be able to order the insulin correctly, assess whether patients are still able and willing to continue on their pump despite their illness, formulate alternative insulin regimens as needed, or help empower patients who may have forgotten some aspects of pump management:

• Insulin pump manufacturer/model.
• Insulin used in the pump.
• Often fast‐acting insulin (lispro, aspart, or glulisine).
• Some patients use regular insulin.
• A few patients use U500 insulin (5 more potent than other insulins).
• Insulin doses/pump settings.

Patients are assessed for:

• Hypoglycemia awareness.
• Previous glucose control.
• Bolus calculation (either using the built‐in calculator, computing this mentally, or using a different calculator).
• Ability to deliver a bolus (including vision and dexterity challenges).
• Ability to change the basal rate, or set a temporary rate, and suspend insulin delivery.

Discussion on Options for Inpatient Management

After assessment, education is provided as needed. If there are concerns on the part of the patient, the primary team, or the Endocrinology team about safe continuation of CSII during the hospitalization, then alternative insulin regimens are discussed. Patients who cannot access their basal rates and cannot adjust the doses are not able to self‐manage; they should be taken off the pump and treated with multiple subcutaneous insulin doses. Conversion to MDII is discussed under Interruption of Continuous Subcutaneous Insulin Infusion for Short and Prolonged Periods.

Provision of Pump Information for Hospital Healthcare Providers

Users of CSII, even if perfectly competent using their pumps in the ambulatory setting, may need assistance in the hospital for various reasons. They may not know what to do for surgical or radiologic procedures (discussed below) and may not be familiar with hospital policies involving CSII. Hospital providers trained on insulin pumps may need a refresher on locating a particular pump setting.

The provider can call the toll‐free number on the back of the pump for assistance (Table 3). Insulin‐pump companies also have menu maps to aid in finding information on pump settings (samples shown in Figures 1 and 2).[12, 13] Documentation of the patient's pump settings will assist in CSII dose changes during the acute illness or assist in switching to MDII if needed. The following information need to be collected:

Point‐of‐Care Testing for Glucose Monitoring

Our policy specifies that the hospital glucose meter is the meter of record upon which dose adjustments are based. Point‐of‐care testing is performed by our patient care nursing assistants or bedside nurses. The timing is typically before meals, at bedtime, between 2 and 3 AM, and with allowance for other times that patients are used to checking when they were at home, such as after meals. Frequent POCT has to be especially borne in mind for patients with hypoglycemia unawareness. Some patients are used to checking with their own home glucose meters in between these times, and we do work with them with the understanding that dose‐change decisions are based on the hospital glucose meter readings.

Dose Adjustments

Continuous subcutaneous insulin infusion dose adjustments for hypoglycemia and hyperglycemia are usually done in 10% to 20% decrements/emncrements. Our Endocrinology team discusses these with the patients and ensures that the new settings are entered into the pump and into the order set.

Basal Dose

In most cases, these situations stretch for greater than 24 hours, such as surgery and the anticipated recovery from anesthesia. We favor long‐acting insulin for basal needs, given 2 hours before discontinuing the pump. The total basal insulin dose per day can be given as the starting long‐acting insulin dose and then adjusted as needed. The total daily basal insulin dose can be retrieved from the insulin pump. In one study on T1DM patients using insulin lispro on the pump, total daily basal dose was given as insulin glargine without adverse effects.[16] If there is concern for hypoglycemia, then the dose can be reduced by 10% to 20%. Care should be made to ascertain that the basal insulin delivered via the pump is appropriate.

There are several ways to estimate this:

• If the daily total basal and the total prandial insulin requirements approximate a 50:50 ratio, then the basal rate via the pump is appropriate.
• If the daily total basal rate via the pump is similar to a weight‐based estimate of the basal dose (often 0.150.2 units/kg/day in patients with T1DM, 0.20.3 units/kg/day for T2DM, and higher in both cases with longer duration of DM or greater insulin resistance), then the basal rate via the pump is appropriate.

Bolus Dose

Patients can still continue to calculate their sensitivity factor and carbohydrate ratio and request for the equivalent dose of insulin. In the ideal situation, bedside nurses would be taught how to calculate this ratio and dose rapid‐acting insulin accordingly should the patient need to come off the insulin pump. Because of the logistic difficulties of making this uniform in our institution, we have worked instead on providing patients with information on the carbohydrate content of their meal tray. If the pump is discontinued, the patient would continue to calculate their prandial insulin based on their carbohydrates ratio and indicate to the nurse how much he/she would need. Our subcutaneous insulin orders for MDII allow for us to put a range of insulin doses based on the patient's typical insulin needs for mealtime.

Pump Removal for Certain Hospital Procedures

Patients may not remember that the pump has to be removed before entering high‐radiation areas. The pump owner manuals tell patients not to use the pump when going for magnetic resonance imaging, CT scans, or x‐rays, or near equipment with high electromagnetic fields.

Interrogation of the Pump

If there is concern about pump malfunction, patients should be switched to MDII. The pump company can be contacted for pump interrogation and provision of a temporary pump (Table 3).

Pump Disconnection

In our institution, the Radiology department has signage instructing patients to inform the technician if they are wearing an insulin pump. The pump is handed off to a family member or stored until the procedure is over. Another option is to leave it with the bedside nurse or the floor nurse manager for safekeeping. This is less ideal, because the wait for the radiologic procedure might take longer than expected and the patient is left without any insulin on board.

Interruption of Continuous Subcutaneous Insulin Infusion for Surgical Procedures

In our institution, the anesthesiology department has worked with the Endocrinology, Surgery and Medicine departments regarding patients with insulin pumps. Discontinuation of CSII is recommended for surgical procedures longer than 1 hour; patients are asked to continue on the insulin pump until they are taken to the preoperative suite, at which point they are placed on IV insulin infusion. Ideally, there should be an overlap of 1530 minutes. Providing an alternative continuous source of insulin during pump interruption is important, especially for patients with T1DM.[17]

Pump Resumption

Once the patient is ready to resume the pump, any subcutaneous insulin that was delivered and might still be active has to be accounted for and subtracted from the basal pump dose so that hypoglycemia is avoided. An alternative would be to wait until the last basal subcutaneous insulin dose is expected to be cleared before restarting CSII.

The Patient Needing Transition to Multiple Daily Insulin Injections

A 29‐year‐old male with T1DM, on CSII, was admitted for fever and chills. His latest glycated hemoglobin (HbA_1c) level was 6.8%. His glucose levels started rising, and he wished to be taken off the insulin pump. He was started by the primary team on multiple daily insulin injections (MDII) with insulin glargine and insulin lispro. His glucose levels continued to rise, so an intravenous (IV) insulin infusion was started. Endocrinology was then consulted. The patient's condition was concerning for the potential development of DKA, so he was kept on IV insulin. When he was ready for transitioning to subcutaneous insulin, the pump had been taken home by a family member, and the patient could not recall his CSII basal rates but knew his total basal insulin dose, carbohydrate ratio, and sensitivity factor. Endocrinology assisted in transitioning him from the insulin infusion to MDII based on these recalled doses. When the insulin pump was available, the pump settings were interrogated, and he was transitioned back to it.

Key points:

• Having key hospital personnel trained on CSII, including interrogating the pump's settings, facilitates the transitioning of these patients from one hospital unit, or level of care, to another.
• Accessing the pump's settings involves pushing several buttons on the insulin pump. Because key hospital personnel will encounter patients on different insulin pumps, it may be helpful to keep menu maps handy as a quick reference. A menu map will show at a glance where certain information can be found, such as the basal insulin rate or the sensitivity factor (see examples in Figures 1 and 2).[12, 13]
• Knowing the HbA_1c will help determine if pump use has been effective.

Figure 1

Figure 2

The Patient With Technical Problems

An 84‐year‐old gentleman with T2DM was admitted for heart failure and aortic valve replacement. His HbA_1c was 6.2%, and he had had several outpatient hypoglycemic events. While on CSII in the hospital, his point of care testing (POCT) glucose readings ranged from 105 to 260 mg/dL. On the afternoon of the third hospital day, POCT readings stayed above 220 mg/dL and rose to 348 mg/dL on the fourth hospital morning, despite multiple blousing and changing the insulin, insertion site, reservoir cartridge, and pump tubing. There was no evidence of infection and no medication change that would have impacted glucose levels. Review of his procedures revealed that he had undergone computed tomography (CT) on the morning of hospital day 3 and wore his pump while being scanned. The pump company was notified.

Key points:

• Patients, and medical and nursing staff, should be reminded to remove insulin pumps for CT scans, magnetic resonance imaging, x‐rays, or other tests with high electromagnetic fields.
• If there is a suspicion of pump malfunction from such a procedure, notify the pump company.

The Patient Who Can Benefit From Inpatient Education

A 70‐year‐old female with T1DM was admitted for heart failure. The patient had been using CSII for 20 years. Her latest HbA_1c was 6.9%. She had 1 hypoglycemic event every 1 to 2 weeks. In the hospital, she experienced 2 hypoglycemic events within 3 days, both around bedtime. It was discovered that the patient was giving a bolus of insulin for elevated glucose levels based on the hospital POCT, and when the meal arrived (3060 minutes later), she again delivered a bolus based on her own glucometer reading plus insulin based on the carbohydrates in her meal. The patient was then instructed to request the POCT when her meal tray arrived, and she was taught how to use the pump's built‐in calculator. Glucose excursions improved.

Key points:

• Patients on CSII, though able to exercise autonomy in managing their insulin doses, may also need assistance in dosing insulin properly.
• Although pump education is ideally done on an outpatient basis, hospital‐based providers may encounter patients who need reinforcement of their training while hospitalized. Hospital personnel trained on insulin pumps (such as physicians, nurse practitioners, physician assistants, and certified diabetes educators) can help augment the patient's knowledge while in the hospital. In the absence of such key personnel, patient safety has to be addressed with re‐evaluation of the need to discontinue the pump and switch to multiple doses of subcutaneous insulin.

Initial Patient Assessment

On admission, patients are asked whether they use an insulin pump. This is included in the nursing assessment form. If they do, the physician is notified.

The insulin pump might be missed unless specifically asked for because (1) the insulin pump may be thought of more as a device rather than a medication, and (2) the insulin pump may be worn in less obvious areas, not only on the abdomen where providers are more apt to detect it.

Hospital Policy and Insulin Orders

Written hospital policies on how to safely manage patients presenting with an insulin pump will delineate patients who can safely be allowed to continue on the pump, and the responsibilities that come along with this. Our institution has such a policy. Experts from both the legal and biomedical engineering departments were consulted when the policy was crafted. Patients must be fully alert, able, and willing to self‐manage the pump. General contraindications to pump use in the hospital, such as altered mental status or DKA, are listed in Table 1. In addition, patients in the intensive care units are best managed on an IV insulin infusion during their critical illness, in keeping with several society guidelines.[14] Controlling severe hyperglycemia and DKA with multiple boluses through the insulin pump can potentially lead to stacking of insulin with subsequent hypoglycemia.

In our institution, a computerized insulin pump order set has to be activated. Apart from insulin, POCT, and hypoglycemia‐management orders, this order set contains documents aimed at balancing patient autonomy with delivery of appropriate and safe medical care that the bedside nurse goes over with the patient (Table 2). By policy at our institution, insulin should be dispensed from the hospital's pharmacy (except for that already in the pump), so the order set is linked to the pharmacy and a 3‐mL insulin vial is delivered to the hospital floor and stored in the patient‐specific medication bin. The order set triggers an Endocrinology consult so that the patient can be assessed by key trained personnel.

Patient Diagnosis

It is important to try to distinguish T1DM vs T2DM, as patients with T1DM are prone to ketoacidosis when the pump is disconnected.

Patient Assessment by the Endocrinology Consult Service

Hospitalized patients on the pump have varying degrees of pump knowledge and skill sets. We have encountered highly trained patients who meticulously count their carbohydrates and double‐check the insulin doses calculated by the built‐in pump calculator, and those who have knowledge gaps because their physicians, and not they themselves, change pump settings at the clinic visits.

Therefore, the Endocrinology consult team members (comprising physicians, nurse practitioners, and certified diabetes educators) go through the following items to be able to order the insulin correctly, assess whether patients are still able and willing to continue on their pump despite their illness, formulate alternative insulin regimens as needed, or help empower patients who may have forgotten some aspects of pump management:

• Insulin pump manufacturer/model.
• Insulin used in the pump.
• Often fast‐acting insulin (lispro, aspart, or glulisine).
• Some patients use regular insulin.
• A few patients use U500 insulin (5 more potent than other insulins).
• Insulin doses/pump settings.

Patients are assessed for:

• Hypoglycemia awareness.
• Previous glucose control.
• Bolus calculation (either using the built‐in calculator, computing this mentally, or using a different calculator).
• Ability to deliver a bolus (including vision and dexterity challenges).
• Ability to change the basal rate, or set a temporary rate, and suspend insulin delivery.

Discussion on Options for Inpatient Management

After assessment, education is provided as needed. If there are concerns on the part of the patient, the primary team, or the Endocrinology team about safe continuation of CSII during the hospitalization, then alternative insulin regimens are discussed. Patients who cannot access their basal rates and cannot adjust the doses are not able to self‐manage; they should be taken off the pump and treated with multiple subcutaneous insulin doses. Conversion to MDII is discussed under Interruption of Continuous Subcutaneous Insulin Infusion for Short and Prolonged Periods.

Provision of Pump Information for Hospital Healthcare Providers

Users of CSII, even if perfectly competent using their pumps in the ambulatory setting, may need assistance in the hospital for various reasons. They may not know what to do for surgical or radiologic procedures (discussed below) and may not be familiar with hospital policies involving CSII. Hospital providers trained on insulin pumps may need a refresher on locating a particular pump setting.

The provider can call the toll‐free number on the back of the pump for assistance (Table 3). Insulin‐pump companies also have menu maps to aid in finding information on pump settings (samples shown in Figures 1 and 2).[12, 13] Documentation of the patient's pump settings will assist in CSII dose changes during the acute illness or assist in switching to MDII if needed. The following information need to be collected:

Point‐of‐Care Testing for Glucose Monitoring

Our policy specifies that the hospital glucose meter is the meter of record upon which dose adjustments are based. Point‐of‐care testing is performed by our patient care nursing assistants or bedside nurses. The timing is typically before meals, at bedtime, between 2 and 3 AM, and with allowance for other times that patients are used to checking when they were at home, such as after meals. Frequent POCT has to be especially borne in mind for patients with hypoglycemia unawareness. Some patients are used to checking with their own home glucose meters in between these times, and we do work with them with the understanding that dose‐change decisions are based on the hospital glucose meter readings.

Dose Adjustments

Continuous subcutaneous insulin infusion dose adjustments for hypoglycemia and hyperglycemia are usually done in 10% to 20% decrements/emncrements. Our Endocrinology team discusses these with the patients and ensures that the new settings are entered into the pump and into the order set.

Basal Dose

In most cases, these situations stretch for greater than 24 hours, such as surgery and the anticipated recovery from anesthesia. We favor long‐acting insulin for basal needs, given 2 hours before discontinuing the pump. The total basal insulin dose per day can be given as the starting long‐acting insulin dose and then adjusted as needed. The total daily basal insulin dose can be retrieved from the insulin pump. In one study on T1DM patients using insulin lispro on the pump, total daily basal dose was given as insulin glargine without adverse effects.[16] If there is concern for hypoglycemia, then the dose can be reduced by 10% to 20%. Care should be made to ascertain that the basal insulin delivered via the pump is appropriate.

There are several ways to estimate this:

• If the daily total basal and the total prandial insulin requirements approximate a 50:50 ratio, then the basal rate via the pump is appropriate.
• If the daily total basal rate via the pump is similar to a weight‐based estimate of the basal dose (often 0.150.2 units/kg/day in patients with T1DM, 0.20.3 units/kg/day for T2DM, and higher in both cases with longer duration of DM or greater insulin resistance), then the basal rate via the pump is appropriate.

Bolus Dose

Patients can still continue to calculate their sensitivity factor and carbohydrate ratio and request for the equivalent dose of insulin. In the ideal situation, bedside nurses would be taught how to calculate this ratio and dose rapid‐acting insulin accordingly should the patient need to come off the insulin pump. Because of the logistic difficulties of making this uniform in our institution, we have worked instead on providing patients with information on the carbohydrate content of their meal tray. If the pump is discontinued, the patient would continue to calculate their prandial insulin based on their carbohydrates ratio and indicate to the nurse how much he/she would need. Our subcutaneous insulin orders for MDII allow for us to put a range of insulin doses based on the patient's typical insulin needs for mealtime.

Pump Removal for Certain Hospital Procedures

Patients may not remember that the pump has to be removed before entering high‐radiation areas. The pump owner manuals tell patients not to use the pump when going for magnetic resonance imaging, CT scans, or x‐rays, or near equipment with high electromagnetic fields.

Interrogation of the Pump

If there is concern about pump malfunction, patients should be switched to MDII. The pump company can be contacted for pump interrogation and provision of a temporary pump (Table 3).

Pump Disconnection

In our institution, the Radiology department has signage instructing patients to inform the technician if they are wearing an insulin pump. The pump is handed off to a family member or stored until the procedure is over. Another option is to leave it with the bedside nurse or the floor nurse manager for safekeeping. This is less ideal, because the wait for the radiologic procedure might take longer than expected and the patient is left without any insulin on board.

Interruption of Continuous Subcutaneous Insulin Infusion for Surgical Procedures

In our institution, the anesthesiology department has worked with the Endocrinology, Surgery and Medicine departments regarding patients with insulin pumps. Discontinuation of CSII is recommended for surgical procedures longer than 1 hour; patients are asked to continue on the insulin pump until they are taken to the preoperative suite, at which point they are placed on IV insulin infusion. Ideally, there should be an overlap of 1530 minutes. Providing an alternative continuous source of insulin during pump interruption is important, especially for patients with T1DM.[17]

Pump Resumption

Once the patient is ready to resume the pump, any subcutaneous insulin that was delivered and might still be active has to be accounted for and subtracted from the basal pump dose so that hypoglycemia is avoided. An alternative would be to wait until the last basal subcutaneous insulin dose is expected to be cleared before restarting CSII.

Setting

C. S. Mott Children's Hospital at the University of Michigan is a suburban academic, tertiary care hospital located in Ann Arbor, Michigan. The pediatric emergency department had approximately 20,000 visits per year during the study period, and children were treated in an 11‐bed area. There were 106 pediatric inpatient general care beds during the study years.

Visit Selection

ED visits made by children <18 years old during the 3‐year period between May 1, 2007 and April 30, 2010 were considered eligible for study. Visits were included in the study based on the presence of International Classification of Disease, 9th Revision, Clinical Modification (ICD‐9‐CM) codes for respiratory conditions (asthma [493.xx], pneumonia [480.x‐486.x], croup [464.4], bronchiolitis [466.x], acute bronchospasm [519.11], wheezing [786.07]) in the top 3 ED diagnoses. Visits were excluded if the hospital discharge diagnoses included ICD‐9‐CM codes determined a priori by the study team to represent severe illness (e.g., respiratory failure) or complicated medical conditions (e.g., ventilator dependence) unlikely to respond to short stay care (see Supporting Appendices 1 through 3 in the online version of this article). ED visits that resulted in an admission to the intensive care unit (ICU), a procedure, or telemetry were excluded because these children would be unlikely to respond to short stay care.

Variables

Subject demographic characteristics included age, gender, race, and payer. Visit‐related variables included date and time of ED arrival, date and time of ED disposition (to home or admission), and date and time of hospital discharge if admitted. Arrival date and time were used to determine season (winter: DecemberMarch, spring/summer: AprilAugust, fall: SeptemberNovember) and shift (day: 8 _am4 _pm, evening: 4 _pmmidnight, night: midnight8 _am).

Outcome

LOS was the primary outcome. ED LOS was calculated by subtracting the date and time of ED arrival from the date and time of ED disposition. Inpatient LOS was calculated by subtracting the date and time of arrival to the inpatient unit from the date and time of discharge from the hospital admission, discharge, transfer (ADT) system. Total LOS was calculated by subtracting the date and time of ED arrival from the date and time of discharge from either the ED or hospital as appropriate.

Analyses

Descriptive statistics were calculated for patient demographics and actual LOS. To describe discharge timing patterns in the ED and inpatient unit, 24 24, 576‐cell matrices were produced with arrival hour along a horizontal axis and discharge hour along a vertical axis. Visits were grouped by hour of ED or inpatient arrival and the number of these visits that were discharged in each hour was plotted in the appropriate cell in the matrix. Cells were then shaded according to the proportion of discharges that occurred at that hour, for each hour of arrival.

Finally, Monte Carlo simulations were designed to illustrate the impact of the timing and frequency of rounds on inpatient LOS. Historical arrival rates were utilized in the simulations. Because medication administration times were not available in the dataset, active treatment times were fixed at 12 hours and 20 hours. The fixed treatment time of 12 hours was based on research by McConnochie et al.,[11] which found that roughly 65% of children hospitalized for asthma received frequent nebulized treatments for 16 hours (two 8‐hour nursing shifts). Four hours were subtracted from this treatment time because our sample included ED treatment times, and local experience demonstrates that children receive at least 4 hours of ED treatment prior to transfer to an inpatient bed. The 20‐hour fixed treatment time was selected to simulate a sicker population of children.

We assumed that physician‐led rounds were the decision point leading to hospital discharge, and that, at minimum, rounds occurred once each morning. The simulation began with 1 physician‐led rounding session occurring at 9 _am. Rounding sessions were then added to the model up to a maximum of 24 hourly rounds. Rounding every hour was considered to be analogous to the 24‐hour operation of an ED, in which patients are discharged at all hours. A 4‐hour lag time between the start of first morning rounds and patient discharge was assumed based on historical discharge times. The lag time for additional rounding times was varied, from a minimum of 2 hours for afternoon rounds to a maximum of 10 hours for evening rounds to allow the patient to sleep. Each simulation completed 10,000 iterations and 95% confidence intervals (CIs) were calculated. All analyses were conducted in Excel 2010 (Microsoft Corp., Redmond, WA).

Study Sample Characteristics

Of the 57,639 pediatric ED visits during the 3‐year study period, 5699 were diagnosed with respiratory conditions considered eligible for this study, and 5503 (96.6%) were included in analyses. Hospitalization was the outcome for 1285 (23.4%) ED visits. Demographic characteristics of the study sample are presented in Table 1. Almost half of the visits were by children <3 years old. Visits for respiratory conditions peaked in the fall and winter months. Daily peaks in ED arrivals occurred in the late afternoon.

Length of Stay

For visits meeting study criteria, median ED LOS was 3.0 hours (interquartile range [IQR] 2.1‐4.1) among ED discharges and 5.1 hours (IQR 4.0‐6.6) among admissions. Of the inpatients, 8.4% were admitted and discharged on the same day, 37.3% were admitted for 1 night, 27.0% were admitted for 2 nights, and 27.3% were admitted for 3 or more nights. Median inpatient LOS was 41.2 hours (IQR 23.4‐66.4), with median total LOS of 46.7 hours (IQR 29.3‐71.5).

Arrival and Discharge Patterns

Figure 1 illustrates the relationship between the ED arrival and disposition times for all included visits. Figure 2 illustrates the relationship between the inpatient arrival and discharge times among the subset of visits that received inpatient care. Children with respiratory illness arrived to the ED at all hours of the day and night and generally were discharged 2 to 5 hours after arrival regardless of the time of day. In contrast, children admitted to the inpatient setting were most commonly discharged between 11 _am and 6 _pm, regardless of the time of admission.

Figure 1

Figure 2

Simulation

The influence of timing and frequency of rounds on inpatient LOS is shown in Table 2. Once‐daily rounds resulted in a predicted LOS of 27.24 hours (95% CI: 27.0727.41) for the 12‐hour active treatment time simulation and 36.93 hours (95% CI: 36.70‐37.16) for the 20‐hour active treatment time. There were incremental reductions in predicted LOS with each additional rounding session. When the number of rounding sessions reached 24, predicted LOS was 14.5 hours (95% CI: 14.49‐14.51) for the 12‐hour active treatment time simulation and 22.5 hours (95% CI: 22.49‐22.51) for the 20‐hour active treatment time.

Limitations

There are several limitations to our study. First, the time of discharge is recorded differently for the ED and inpatient settings in our administrative dataset as a function of the electronic medical record systems in these 2 environments. Discharge from the ED reflects that time the patient left the ED but discharge from the inpatient setting reflects the time the patient exits the hospital ADT system. This likely biases our results toward longer LOS in the inpatient setting. We expect this would shift the band of inpatient discharges later in the day but do not expect this to alter the observed difference in discharge patterns between the ED and inpatient settings.

Second, administrative data do not provide information about the physicians making the discharge decisions or valid reasons a child would remain in the hospital after they improved clinically including time for teaching or care coordination. Our results therefore overestimate the amount of excess time associated with inpatient care. Future research is needed to determine the actual duration of time between when a child is clinically well enough for discharge home and when a child is actually discharged. There is also a need to understand discharge decision making and to identify the nonclinical factors that contribute to discharge delays.

Third, our data sample was taken from visits made by children to a pediatric ED nested within a general ED where children are admitted to an adjoining tertiary care academic children's hospital. Our results may not be generalizable to other settings in which children receive hospital‐based care, including freestanding children's hospitals and general hospitals that admit children.

Fourth, we focused our analyses on children with respiratory conditions, excluding visits with diagnoses suggestive of complex comorbid conditions and severe illnesses. Although we anticipate that children with these respiratory conditions treated in other hospitals will be similar to our population, our results may not hold for other conditions or across the full spectrum of severity for acute respiratory illness treated within EDs and inpatient units.

Finally, our simulation model makes assumptions, such as a fixed treatment times and discharge process length, which do not capture the clinical nuances of an individual child's response to hospital‐based treatment. Though much of the clinical complexity of hospital operations is not taken into account in the model, our intended purpose, to explore the influence of different rounding times on LOS, remains valid.

Setting

C. S. Mott Children's Hospital at the University of Michigan is a suburban academic, tertiary care hospital located in Ann Arbor, Michigan. The pediatric emergency department had approximately 20,000 visits per year during the study period, and children were treated in an 11‐bed area. There were 106 pediatric inpatient general care beds during the study years.

Visit Selection

ED visits made by children <18 years old during the 3‐year period between May 1, 2007 and April 30, 2010 were considered eligible for study. Visits were included in the study based on the presence of International Classification of Disease, 9th Revision, Clinical Modification (ICD‐9‐CM) codes for respiratory conditions (asthma [493.xx], pneumonia [480.x‐486.x], croup [464.4], bronchiolitis [466.x], acute bronchospasm [519.11], wheezing [786.07]) in the top 3 ED diagnoses. Visits were excluded if the hospital discharge diagnoses included ICD‐9‐CM codes determined a priori by the study team to represent severe illness (e.g., respiratory failure) or complicated medical conditions (e.g., ventilator dependence) unlikely to respond to short stay care (see Supporting Appendices 1 through 3 in the online version of this article). ED visits that resulted in an admission to the intensive care unit (ICU), a procedure, or telemetry were excluded because these children would be unlikely to respond to short stay care.

Variables

Subject demographic characteristics included age, gender, race, and payer. Visit‐related variables included date and time of ED arrival, date and time of ED disposition (to home or admission), and date and time of hospital discharge if admitted. Arrival date and time were used to determine season (winter: DecemberMarch, spring/summer: AprilAugust, fall: SeptemberNovember) and shift (day: 8 _am4 _pm, evening: 4 _pmmidnight, night: midnight8 _am).

Outcome

LOS was the primary outcome. ED LOS was calculated by subtracting the date and time of ED arrival from the date and time of ED disposition. Inpatient LOS was calculated by subtracting the date and time of arrival to the inpatient unit from the date and time of discharge from the hospital admission, discharge, transfer (ADT) system. Total LOS was calculated by subtracting the date and time of ED arrival from the date and time of discharge from either the ED or hospital as appropriate.

Analyses

Descriptive statistics were calculated for patient demographics and actual LOS. To describe discharge timing patterns in the ED and inpatient unit, 24 24, 576‐cell matrices were produced with arrival hour along a horizontal axis and discharge hour along a vertical axis. Visits were grouped by hour of ED or inpatient arrival and the number of these visits that were discharged in each hour was plotted in the appropriate cell in the matrix. Cells were then shaded according to the proportion of discharges that occurred at that hour, for each hour of arrival.

Finally, Monte Carlo simulations were designed to illustrate the impact of the timing and frequency of rounds on inpatient LOS. Historical arrival rates were utilized in the simulations. Because medication administration times were not available in the dataset, active treatment times were fixed at 12 hours and 20 hours. The fixed treatment time of 12 hours was based on research by McConnochie et al.,[11] which found that roughly 65% of children hospitalized for asthma received frequent nebulized treatments for 16 hours (two 8‐hour nursing shifts). Four hours were subtracted from this treatment time because our sample included ED treatment times, and local experience demonstrates that children receive at least 4 hours of ED treatment prior to transfer to an inpatient bed. The 20‐hour fixed treatment time was selected to simulate a sicker population of children.

We assumed that physician‐led rounds were the decision point leading to hospital discharge, and that, at minimum, rounds occurred once each morning. The simulation began with 1 physician‐led rounding session occurring at 9 _am. Rounding sessions were then added to the model up to a maximum of 24 hourly rounds. Rounding every hour was considered to be analogous to the 24‐hour operation of an ED, in which patients are discharged at all hours. A 4‐hour lag time between the start of first morning rounds and patient discharge was assumed based on historical discharge times. The lag time for additional rounding times was varied, from a minimum of 2 hours for afternoon rounds to a maximum of 10 hours for evening rounds to allow the patient to sleep. Each simulation completed 10,000 iterations and 95% confidence intervals (CIs) were calculated. All analyses were conducted in Excel 2010 (Microsoft Corp., Redmond, WA).

Study Sample Characteristics

Of the 57,639 pediatric ED visits during the 3‐year study period, 5699 were diagnosed with respiratory conditions considered eligible for this study, and 5503 (96.6%) were included in analyses. Hospitalization was the outcome for 1285 (23.4%) ED visits. Demographic characteristics of the study sample are presented in Table 1. Almost half of the visits were by children <3 years old. Visits for respiratory conditions peaked in the fall and winter months. Daily peaks in ED arrivals occurred in the late afternoon.

Length of Stay

For visits meeting study criteria, median ED LOS was 3.0 hours (interquartile range [IQR] 2.1‐4.1) among ED discharges and 5.1 hours (IQR 4.0‐6.6) among admissions. Of the inpatients, 8.4% were admitted and discharged on the same day, 37.3% were admitted for 1 night, 27.0% were admitted for 2 nights, and 27.3% were admitted for 3 or more nights. Median inpatient LOS was 41.2 hours (IQR 23.4‐66.4), with median total LOS of 46.7 hours (IQR 29.3‐71.5).

Arrival and Discharge Patterns

Figure 1 illustrates the relationship between the ED arrival and disposition times for all included visits. Figure 2 illustrates the relationship between the inpatient arrival and discharge times among the subset of visits that received inpatient care. Children with respiratory illness arrived to the ED at all hours of the day and night and generally were discharged 2 to 5 hours after arrival regardless of the time of day. In contrast, children admitted to the inpatient setting were most commonly discharged between 11 _am and 6 _pm, regardless of the time of admission.

Figure 1

Figure 2

Simulation

The influence of timing and frequency of rounds on inpatient LOS is shown in Table 2. Once‐daily rounds resulted in a predicted LOS of 27.24 hours (95% CI: 27.0727.41) for the 12‐hour active treatment time simulation and 36.93 hours (95% CI: 36.70‐37.16) for the 20‐hour active treatment time. There were incremental reductions in predicted LOS with each additional rounding session. When the number of rounding sessions reached 24, predicted LOS was 14.5 hours (95% CI: 14.49‐14.51) for the 12‐hour active treatment time simulation and 22.5 hours (95% CI: 22.49‐22.51) for the 20‐hour active treatment time.

Limitations

There are several limitations to our study. First, the time of discharge is recorded differently for the ED and inpatient settings in our administrative dataset as a function of the electronic medical record systems in these 2 environments. Discharge from the ED reflects that time the patient left the ED but discharge from the inpatient setting reflects the time the patient exits the hospital ADT system. This likely biases our results toward longer LOS in the inpatient setting. We expect this would shift the band of inpatient discharges later in the day but do not expect this to alter the observed difference in discharge patterns between the ED and inpatient settings.

Second, administrative data do not provide information about the physicians making the discharge decisions or valid reasons a child would remain in the hospital after they improved clinically including time for teaching or care coordination. Our results therefore overestimate the amount of excess time associated with inpatient care. Future research is needed to determine the actual duration of time between when a child is clinically well enough for discharge home and when a child is actually discharged. There is also a need to understand discharge decision making and to identify the nonclinical factors that contribute to discharge delays.

Third, our data sample was taken from visits made by children to a pediatric ED nested within a general ED where children are admitted to an adjoining tertiary care academic children's hospital. Our results may not be generalizable to other settings in which children receive hospital‐based care, including freestanding children's hospitals and general hospitals that admit children.

Fourth, we focused our analyses on children with respiratory conditions, excluding visits with diagnoses suggestive of complex comorbid conditions and severe illnesses. Although we anticipate that children with these respiratory conditions treated in other hospitals will be similar to our population, our results may not hold for other conditions or across the full spectrum of severity for acute respiratory illness treated within EDs and inpatient units.

Finally, our simulation model makes assumptions, such as a fixed treatment times and discharge process length, which do not capture the clinical nuances of an individual child's response to hospital‐based treatment. Though much of the clinical complexity of hospital operations is not taken into account in the model, our intended purpose, to explore the influence of different rounding times on LOS, remains valid.

Setting

C. S. Mott Children's Hospital at the University of Michigan is a suburban academic, tertiary care hospital located in Ann Arbor, Michigan. The pediatric emergency department had approximately 20,000 visits per year during the study period, and children were treated in an 11‐bed area. There were 106 pediatric inpatient general care beds during the study years.

Visit Selection

ED visits made by children <18 years old during the 3‐year period between May 1, 2007 and April 30, 2010 were considered eligible for study. Visits were included in the study based on the presence of International Classification of Disease, 9th Revision, Clinical Modification (ICD‐9‐CM) codes for respiratory conditions (asthma [493.xx], pneumonia [480.x‐486.x], croup [464.4], bronchiolitis [466.x], acute bronchospasm [519.11], wheezing [786.07]) in the top 3 ED diagnoses. Visits were excluded if the hospital discharge diagnoses included ICD‐9‐CM codes determined a priori by the study team to represent severe illness (e.g., respiratory failure) or complicated medical conditions (e.g., ventilator dependence) unlikely to respond to short stay care (see Supporting Appendices 1 through 3 in the online version of this article). ED visits that resulted in an admission to the intensive care unit (ICU), a procedure, or telemetry were excluded because these children would be unlikely to respond to short stay care.

Variables

Subject demographic characteristics included age, gender, race, and payer. Visit‐related variables included date and time of ED arrival, date and time of ED disposition (to home or admission), and date and time of hospital discharge if admitted. Arrival date and time were used to determine season (winter: DecemberMarch, spring/summer: AprilAugust, fall: SeptemberNovember) and shift (day: 8 _am4 _pm, evening: 4 _pmmidnight, night: midnight8 _am).

Outcome

LOS was the primary outcome. ED LOS was calculated by subtracting the date and time of ED arrival from the date and time of ED disposition. Inpatient LOS was calculated by subtracting the date and time of arrival to the inpatient unit from the date and time of discharge from the hospital admission, discharge, transfer (ADT) system. Total LOS was calculated by subtracting the date and time of ED arrival from the date and time of discharge from either the ED or hospital as appropriate.

Analyses

Descriptive statistics were calculated for patient demographics and actual LOS. To describe discharge timing patterns in the ED and inpatient unit, 24 24, 576‐cell matrices were produced with arrival hour along a horizontal axis and discharge hour along a vertical axis. Visits were grouped by hour of ED or inpatient arrival and the number of these visits that were discharged in each hour was plotted in the appropriate cell in the matrix. Cells were then shaded according to the proportion of discharges that occurred at that hour, for each hour of arrival.

Finally, Monte Carlo simulations were designed to illustrate the impact of the timing and frequency of rounds on inpatient LOS. Historical arrival rates were utilized in the simulations. Because medication administration times were not available in the dataset, active treatment times were fixed at 12 hours and 20 hours. The fixed treatment time of 12 hours was based on research by McConnochie et al.,[11] which found that roughly 65% of children hospitalized for asthma received frequent nebulized treatments for 16 hours (two 8‐hour nursing shifts). Four hours were subtracted from this treatment time because our sample included ED treatment times, and local experience demonstrates that children receive at least 4 hours of ED treatment prior to transfer to an inpatient bed. The 20‐hour fixed treatment time was selected to simulate a sicker population of children.

We assumed that physician‐led rounds were the decision point leading to hospital discharge, and that, at minimum, rounds occurred once each morning. The simulation began with 1 physician‐led rounding session occurring at 9 _am. Rounding sessions were then added to the model up to a maximum of 24 hourly rounds. Rounding every hour was considered to be analogous to the 24‐hour operation of an ED, in which patients are discharged at all hours. A 4‐hour lag time between the start of first morning rounds and patient discharge was assumed based on historical discharge times. The lag time for additional rounding times was varied, from a minimum of 2 hours for afternoon rounds to a maximum of 10 hours for evening rounds to allow the patient to sleep. Each simulation completed 10,000 iterations and 95% confidence intervals (CIs) were calculated. All analyses were conducted in Excel 2010 (Microsoft Corp., Redmond, WA).

Study Sample Characteristics

Of the 57,639 pediatric ED visits during the 3‐year study period, 5699 were diagnosed with respiratory conditions considered eligible for this study, and 5503 (96.6%) were included in analyses. Hospitalization was the outcome for 1285 (23.4%) ED visits. Demographic characteristics of the study sample are presented in Table 1. Almost half of the visits were by children <3 years old. Visits for respiratory conditions peaked in the fall and winter months. Daily peaks in ED arrivals occurred in the late afternoon.

Length of Stay

For visits meeting study criteria, median ED LOS was 3.0 hours (interquartile range [IQR] 2.1‐4.1) among ED discharges and 5.1 hours (IQR 4.0‐6.6) among admissions. Of the inpatients, 8.4% were admitted and discharged on the same day, 37.3% were admitted for 1 night, 27.0% were admitted for 2 nights, and 27.3% were admitted for 3 or more nights. Median inpatient LOS was 41.2 hours (IQR 23.4‐66.4), with median total LOS of 46.7 hours (IQR 29.3‐71.5).

Arrival and Discharge Patterns

Figure 1 illustrates the relationship between the ED arrival and disposition times for all included visits. Figure 2 illustrates the relationship between the inpatient arrival and discharge times among the subset of visits that received inpatient care. Children with respiratory illness arrived to the ED at all hours of the day and night and generally were discharged 2 to 5 hours after arrival regardless of the time of day. In contrast, children admitted to the inpatient setting were most commonly discharged between 11 _am and 6 _pm, regardless of the time of admission.

Figure 1

Figure 2

Simulation

The influence of timing and frequency of rounds on inpatient LOS is shown in Table 2. Once‐daily rounds resulted in a predicted LOS of 27.24 hours (95% CI: 27.0727.41) for the 12‐hour active treatment time simulation and 36.93 hours (95% CI: 36.70‐37.16) for the 20‐hour active treatment time. There were incremental reductions in predicted LOS with each additional rounding session. When the number of rounding sessions reached 24, predicted LOS was 14.5 hours (95% CI: 14.49‐14.51) for the 12‐hour active treatment time simulation and 22.5 hours (95% CI: 22.49‐22.51) for the 20‐hour active treatment time.

Limitations

There are several limitations to our study. First, the time of discharge is recorded differently for the ED and inpatient settings in our administrative dataset as a function of the electronic medical record systems in these 2 environments. Discharge from the ED reflects that time the patient left the ED but discharge from the inpatient setting reflects the time the patient exits the hospital ADT system. This likely biases our results toward longer LOS in the inpatient setting. We expect this would shift the band of inpatient discharges later in the day but do not expect this to alter the observed difference in discharge patterns between the ED and inpatient settings.

Second, administrative data do not provide information about the physicians making the discharge decisions or valid reasons a child would remain in the hospital after they improved clinically including time for teaching or care coordination. Our results therefore overestimate the amount of excess time associated with inpatient care. Future research is needed to determine the actual duration of time between when a child is clinically well enough for discharge home and when a child is actually discharged. There is also a need to understand discharge decision making and to identify the nonclinical factors that contribute to discharge delays.

Third, our data sample was taken from visits made by children to a pediatric ED nested within a general ED where children are admitted to an adjoining tertiary care academic children's hospital. Our results may not be generalizable to other settings in which children receive hospital‐based care, including freestanding children's hospitals and general hospitals that admit children.

Fourth, we focused our analyses on children with respiratory conditions, excluding visits with diagnoses suggestive of complex comorbid conditions and severe illnesses. Although we anticipate that children with these respiratory conditions treated in other hospitals will be similar to our population, our results may not hold for other conditions or across the full spectrum of severity for acute respiratory illness treated within EDs and inpatient units.

Finally, our simulation model makes assumptions, such as a fixed treatment times and discharge process length, which do not capture the clinical nuances of an individual child's response to hospital‐based treatment. Though much of the clinical complexity of hospital operations is not taken into account in the model, our intended purpose, to explore the influence of different rounding times on LOS, remains valid.