Catherine L. Liang, MPH

Diabetes mellitus and/or inpatient hyperglycemia are common comorbid conditions in hospitalized patients. Recent surveys show that over 90% of hospitalized diabetic patients experience hyperglycemia (>200 mg/dL), and in nearly 1 in 5 of these patients hyperglycemia persists for 3 days or more.1 Hyperglycemia among inpatients without a previous history of diabetes mellitus is also very common.2 Observational studies have shown that hyperglycemia in hospitalized patients is associated with adverse outcomes including infectious complications, increased length of stay, and increased mortality.27 Recent randomized controlled trials have demonstrated that aggressive treatment of inpatient hyperglycemia improves outcomes in surgical and medical intensive care units.8, 9

Based on the available data, the American Diabetes Association (ADA) now advocates good metabolic control, defined as preprandial glucose levels of 90 to 130 mg/dL and peak postprandial glucose levels <180 mg/dL in hospitalized nonintensive care unit (ICU) patients.10 To reach these targets, the ADA and American College of Endocrinology (ACE) suggest that multidisciplinary teams develop and implement hyperglycemia management guidelines and protocols.11 Protocols should promote the use of continuous intravenous insulin infusions or scheduled basal‐bolus subcutaneous insulin regimens. Subcutaneous insulin protocols should include target glucose levels, basal, nutritional, and supplemental insulin, and daily dose adjustments.6 A recent randomized controlled trial of non‐ICU inpatients demonstrated that such a basal‐bolus insulin regimen results in improved glucose control compared with a sliding scale only regimen.12

To date, few published studies have investigated the best ways to implement such management protocols; those that have are often resource‐intensive, for example involving daily involvement of nurse practitioners or diabetologists.13, 14 It is therefore not known how best to implement an inpatient diabetes management program that is effective, efficient, and self‐perpetuating. At Brigham and Women's Hospital (BWH), we have been refining a subcutaneous insulin protocol, focused provider education, and more recently a computerized order set to overcome barriers related to fear of hypoglycemia, delays in insulin prescribing, and unfamiliarity with inpatient glucose management.15 The aims of this current trial were to evaluate the effects of these interventions on a geographically localized general medical service previously naive to these interventions to evaluate their effects on glycemic control, patient safety, and processes of care. We hypothesized that these interventions would improve glycemic control and increase use of basal‐bolus insulin orders without increasing the rate of hypoglycemia.

METHODS

RESULTS

We prospectively identified 248 potential patients for the study. We subsequently excluded 79 patients for the following reasons: no glucose readings beyond hospital day 1 while on PACE service (34 patients); never admitted to PACE service (15 patients); no diabetes or inpatient hyperglycemia (9 patients, mostly patients prescribed an insulin sliding scale prophylactically to avoid steroid‐induced hyperglycemia); type 1 diabetes (13 patients); TPN, DKA, or HHS (5 patients); and palliative care (3 patients). The remaining 169 patients included 63 from the preintervention period(out of 489 total admissions to the PACE service; 13%) and 106 patients in the postintervention period (out of 565 admissions; 19%). These patients had 2447 glucose readings, or an average of 3.6 glucose readings per monitored patient‐day in the preintervention period and 3.3 glucose readings per patient‐day in the postintervention period. Even including the 34 patients who were excluded for lack of glucose readings, glucose data were still available for 717 out of a potential 775 patient‐days (93%). Characteristics for all included patients are shown in Table 2. The mean admission glucose was 197 mg/dL, mean A1C was 8.4%, 54% of the patients were prescribed insulin prior to admission, and 7% had no prior diagnosis of diabetes. There were no significant differences in baseline characteristics between the 2 patient groups except for Charlson score, which was higher in the preintervention group (87% versus 74% with score 2 or higher; Table 2). The top diagnosis‐related groups for the entire cohort included: heart failure and shock (12 patients); kidney and urinary tract infections (12 patients); esophagitis, gastroenteritis, and miscellaneous digestive disorders (11 patients); chronic obstructive pulmonary disease (10 patients); renal failure (10 patients); simple pneumonia and pleurisy (7 patients); disorders of the pancreas except malignancy (6 patients); chest pain (5 patients); and cellulitis (5 patients).

With respect to insulin ordering practices, there was no significant difference in the use of basal insulin in hyperglycemic patients between the preintervention period and postintervention period (81% versus 91%; P = 0.17), nor in the dose of basal insulin prescribed (results not shown), but there was an increase in the use of scheduled nutritional insulin for those patients with hyperglycemia receiving nutrition: 40% versus 75%, P < 0.001 (Table 3). The percent of patients receiving supplemental (sliding scale) insulin by itself (ie, without ever receiving basal or nutritional insulin) was lower during the postintervention period (29% versus 8%, P < 0.001). Nonsignificant differences were seen in the rates of prescribing an appropriate dose and type of nutritional insulin. Notably, there was no difference at all in the proportion of patient‐days in which insulin adjustments were made when 2 or more episodes of hyperglycemia or hypoglycemia were present during the previous day (56% of patient‐days in both groups; P = 0.90).

The primary outcome, the mean percent of glucose readings between 60 and 180 mg/dL per patient, was 59.1% in the preintervention period and 64.7% in the postintervention (P = 0.13 in unadjusted analysis; Table 3). When adjusted for A1C, admission glucose, and insulin use prior to admission, the adjusted absolute difference in the percent of glucose readings within range was 9.7% (95% confidence interval [CI], 0.6%‐18.8%; P = 0.04; Table 3). Regarding other measures of glucose control, the patient‐day weighted mean glucose was 174.7 mg/dL in the preintervention period and 164.6 mg/dL postintervention (P = 0.02), and there was no significant difference in the percent of patient‐days with any hypoglycemia (glucose <60 mg/dL) or severe hypoglycemia (glucose <40 mg/dL; Table 3). There were also no significant differences in the mean number of hypoglycemic events per patient‐day (6.8 versus 6.6 per 100 patient‐days; relative risk, 0.95; 95% CI, 0.541.67; P = 0.87) or severe hypoglycemic events per patient‐day (1.0 versus 1.4 per 100 patient‐days; relative risk, 1.38; 95% CI, 0.355.53; P = 0.65).

We also compared hospital length of stay in hours between the study groups (Table 3). Length of stay (LOS) was shorter in the postintervention arm in unadjusted analyses (112 versus 86 hours; P < 0.001), and this difference persisted when adjusted for patient insurance, race, gender, and Charlson comorbidity score (25% shorter; 95% CI, 6%‐44%). A comparison of LOS among nonstudy patients on the PACE service during these 2 time periods revealed no difference (105 versus 101 hours). When the length of stay analysis was limited to study patients with a known diagnosis of diabetes, the adjusted effect size was a 31% relative decrease in length of stay.

Figure 1A shows the percent glucose readings within range per patient by hospital day. The greatest differences between groups can be seen on hospital days 2 and 3 (11% absolute differences on both days). Similarly, Figure 1B shows the mean glucose per patient by hospital day. Again, the biggest differences are seen on hospital days 2 and 3 (20 and 23 mg/dL difference between groups, respectively). In both cases, only the day 3 comparisons were significantly different between study groups.

Figure 1

DISCUSSION

In this before‐after study, we found that a multifaceted intervention consisting of a subcutaneous insulin protocol, focused education, and an order set built into the hospital's CPOE system was associated with a significantly higher percentage of glucose readings within range per patient in analyses adjusted for patient demographics and severity of diabetes. We also found a significant decrease in patient‐day weighted mean glucose, a marked increase in appropriate use of scheduled nutritional insulin, and a concomitant decrease in sliding scale insulin only regimens during the postintervention period. Moreover, we found a shorter length of stay during the postintervention period that persisted after adjustment for several clinical factors. Importantly, the interventions described in this study require very few resources to continue indefinitely: printing costs for the management protocol, 4 hours of education delivered per year, and routine upkeep of an electronic order set.

Because this was a before‐after study, we cannot exclude the possibility that these improvements in process and outcome were due to cointerventions and/or temporal trends. However, we know of no other interventions aimed at improving diabetes care in this self‐contained service of nurses, PAs, and hospitalists. Moreover, the process improvements, especially the increase in scheduled nutritional insulin, were rather marked, unlikely to be due to temporal trends alone, and likely capable of producing the corresponding improvements in glucose control. That glucose control stopped improving after hospital day 3 may be due to the fact that subsequent adjustment to insulin orders occurred infrequently and no more often than prior to the intervention. That we did not see greater improvements in glycemic control overall may also reflect the fact that 81% of study patients with inpatient hyperglycemia received basal insulin prior to the intervention.

The reduction in patient LOS was somewhat surprising given the relatively small sample size. However, the results are consistent with those of other studies linking hyperglycemia to LOS18, 19 and we found no evidence for a temporal trend toward lower LOS on the PACE service as a whole during the same time period. While a greater proportion of patients on the PACE service were in the study in the post‐intervention period compared with the preintervention period, we found no evidence that the difference in length of stay was due to increased surveillance for nondiabetics, especially because eligibility criteria depended on phlebotomy glucose values, which were uniformly tested in all inpatients. Also, effects on length of stay were actually stronger when limited to patients with known diabetes. Finally, we controlled for several predictors of length of stay, although we still cannot exclude the possibility of unmeasured confounding between groups.

Since ADA and ACE issued guidelines for inpatient management of diabetes and hyperglycemia, many institutions have developed subcutaneous insulin algorithms, educational curricula, and/or order sets to increase compliance with these guidelines and improve glycemic control. Some of these efforts have been studied and some have been successful in their efforts.13, 14, 2023 Unfortunately, most of these programs have not rigorously assessed their impact on process and outcomes, and the most effective studies published to date have involved interventions much more intensive than those described here. For example, Rush University's intervention was associated with a 50 mg/dL decrease in mean blood glucose but involved an endocrinologist rounding twice daily with house officers for 2 weeks at a time.13 At Northwestern University, a diabetes management service run by nurse practitioners was established, and the focus was on the conversion from intravenous to subcutaneous insulin regimens.14 The RABBIT 2 study that demonstrated the benefits of a basal‐bolus insulin regimen used daily rounding with an endocrinologist.12 More modestly, a program in Pitt County Memorial Hospital in Greenville, NC, relied mostly on diabetes nurse case managers, a strategy which reduced hospital‐wide mean glucose levels as well as LOS, although the greatest improvements in glycemic control were seen in the ICU.19 Our findings are much more consistent with those from University of California San Diego, as yet unpublished, which also used an algorithm, computerized order set, education, as well as continuous quality improvement methods to achieve its aims.22

Our study has several limitations, including being conducted on 1 general medicine service at 1 academic medical center. Moreover, this service, using a physician assistant/hospitalist model, a closed geographic unit, and fairly generous staffing ratio, is likely different from those in many settings and may limit the generalizability of our findings. However, this model allowed us to conduct the study in a laboratory relatively untouched by other cointerventions. Furthermore, the use of PAs in this way may become more common as both academic and community hospitals rely more on mid‐level providers. Our study had a relatively low percentage of patients without a known diagnosis of diabetes compared with other studies, again potentially but not necessarily limiting generalizability. This finding has been shown in other studies at our institution24 and may be due to the high rate of screening for diabetes in the community. Another limitation is that this was a nonrandomized, before‐after trial. However, all subjects were prospectively enrolled to improve comparability, and we performed rigorous adjustment for multiple potential confounding factors. Also, this study had limited statistical power to detect differences in hypoglycemia rates. The preintervention arm was smaller than planned due to fewer diabetic patients than expected on the service and a higher number of exclusions; we prolonged the postintervention period to achieve the desired sample size for that arm of the study.

Our study also has several strengths, including electronic capture of many processes of care and a methodology to operationalize them into measures of protocol adherence. Our metrics of glycemic control were rigorously designed and based on a national task force on inpatient glycemic control sponsored by the Society of Hospital Medicine, with representation from the ADA and AACE.25

Potential future improvements to this intervention include modifications to the daily adjustment algorithm to improve its usability and ability to improve glucose control. Another is the use of high‐reliability methods to improve order set use and daily insulin adjustment, including alerts within the CPOE system and nurse empowerment to contact medical teams if glucose levels are out of range (eg, if greater than 180 mg/dL, not just if greater than 350 or 400 mg/dL). Future research directions include multicenter, randomized controlled trials of these types of interventions and an analysis of more distal patient outcomes including total healthcare utilization, infection rates, end‐organ damage, and mortality.

In conclusion, we found a relationship between a relatively low‐cost quality improvement intervention and improved glycemic control in the non‐ICU general medical setting. Such a finding suggests the benefits of the algorithm itself to improve glucose control and of our implementation strategy. Other institutions may find this intervention a useful starting point for their own quality improvement efforts. Both the algorithm and implementation strategy are deserving of further improvements and future study.

Diabetes mellitus and/or inpatient hyperglycemia are common comorbid conditions in hospitalized patients. Recent surveys show that over 90% of hospitalized diabetic patients experience hyperglycemia (>200 mg/dL), and in nearly 1 in 5 of these patients hyperglycemia persists for 3 days or more.1 Hyperglycemia among inpatients without a previous history of diabetes mellitus is also very common.2 Observational studies have shown that hyperglycemia in hospitalized patients is associated with adverse outcomes including infectious complications, increased length of stay, and increased mortality.27 Recent randomized controlled trials have demonstrated that aggressive treatment of inpatient hyperglycemia improves outcomes in surgical and medical intensive care units.8, 9

Based on the available data, the American Diabetes Association (ADA) now advocates good metabolic control, defined as preprandial glucose levels of 90 to 130 mg/dL and peak postprandial glucose levels <180 mg/dL in hospitalized nonintensive care unit (ICU) patients.10 To reach these targets, the ADA and American College of Endocrinology (ACE) suggest that multidisciplinary teams develop and implement hyperglycemia management guidelines and protocols.11 Protocols should promote the use of continuous intravenous insulin infusions or scheduled basal‐bolus subcutaneous insulin regimens. Subcutaneous insulin protocols should include target glucose levels, basal, nutritional, and supplemental insulin, and daily dose adjustments.6 A recent randomized controlled trial of non‐ICU inpatients demonstrated that such a basal‐bolus insulin regimen results in improved glucose control compared with a sliding scale only regimen.12

To date, few published studies have investigated the best ways to implement such management protocols; those that have are often resource‐intensive, for example involving daily involvement of nurse practitioners or diabetologists.13, 14 It is therefore not known how best to implement an inpatient diabetes management program that is effective, efficient, and self‐perpetuating. At Brigham and Women's Hospital (BWH), we have been refining a subcutaneous insulin protocol, focused provider education, and more recently a computerized order set to overcome barriers related to fear of hypoglycemia, delays in insulin prescribing, and unfamiliarity with inpatient glucose management.15 The aims of this current trial were to evaluate the effects of these interventions on a geographically localized general medical service previously naive to these interventions to evaluate their effects on glycemic control, patient safety, and processes of care. We hypothesized that these interventions would improve glycemic control and increase use of basal‐bolus insulin orders without increasing the rate of hypoglycemia.

METHODS

RESULTS

We prospectively identified 248 potential patients for the study. We subsequently excluded 79 patients for the following reasons: no glucose readings beyond hospital day 1 while on PACE service (34 patients); never admitted to PACE service (15 patients); no diabetes or inpatient hyperglycemia (9 patients, mostly patients prescribed an insulin sliding scale prophylactically to avoid steroid‐induced hyperglycemia); type 1 diabetes (13 patients); TPN, DKA, or HHS (5 patients); and palliative care (3 patients). The remaining 169 patients included 63 from the preintervention period(out of 489 total admissions to the PACE service; 13%) and 106 patients in the postintervention period (out of 565 admissions; 19%). These patients had 2447 glucose readings, or an average of 3.6 glucose readings per monitored patient‐day in the preintervention period and 3.3 glucose readings per patient‐day in the postintervention period. Even including the 34 patients who were excluded for lack of glucose readings, glucose data were still available for 717 out of a potential 775 patient‐days (93%). Characteristics for all included patients are shown in Table 2. The mean admission glucose was 197 mg/dL, mean A1C was 8.4%, 54% of the patients were prescribed insulin prior to admission, and 7% had no prior diagnosis of diabetes. There were no significant differences in baseline characteristics between the 2 patient groups except for Charlson score, which was higher in the preintervention group (87% versus 74% with score 2 or higher; Table 2). The top diagnosis‐related groups for the entire cohort included: heart failure and shock (12 patients); kidney and urinary tract infections (12 patients); esophagitis, gastroenteritis, and miscellaneous digestive disorders (11 patients); chronic obstructive pulmonary disease (10 patients); renal failure (10 patients); simple pneumonia and pleurisy (7 patients); disorders of the pancreas except malignancy (6 patients); chest pain (5 patients); and cellulitis (5 patients).

With respect to insulin ordering practices, there was no significant difference in the use of basal insulin in hyperglycemic patients between the preintervention period and postintervention period (81% versus 91%; P = 0.17), nor in the dose of basal insulin prescribed (results not shown), but there was an increase in the use of scheduled nutritional insulin for those patients with hyperglycemia receiving nutrition: 40% versus 75%, P < 0.001 (Table 3). The percent of patients receiving supplemental (sliding scale) insulin by itself (ie, without ever receiving basal or nutritional insulin) was lower during the postintervention period (29% versus 8%, P < 0.001). Nonsignificant differences were seen in the rates of prescribing an appropriate dose and type of nutritional insulin. Notably, there was no difference at all in the proportion of patient‐days in which insulin adjustments were made when 2 or more episodes of hyperglycemia or hypoglycemia were present during the previous day (56% of patient‐days in both groups; P = 0.90).

The primary outcome, the mean percent of glucose readings between 60 and 180 mg/dL per patient, was 59.1% in the preintervention period and 64.7% in the postintervention (P = 0.13 in unadjusted analysis; Table 3). When adjusted for A1C, admission glucose, and insulin use prior to admission, the adjusted absolute difference in the percent of glucose readings within range was 9.7% (95% confidence interval [CI], 0.6%‐18.8%; P = 0.04; Table 3). Regarding other measures of glucose control, the patient‐day weighted mean glucose was 174.7 mg/dL in the preintervention period and 164.6 mg/dL postintervention (P = 0.02), and there was no significant difference in the percent of patient‐days with any hypoglycemia (glucose <60 mg/dL) or severe hypoglycemia (glucose <40 mg/dL; Table 3). There were also no significant differences in the mean number of hypoglycemic events per patient‐day (6.8 versus 6.6 per 100 patient‐days; relative risk, 0.95; 95% CI, 0.541.67; P = 0.87) or severe hypoglycemic events per patient‐day (1.0 versus 1.4 per 100 patient‐days; relative risk, 1.38; 95% CI, 0.355.53; P = 0.65).

We also compared hospital length of stay in hours between the study groups (Table 3). Length of stay (LOS) was shorter in the postintervention arm in unadjusted analyses (112 versus 86 hours; P < 0.001), and this difference persisted when adjusted for patient insurance, race, gender, and Charlson comorbidity score (25% shorter; 95% CI, 6%‐44%). A comparison of LOS among nonstudy patients on the PACE service during these 2 time periods revealed no difference (105 versus 101 hours). When the length of stay analysis was limited to study patients with a known diagnosis of diabetes, the adjusted effect size was a 31% relative decrease in length of stay.

Figure 1A shows the percent glucose readings within range per patient by hospital day. The greatest differences between groups can be seen on hospital days 2 and 3 (11% absolute differences on both days). Similarly, Figure 1B shows the mean glucose per patient by hospital day. Again, the biggest differences are seen on hospital days 2 and 3 (20 and 23 mg/dL difference between groups, respectively). In both cases, only the day 3 comparisons were significantly different between study groups.

Figure 1

DISCUSSION

In this before‐after study, we found that a multifaceted intervention consisting of a subcutaneous insulin protocol, focused education, and an order set built into the hospital's CPOE system was associated with a significantly higher percentage of glucose readings within range per patient in analyses adjusted for patient demographics and severity of diabetes. We also found a significant decrease in patient‐day weighted mean glucose, a marked increase in appropriate use of scheduled nutritional insulin, and a concomitant decrease in sliding scale insulin only regimens during the postintervention period. Moreover, we found a shorter length of stay during the postintervention period that persisted after adjustment for several clinical factors. Importantly, the interventions described in this study require very few resources to continue indefinitely: printing costs for the management protocol, 4 hours of education delivered per year, and routine upkeep of an electronic order set.

Because this was a before‐after study, we cannot exclude the possibility that these improvements in process and outcome were due to cointerventions and/or temporal trends. However, we know of no other interventions aimed at improving diabetes care in this self‐contained service of nurses, PAs, and hospitalists. Moreover, the process improvements, especially the increase in scheduled nutritional insulin, were rather marked, unlikely to be due to temporal trends alone, and likely capable of producing the corresponding improvements in glucose control. That glucose control stopped improving after hospital day 3 may be due to the fact that subsequent adjustment to insulin orders occurred infrequently and no more often than prior to the intervention. That we did not see greater improvements in glycemic control overall may also reflect the fact that 81% of study patients with inpatient hyperglycemia received basal insulin prior to the intervention.

The reduction in patient LOS was somewhat surprising given the relatively small sample size. However, the results are consistent with those of other studies linking hyperglycemia to LOS18, 19 and we found no evidence for a temporal trend toward lower LOS on the PACE service as a whole during the same time period. While a greater proportion of patients on the PACE service were in the study in the post‐intervention period compared with the preintervention period, we found no evidence that the difference in length of stay was due to increased surveillance for nondiabetics, especially because eligibility criteria depended on phlebotomy glucose values, which were uniformly tested in all inpatients. Also, effects on length of stay were actually stronger when limited to patients with known diabetes. Finally, we controlled for several predictors of length of stay, although we still cannot exclude the possibility of unmeasured confounding between groups.

Since ADA and ACE issued guidelines for inpatient management of diabetes and hyperglycemia, many institutions have developed subcutaneous insulin algorithms, educational curricula, and/or order sets to increase compliance with these guidelines and improve glycemic control. Some of these efforts have been studied and some have been successful in their efforts.13, 14, 2023 Unfortunately, most of these programs have not rigorously assessed their impact on process and outcomes, and the most effective studies published to date have involved interventions much more intensive than those described here. For example, Rush University's intervention was associated with a 50 mg/dL decrease in mean blood glucose but involved an endocrinologist rounding twice daily with house officers for 2 weeks at a time.13 At Northwestern University, a diabetes management service run by nurse practitioners was established, and the focus was on the conversion from intravenous to subcutaneous insulin regimens.14 The RABBIT 2 study that demonstrated the benefits of a basal‐bolus insulin regimen used daily rounding with an endocrinologist.12 More modestly, a program in Pitt County Memorial Hospital in Greenville, NC, relied mostly on diabetes nurse case managers, a strategy which reduced hospital‐wide mean glucose levels as well as LOS, although the greatest improvements in glycemic control were seen in the ICU.19 Our findings are much more consistent with those from University of California San Diego, as yet unpublished, which also used an algorithm, computerized order set, education, as well as continuous quality improvement methods to achieve its aims.22

Our study has several limitations, including being conducted on 1 general medicine service at 1 academic medical center. Moreover, this service, using a physician assistant/hospitalist model, a closed geographic unit, and fairly generous staffing ratio, is likely different from those in many settings and may limit the generalizability of our findings. However, this model allowed us to conduct the study in a laboratory relatively untouched by other cointerventions. Furthermore, the use of PAs in this way may become more common as both academic and community hospitals rely more on mid‐level providers. Our study had a relatively low percentage of patients without a known diagnosis of diabetes compared with other studies, again potentially but not necessarily limiting generalizability. This finding has been shown in other studies at our institution24 and may be due to the high rate of screening for diabetes in the community. Another limitation is that this was a nonrandomized, before‐after trial. However, all subjects were prospectively enrolled to improve comparability, and we performed rigorous adjustment for multiple potential confounding factors. Also, this study had limited statistical power to detect differences in hypoglycemia rates. The preintervention arm was smaller than planned due to fewer diabetic patients than expected on the service and a higher number of exclusions; we prolonged the postintervention period to achieve the desired sample size for that arm of the study.

Our study also has several strengths, including electronic capture of many processes of care and a methodology to operationalize them into measures of protocol adherence. Our metrics of glycemic control were rigorously designed and based on a national task force on inpatient glycemic control sponsored by the Society of Hospital Medicine, with representation from the ADA and AACE.25

Potential future improvements to this intervention include modifications to the daily adjustment algorithm to improve its usability and ability to improve glucose control. Another is the use of high‐reliability methods to improve order set use and daily insulin adjustment, including alerts within the CPOE system and nurse empowerment to contact medical teams if glucose levels are out of range (eg, if greater than 180 mg/dL, not just if greater than 350 or 400 mg/dL). Future research directions include multicenter, randomized controlled trials of these types of interventions and an analysis of more distal patient outcomes including total healthcare utilization, infection rates, end‐organ damage, and mortality.

In conclusion, we found a relationship between a relatively low‐cost quality improvement intervention and improved glycemic control in the non‐ICU general medical setting. Such a finding suggests the benefits of the algorithm itself to improve glucose control and of our implementation strategy. Other institutions may find this intervention a useful starting point for their own quality improvement efforts. Both the algorithm and implementation strategy are deserving of further improvements and future study.

Diabetes mellitus and/or inpatient hyperglycemia are common comorbid conditions in hospitalized patients. Recent surveys show that over 90% of hospitalized diabetic patients experience hyperglycemia (>200 mg/dL), and in nearly 1 in 5 of these patients hyperglycemia persists for 3 days or more.1 Hyperglycemia among inpatients without a previous history of diabetes mellitus is also very common.2 Observational studies have shown that hyperglycemia in hospitalized patients is associated with adverse outcomes including infectious complications, increased length of stay, and increased mortality.27 Recent randomized controlled trials have demonstrated that aggressive treatment of inpatient hyperglycemia improves outcomes in surgical and medical intensive care units.8, 9

Based on the available data, the American Diabetes Association (ADA) now advocates good metabolic control, defined as preprandial glucose levels of 90 to 130 mg/dL and peak postprandial glucose levels <180 mg/dL in hospitalized nonintensive care unit (ICU) patients.10 To reach these targets, the ADA and American College of Endocrinology (ACE) suggest that multidisciplinary teams develop and implement hyperglycemia management guidelines and protocols.11 Protocols should promote the use of continuous intravenous insulin infusions or scheduled basal‐bolus subcutaneous insulin regimens. Subcutaneous insulin protocols should include target glucose levels, basal, nutritional, and supplemental insulin, and daily dose adjustments.6 A recent randomized controlled trial of non‐ICU inpatients demonstrated that such a basal‐bolus insulin regimen results in improved glucose control compared with a sliding scale only regimen.12

To date, few published studies have investigated the best ways to implement such management protocols; those that have are often resource‐intensive, for example involving daily involvement of nurse practitioners or diabetologists.13, 14 It is therefore not known how best to implement an inpatient diabetes management program that is effective, efficient, and self‐perpetuating. At Brigham and Women's Hospital (BWH), we have been refining a subcutaneous insulin protocol, focused provider education, and more recently a computerized order set to overcome barriers related to fear of hypoglycemia, delays in insulin prescribing, and unfamiliarity with inpatient glucose management.15 The aims of this current trial were to evaluate the effects of these interventions on a geographically localized general medical service previously naive to these interventions to evaluate their effects on glycemic control, patient safety, and processes of care. We hypothesized that these interventions would improve glycemic control and increase use of basal‐bolus insulin orders without increasing the rate of hypoglycemia.

METHODS

RESULTS

We prospectively identified 248 potential patients for the study. We subsequently excluded 79 patients for the following reasons: no glucose readings beyond hospital day 1 while on PACE service (34 patients); never admitted to PACE service (15 patients); no diabetes or inpatient hyperglycemia (9 patients, mostly patients prescribed an insulin sliding scale prophylactically to avoid steroid‐induced hyperglycemia); type 1 diabetes (13 patients); TPN, DKA, or HHS (5 patients); and palliative care (3 patients). The remaining 169 patients included 63 from the preintervention period(out of 489 total admissions to the PACE service; 13%) and 106 patients in the postintervention period (out of 565 admissions; 19%). These patients had 2447 glucose readings, or an average of 3.6 glucose readings per monitored patient‐day in the preintervention period and 3.3 glucose readings per patient‐day in the postintervention period. Even including the 34 patients who were excluded for lack of glucose readings, glucose data were still available for 717 out of a potential 775 patient‐days (93%). Characteristics for all included patients are shown in Table 2. The mean admission glucose was 197 mg/dL, mean A1C was 8.4%, 54% of the patients were prescribed insulin prior to admission, and 7% had no prior diagnosis of diabetes. There were no significant differences in baseline characteristics between the 2 patient groups except for Charlson score, which was higher in the preintervention group (87% versus 74% with score 2 or higher; Table 2). The top diagnosis‐related groups for the entire cohort included: heart failure and shock (12 patients); kidney and urinary tract infections (12 patients); esophagitis, gastroenteritis, and miscellaneous digestive disorders (11 patients); chronic obstructive pulmonary disease (10 patients); renal failure (10 patients); simple pneumonia and pleurisy (7 patients); disorders of the pancreas except malignancy (6 patients); chest pain (5 patients); and cellulitis (5 patients).

With respect to insulin ordering practices, there was no significant difference in the use of basal insulin in hyperglycemic patients between the preintervention period and postintervention period (81% versus 91%; P = 0.17), nor in the dose of basal insulin prescribed (results not shown), but there was an increase in the use of scheduled nutritional insulin for those patients with hyperglycemia receiving nutrition: 40% versus 75%, P < 0.001 (Table 3). The percent of patients receiving supplemental (sliding scale) insulin by itself (ie, without ever receiving basal or nutritional insulin) was lower during the postintervention period (29% versus 8%, P < 0.001). Nonsignificant differences were seen in the rates of prescribing an appropriate dose and type of nutritional insulin. Notably, there was no difference at all in the proportion of patient‐days in which insulin adjustments were made when 2 or more episodes of hyperglycemia or hypoglycemia were present during the previous day (56% of patient‐days in both groups; P = 0.90).

The primary outcome, the mean percent of glucose readings between 60 and 180 mg/dL per patient, was 59.1% in the preintervention period and 64.7% in the postintervention (P = 0.13 in unadjusted analysis; Table 3). When adjusted for A1C, admission glucose, and insulin use prior to admission, the adjusted absolute difference in the percent of glucose readings within range was 9.7% (95% confidence interval [CI], 0.6%‐18.8%; P = 0.04; Table 3). Regarding other measures of glucose control, the patient‐day weighted mean glucose was 174.7 mg/dL in the preintervention period and 164.6 mg/dL postintervention (P = 0.02), and there was no significant difference in the percent of patient‐days with any hypoglycemia (glucose <60 mg/dL) or severe hypoglycemia (glucose <40 mg/dL; Table 3). There were also no significant differences in the mean number of hypoglycemic events per patient‐day (6.8 versus 6.6 per 100 patient‐days; relative risk, 0.95; 95% CI, 0.541.67; P = 0.87) or severe hypoglycemic events per patient‐day (1.0 versus 1.4 per 100 patient‐days; relative risk, 1.38; 95% CI, 0.355.53; P = 0.65).

We also compared hospital length of stay in hours between the study groups (Table 3). Length of stay (LOS) was shorter in the postintervention arm in unadjusted analyses (112 versus 86 hours; P < 0.001), and this difference persisted when adjusted for patient insurance, race, gender, and Charlson comorbidity score (25% shorter; 95% CI, 6%‐44%). A comparison of LOS among nonstudy patients on the PACE service during these 2 time periods revealed no difference (105 versus 101 hours). When the length of stay analysis was limited to study patients with a known diagnosis of diabetes, the adjusted effect size was a 31% relative decrease in length of stay.

Figure 1A shows the percent glucose readings within range per patient by hospital day. The greatest differences between groups can be seen on hospital days 2 and 3 (11% absolute differences on both days). Similarly, Figure 1B shows the mean glucose per patient by hospital day. Again, the biggest differences are seen on hospital days 2 and 3 (20 and 23 mg/dL difference between groups, respectively). In both cases, only the day 3 comparisons were significantly different between study groups.

Figure 1

DISCUSSION

In this before‐after study, we found that a multifaceted intervention consisting of a subcutaneous insulin protocol, focused education, and an order set built into the hospital's CPOE system was associated with a significantly higher percentage of glucose readings within range per patient in analyses adjusted for patient demographics and severity of diabetes. We also found a significant decrease in patient‐day weighted mean glucose, a marked increase in appropriate use of scheduled nutritional insulin, and a concomitant decrease in sliding scale insulin only regimens during the postintervention period. Moreover, we found a shorter length of stay during the postintervention period that persisted after adjustment for several clinical factors. Importantly, the interventions described in this study require very few resources to continue indefinitely: printing costs for the management protocol, 4 hours of education delivered per year, and routine upkeep of an electronic order set.

Because this was a before‐after study, we cannot exclude the possibility that these improvements in process and outcome were due to cointerventions and/or temporal trends. However, we know of no other interventions aimed at improving diabetes care in this self‐contained service of nurses, PAs, and hospitalists. Moreover, the process improvements, especially the increase in scheduled nutritional insulin, were rather marked, unlikely to be due to temporal trends alone, and likely capable of producing the corresponding improvements in glucose control. That glucose control stopped improving after hospital day 3 may be due to the fact that subsequent adjustment to insulin orders occurred infrequently and no more often than prior to the intervention. That we did not see greater improvements in glycemic control overall may also reflect the fact that 81% of study patients with inpatient hyperglycemia received basal insulin prior to the intervention.

The reduction in patient LOS was somewhat surprising given the relatively small sample size. However, the results are consistent with those of other studies linking hyperglycemia to LOS18, 19 and we found no evidence for a temporal trend toward lower LOS on the PACE service as a whole during the same time period. While a greater proportion of patients on the PACE service were in the study in the post‐intervention period compared with the preintervention period, we found no evidence that the difference in length of stay was due to increased surveillance for nondiabetics, especially because eligibility criteria depended on phlebotomy glucose values, which were uniformly tested in all inpatients. Also, effects on length of stay were actually stronger when limited to patients with known diabetes. Finally, we controlled for several predictors of length of stay, although we still cannot exclude the possibility of unmeasured confounding between groups.

Since ADA and ACE issued guidelines for inpatient management of diabetes and hyperglycemia, many institutions have developed subcutaneous insulin algorithms, educational curricula, and/or order sets to increase compliance with these guidelines and improve glycemic control. Some of these efforts have been studied and some have been successful in their efforts.13, 14, 2023 Unfortunately, most of these programs have not rigorously assessed their impact on process and outcomes, and the most effective studies published to date have involved interventions much more intensive than those described here. For example, Rush University's intervention was associated with a 50 mg/dL decrease in mean blood glucose but involved an endocrinologist rounding twice daily with house officers for 2 weeks at a time.13 At Northwestern University, a diabetes management service run by nurse practitioners was established, and the focus was on the conversion from intravenous to subcutaneous insulin regimens.14 The RABBIT 2 study that demonstrated the benefits of a basal‐bolus insulin regimen used daily rounding with an endocrinologist.12 More modestly, a program in Pitt County Memorial Hospital in Greenville, NC, relied mostly on diabetes nurse case managers, a strategy which reduced hospital‐wide mean glucose levels as well as LOS, although the greatest improvements in glycemic control were seen in the ICU.19 Our findings are much more consistent with those from University of California San Diego, as yet unpublished, which also used an algorithm, computerized order set, education, as well as continuous quality improvement methods to achieve its aims.22

Our study has several limitations, including being conducted on 1 general medicine service at 1 academic medical center. Moreover, this service, using a physician assistant/hospitalist model, a closed geographic unit, and fairly generous staffing ratio, is likely different from those in many settings and may limit the generalizability of our findings. However, this model allowed us to conduct the study in a laboratory relatively untouched by other cointerventions. Furthermore, the use of PAs in this way may become more common as both academic and community hospitals rely more on mid‐level providers. Our study had a relatively low percentage of patients without a known diagnosis of diabetes compared with other studies, again potentially but not necessarily limiting generalizability. This finding has been shown in other studies at our institution24 and may be due to the high rate of screening for diabetes in the community. Another limitation is that this was a nonrandomized, before‐after trial. However, all subjects were prospectively enrolled to improve comparability, and we performed rigorous adjustment for multiple potential confounding factors. Also, this study had limited statistical power to detect differences in hypoglycemia rates. The preintervention arm was smaller than planned due to fewer diabetic patients than expected on the service and a higher number of exclusions; we prolonged the postintervention period to achieve the desired sample size for that arm of the study.

Our study also has several strengths, including electronic capture of many processes of care and a methodology to operationalize them into measures of protocol adherence. Our metrics of glycemic control were rigorously designed and based on a national task force on inpatient glycemic control sponsored by the Society of Hospital Medicine, with representation from the ADA and AACE.25

Potential future improvements to this intervention include modifications to the daily adjustment algorithm to improve its usability and ability to improve glucose control. Another is the use of high‐reliability methods to improve order set use and daily insulin adjustment, including alerts within the CPOE system and nurse empowerment to contact medical teams if glucose levels are out of range (eg, if greater than 180 mg/dL, not just if greater than 350 or 400 mg/dL). Future research directions include multicenter, randomized controlled trials of these types of interventions and an analysis of more distal patient outcomes including total healthcare utilization, infection rates, end‐organ damage, and mortality.

In conclusion, we found a relationship between a relatively low‐cost quality improvement intervention and improved glycemic control in the non‐ICU general medical setting. Such a finding suggests the benefits of the algorithm itself to improve glucose control and of our implementation strategy. Other institutions may find this intervention a useful starting point for their own quality improvement efforts. Both the algorithm and implementation strategy are deserving of further improvements and future study.

Midlevel providers (physician assistants and nurse practitioners) have long been employed by academic medical centers, predominantly on surgical services, or on medical subspecialty services, where they have typically had a limited scope of practice, focused in a narrowly defined area or set of procedures.17 In contrast, there are relatively few reports of experiences deploying midlevel providers to replace house staff on inpatient general medicine services in academic centers,810 and few studies of the effect of midlevel providers on quality and efficiency of care in the academic setting. Despite this, reductions in house officer duty hours as mandated by the Accreditation Council on Graduate Medical Education (ACGME)11 have resulted in academic centers increasingly using midlevel providers to decrease house staff workload on inpatient services.12, 13 In general, midlevel practitioners on general medicine services have been deployed to: (1) care for a population of patients separate from and in parallel with house staff; this population may be narrowly defined (eg, patients with chest pain) or not; (2) assist with the management of patients cared for by house staff by performing certain tasks (eg, scheduling appointments, discharging patients). Even as midlevel providers become more prevalent on academic general medicine services, the best model of care incorporating them into clinical care remains unclear, and few studies have rigorously examined the care provided on services that use them.

We developed an inpatient general medicine service within a large academic medical center staffed by physician assistants and hospitalists to help our residency program meet ACGME duty hour requirements. We hypothesized that by creating a service that is geographically localized and supervised by full‐time hospitalists, by instituting multidisciplinary rounds, and by investing in the professional development of highly‐skilled physician assistants, we could provide care for medically complex, acutely ill general medicine inpatients with similar quality and efficiency as compared to house staff teams. We report our experience during the first year of implementing the service, and compare quality and efficiency of care on this service with that of our traditional house staff services. We also evaluate the effects of this service on patient satisfaction and self‐reported house staff workload.

PATIENTS AND METHODS

RESULTS

Patient Satisfaction

Patients were similarly satisfied with their care on the PACE service and on the house staff services. In specific areas and globally, percentages of patients satisfied with their physicians and with the discharge process were not different, as measured by the Press‐Ganey survey (Press‐Ganey Associates, South Bend, IN; Figures 1 and 2). The survey distinguishes between attendings and residents, but not physician assistants; therefore, Figure 1 only includes responses to the attending questions. Given the sampling procedure of the Press‐Ganey survey, exact response rates cannot be calculated, but Press‐Ganey reports a response rate of about 40% for the English survey and about 20% for the Spanish survey.

Figure 1

Figure 2

DISCUSSION

As academic centers struggle with developing a workforce to provide patient care no longer provided by residents, questions about the ideal structure of nonhouse staff inpatient services abound. Although solutions to this problem will be determined to some extent by local factors such as institutional culture and resources, some lessons learned in developing such services will be more widely applicable. We found that by implementing a geographically localized, physician assistant‐staffed hospitalist service, we were able to provide care of similar quality and efficiency to that of traditional house staff services, despite inexperienced hospitalists staffing the service and a medical residency program commonly recognized as one of the best in the country. Adjusted total costs were slightly lower on the PACE service, but this difference was small and of borderline statistical significance. Likewise, no significant differences were seen in any of several quality measures or in patient satisfaction.

Our findings add to the available evidence supporting the use of physician assistants on academic general medicine services, and are germane to academic centers facing reductions in house staff availability and seeking alternative models of care for inpatients. Several specific characteristics of the PACE service and the implications of these should be considered:

• 1

  The service accepted all patients, regardless of diagnosis, acuity, or complexity of illness. This was unlike many previously described nonhouse staff services which were more limited in scope, and allowed more flexibility with patient flow. However, in the end, patients on the PACE service did have a modestly lower case mix index and Charlson score, suggesting that, despite a lack of triage guidelines, there was some bias in the triage of admissions, possibly due to a perception that physician assistants should take care of lower complexity patients. If it is desirable to have a similar distribution of higher complexity patients across house staff and nonhouse staff services, extra efforts may be necessary to overcome this perception.

• 2

  The service was geographically regionalized. Geographic regionalization offered many important advantages, especially with regards to communication among staff, nursing, and consultants, and allowed for multidisciplinary rounds. However, it is possible that the modest, but not statistically significant, trend toward an increased LOS seen on the PACE service might be a reflection of geographic admitting (less incentive to discharge since discharging a patient means taking a new admission).

• 3

  The education and professional development of the physician assistants was a priority. Physician assistants had considerable autonomy and responsibility, and rather than being assigned only lower level administrative tasks, performed all aspects of patient care. They also received regular teaching from the hospitalists, attended house staff teaching conferences, and developed nonclinical roles in education and quality improvement. The higher standards expected of the physician assistants were quite possibly a factor in the quality of care delivered, and almost certainly contributed to physician assistant satisfaction and retention.

Our findings contrast with those of Myers et al.,9 who found that a nonteaching service staffed by hospitalists and nurse practitioners had a significantly lower median LOS and hospital charges compared to similar patients on resident‐based services. However, unlike ours, their service cared for a select patient population, and only accepted patients with chest pain at low risk for acute coronary syndrome. Van Rhee et al.10 found that physician assistants on a general medicine service used fewer resources for patients with pneumonia, stroke, and congestive heart failure than resident physicians, and did not exceed the resources used by residents in other diagnoses. The authors did not find a difference in LOS, but did find a significantly higher mortality among patients with pneumonia cared for by physician assistants.

Several limitations should be noted. First, the study was a retrospective analysis of administrative data rather than a randomized trial, and although we employed a standard approach to adjust for a wide range of patient characteristics including severity of illness, there may have been undetected differences in the patient populations studied that may have confounded our results. Second, resident moonlighters admitted patients to the PACE service and, at other times, to the house staff services, and this may have diluted any differences between the groups. However, when we limited our analysis to the subgroup of patients admitted during the day, similar results were obtained, with the exception that the PACE service had a lower rate of 14‐day readmissions, an unexpected finding deserving of further study. Third, the study was conducted in a single academic institution and our findings may not be generalizable to others with different needs and resources; indeed, the costs associated with implementing such a service may be prohibitive for some institutions. Fourth, because of simultaneous changes that were taking place in our residency program, we are unable to accurately assess the impact of the PACE service on resident duty hours. However, resident duty hours did not increase over this time period on the general medicine service, and implementation of the service allowed for redistribution of house staff to other services and positions. Fifth, patient satisfaction data were obtained from responses to the mailed Press‐Ganey survey, to which there is a relatively low response rate. Also, we did not survey providers regarding their satisfaction with the service during the study period. Sixth, the study had limited power to detect clinically important differences in mortality and ICU transfers. Finally, this study is unable to compare this particular model of incorporating midlevel providers into general medical services with other models, only with traditional house staff services.

Future research should focus on determining the most effective and efficient ways to incorporate midlevel providers on academic general medicine services. One important question from the standpoint of house staff training is whether such services should be separate but equal, or should house staff gain experience during residency working with midlevel providers, since they are likely to encounter them in the future whether they stay in academics or not. Different models of care will likely have large implications for the quality and efficiency of patient care, house staff education and satisfaction, and physician assistant job satisfaction and turnover.

In summary, our study demonstrates that a geographically regionalized, multidisciplinary service staffed by hospitalists and physician assistants can be a safe alternative to house staff‐based services for the care of general medicine inpatients in an academic medical center.

Midlevel providers (physician assistants and nurse practitioners) have long been employed by academic medical centers, predominantly on surgical services, or on medical subspecialty services, where they have typically had a limited scope of practice, focused in a narrowly defined area or set of procedures.17 In contrast, there are relatively few reports of experiences deploying midlevel providers to replace house staff on inpatient general medicine services in academic centers,810 and few studies of the effect of midlevel providers on quality and efficiency of care in the academic setting. Despite this, reductions in house officer duty hours as mandated by the Accreditation Council on Graduate Medical Education (ACGME)11 have resulted in academic centers increasingly using midlevel providers to decrease house staff workload on inpatient services.12, 13 In general, midlevel practitioners on general medicine services have been deployed to: (1) care for a population of patients separate from and in parallel with house staff; this population may be narrowly defined (eg, patients with chest pain) or not; (2) assist with the management of patients cared for by house staff by performing certain tasks (eg, scheduling appointments, discharging patients). Even as midlevel providers become more prevalent on academic general medicine services, the best model of care incorporating them into clinical care remains unclear, and few studies have rigorously examined the care provided on services that use them.

We developed an inpatient general medicine service within a large academic medical center staffed by physician assistants and hospitalists to help our residency program meet ACGME duty hour requirements. We hypothesized that by creating a service that is geographically localized and supervised by full‐time hospitalists, by instituting multidisciplinary rounds, and by investing in the professional development of highly‐skilled physician assistants, we could provide care for medically complex, acutely ill general medicine inpatients with similar quality and efficiency as compared to house staff teams. We report our experience during the first year of implementing the service, and compare quality and efficiency of care on this service with that of our traditional house staff services. We also evaluate the effects of this service on patient satisfaction and self‐reported house staff workload.

PATIENTS AND METHODS

RESULTS

Patient Satisfaction

Patients were similarly satisfied with their care on the PACE service and on the house staff services. In specific areas and globally, percentages of patients satisfied with their physicians and with the discharge process were not different, as measured by the Press‐Ganey survey (Press‐Ganey Associates, South Bend, IN; Figures 1 and 2). The survey distinguishes between attendings and residents, but not physician assistants; therefore, Figure 1 only includes responses to the attending questions. Given the sampling procedure of the Press‐Ganey survey, exact response rates cannot be calculated, but Press‐Ganey reports a response rate of about 40% for the English survey and about 20% for the Spanish survey.

Figure 1

Figure 2

DISCUSSION

As academic centers struggle with developing a workforce to provide patient care no longer provided by residents, questions about the ideal structure of nonhouse staff inpatient services abound. Although solutions to this problem will be determined to some extent by local factors such as institutional culture and resources, some lessons learned in developing such services will be more widely applicable. We found that by implementing a geographically localized, physician assistant‐staffed hospitalist service, we were able to provide care of similar quality and efficiency to that of traditional house staff services, despite inexperienced hospitalists staffing the service and a medical residency program commonly recognized as one of the best in the country. Adjusted total costs were slightly lower on the PACE service, but this difference was small and of borderline statistical significance. Likewise, no significant differences were seen in any of several quality measures or in patient satisfaction.

Our findings add to the available evidence supporting the use of physician assistants on academic general medicine services, and are germane to academic centers facing reductions in house staff availability and seeking alternative models of care for inpatients. Several specific characteristics of the PACE service and the implications of these should be considered:

• 1

  The service accepted all patients, regardless of diagnosis, acuity, or complexity of illness. This was unlike many previously described nonhouse staff services which were more limited in scope, and allowed more flexibility with patient flow. However, in the end, patients on the PACE service did have a modestly lower case mix index and Charlson score, suggesting that, despite a lack of triage guidelines, there was some bias in the triage of admissions, possibly due to a perception that physician assistants should take care of lower complexity patients. If it is desirable to have a similar distribution of higher complexity patients across house staff and nonhouse staff services, extra efforts may be necessary to overcome this perception.

• 2

  The service was geographically regionalized. Geographic regionalization offered many important advantages, especially with regards to communication among staff, nursing, and consultants, and allowed for multidisciplinary rounds. However, it is possible that the modest, but not statistically significant, trend toward an increased LOS seen on the PACE service might be a reflection of geographic admitting (less incentive to discharge since discharging a patient means taking a new admission).

• 3

  The education and professional development of the physician assistants was a priority. Physician assistants had considerable autonomy and responsibility, and rather than being assigned only lower level administrative tasks, performed all aspects of patient care. They also received regular teaching from the hospitalists, attended house staff teaching conferences, and developed nonclinical roles in education and quality improvement. The higher standards expected of the physician assistants were quite possibly a factor in the quality of care delivered, and almost certainly contributed to physician assistant satisfaction and retention.

Our findings contrast with those of Myers et al.,9 who found that a nonteaching service staffed by hospitalists and nurse practitioners had a significantly lower median LOS and hospital charges compared to similar patients on resident‐based services. However, unlike ours, their service cared for a select patient population, and only accepted patients with chest pain at low risk for acute coronary syndrome. Van Rhee et al.10 found that physician assistants on a general medicine service used fewer resources for patients with pneumonia, stroke, and congestive heart failure than resident physicians, and did not exceed the resources used by residents in other diagnoses. The authors did not find a difference in LOS, but did find a significantly higher mortality among patients with pneumonia cared for by physician assistants.

Several limitations should be noted. First, the study was a retrospective analysis of administrative data rather than a randomized trial, and although we employed a standard approach to adjust for a wide range of patient characteristics including severity of illness, there may have been undetected differences in the patient populations studied that may have confounded our results. Second, resident moonlighters admitted patients to the PACE service and, at other times, to the house staff services, and this may have diluted any differences between the groups. However, when we limited our analysis to the subgroup of patients admitted during the day, similar results were obtained, with the exception that the PACE service had a lower rate of 14‐day readmissions, an unexpected finding deserving of further study. Third, the study was conducted in a single academic institution and our findings may not be generalizable to others with different needs and resources; indeed, the costs associated with implementing such a service may be prohibitive for some institutions. Fourth, because of simultaneous changes that were taking place in our residency program, we are unable to accurately assess the impact of the PACE service on resident duty hours. However, resident duty hours did not increase over this time period on the general medicine service, and implementation of the service allowed for redistribution of house staff to other services and positions. Fifth, patient satisfaction data were obtained from responses to the mailed Press‐Ganey survey, to which there is a relatively low response rate. Also, we did not survey providers regarding their satisfaction with the service during the study period. Sixth, the study had limited power to detect clinically important differences in mortality and ICU transfers. Finally, this study is unable to compare this particular model of incorporating midlevel providers into general medical services with other models, only with traditional house staff services.

Future research should focus on determining the most effective and efficient ways to incorporate midlevel providers on academic general medicine services. One important question from the standpoint of house staff training is whether such services should be separate but equal, or should house staff gain experience during residency working with midlevel providers, since they are likely to encounter them in the future whether they stay in academics or not. Different models of care will likely have large implications for the quality and efficiency of patient care, house staff education and satisfaction, and physician assistant job satisfaction and turnover.

In summary, our study demonstrates that a geographically regionalized, multidisciplinary service staffed by hospitalists and physician assistants can be a safe alternative to house staff‐based services for the care of general medicine inpatients in an academic medical center.

Midlevel providers (physician assistants and nurse practitioners) have long been employed by academic medical centers, predominantly on surgical services, or on medical subspecialty services, where they have typically had a limited scope of practice, focused in a narrowly defined area or set of procedures.17 In contrast, there are relatively few reports of experiences deploying midlevel providers to replace house staff on inpatient general medicine services in academic centers,810 and few studies of the effect of midlevel providers on quality and efficiency of care in the academic setting. Despite this, reductions in house officer duty hours as mandated by the Accreditation Council on Graduate Medical Education (ACGME)11 have resulted in academic centers increasingly using midlevel providers to decrease house staff workload on inpatient services.12, 13 In general, midlevel practitioners on general medicine services have been deployed to: (1) care for a population of patients separate from and in parallel with house staff; this population may be narrowly defined (eg, patients with chest pain) or not; (2) assist with the management of patients cared for by house staff by performing certain tasks (eg, scheduling appointments, discharging patients). Even as midlevel providers become more prevalent on academic general medicine services, the best model of care incorporating them into clinical care remains unclear, and few studies have rigorously examined the care provided on services that use them.

We developed an inpatient general medicine service within a large academic medical center staffed by physician assistants and hospitalists to help our residency program meet ACGME duty hour requirements. We hypothesized that by creating a service that is geographically localized and supervised by full‐time hospitalists, by instituting multidisciplinary rounds, and by investing in the professional development of highly‐skilled physician assistants, we could provide care for medically complex, acutely ill general medicine inpatients with similar quality and efficiency as compared to house staff teams. We report our experience during the first year of implementing the service, and compare quality and efficiency of care on this service with that of our traditional house staff services. We also evaluate the effects of this service on patient satisfaction and self‐reported house staff workload.

PATIENTS AND METHODS

RESULTS

Patient Satisfaction

Patients were similarly satisfied with their care on the PACE service and on the house staff services. In specific areas and globally, percentages of patients satisfied with their physicians and with the discharge process were not different, as measured by the Press‐Ganey survey (Press‐Ganey Associates, South Bend, IN; Figures 1 and 2). The survey distinguishes between attendings and residents, but not physician assistants; therefore, Figure 1 only includes responses to the attending questions. Given the sampling procedure of the Press‐Ganey survey, exact response rates cannot be calculated, but Press‐Ganey reports a response rate of about 40% for the English survey and about 20% for the Spanish survey.

Figure 1

Figure 2

DISCUSSION

As academic centers struggle with developing a workforce to provide patient care no longer provided by residents, questions about the ideal structure of nonhouse staff inpatient services abound. Although solutions to this problem will be determined to some extent by local factors such as institutional culture and resources, some lessons learned in developing such services will be more widely applicable. We found that by implementing a geographically localized, physician assistant‐staffed hospitalist service, we were able to provide care of similar quality and efficiency to that of traditional house staff services, despite inexperienced hospitalists staffing the service and a medical residency program commonly recognized as one of the best in the country. Adjusted total costs were slightly lower on the PACE service, but this difference was small and of borderline statistical significance. Likewise, no significant differences were seen in any of several quality measures or in patient satisfaction.

Our findings add to the available evidence supporting the use of physician assistants on academic general medicine services, and are germane to academic centers facing reductions in house staff availability and seeking alternative models of care for inpatients. Several specific characteristics of the PACE service and the implications of these should be considered:

• 1

  The service accepted all patients, regardless of diagnosis, acuity, or complexity of illness. This was unlike many previously described nonhouse staff services which were more limited in scope, and allowed more flexibility with patient flow. However, in the end, patients on the PACE service did have a modestly lower case mix index and Charlson score, suggesting that, despite a lack of triage guidelines, there was some bias in the triage of admissions, possibly due to a perception that physician assistants should take care of lower complexity patients. If it is desirable to have a similar distribution of higher complexity patients across house staff and nonhouse staff services, extra efforts may be necessary to overcome this perception.

• 2

  The service was geographically regionalized. Geographic regionalization offered many important advantages, especially with regards to communication among staff, nursing, and consultants, and allowed for multidisciplinary rounds. However, it is possible that the modest, but not statistically significant, trend toward an increased LOS seen on the PACE service might be a reflection of geographic admitting (less incentive to discharge since discharging a patient means taking a new admission).

• 3

  The education and professional development of the physician assistants was a priority. Physician assistants had considerable autonomy and responsibility, and rather than being assigned only lower level administrative tasks, performed all aspects of patient care. They also received regular teaching from the hospitalists, attended house staff teaching conferences, and developed nonclinical roles in education and quality improvement. The higher standards expected of the physician assistants were quite possibly a factor in the quality of care delivered, and almost certainly contributed to physician assistant satisfaction and retention.

Our findings contrast with those of Myers et al.,9 who found that a nonteaching service staffed by hospitalists and nurse practitioners had a significantly lower median LOS and hospital charges compared to similar patients on resident‐based services. However, unlike ours, their service cared for a select patient population, and only accepted patients with chest pain at low risk for acute coronary syndrome. Van Rhee et al.10 found that physician assistants on a general medicine service used fewer resources for patients with pneumonia, stroke, and congestive heart failure than resident physicians, and did not exceed the resources used by residents in other diagnoses. The authors did not find a difference in LOS, but did find a significantly higher mortality among patients with pneumonia cared for by physician assistants.

Several limitations should be noted. First, the study was a retrospective analysis of administrative data rather than a randomized trial, and although we employed a standard approach to adjust for a wide range of patient characteristics including severity of illness, there may have been undetected differences in the patient populations studied that may have confounded our results. Second, resident moonlighters admitted patients to the PACE service and, at other times, to the house staff services, and this may have diluted any differences between the groups. However, when we limited our analysis to the subgroup of patients admitted during the day, similar results were obtained, with the exception that the PACE service had a lower rate of 14‐day readmissions, an unexpected finding deserving of further study. Third, the study was conducted in a single academic institution and our findings may not be generalizable to others with different needs and resources; indeed, the costs associated with implementing such a service may be prohibitive for some institutions. Fourth, because of simultaneous changes that were taking place in our residency program, we are unable to accurately assess the impact of the PACE service on resident duty hours. However, resident duty hours did not increase over this time period on the general medicine service, and implementation of the service allowed for redistribution of house staff to other services and positions. Fifth, patient satisfaction data were obtained from responses to the mailed Press‐Ganey survey, to which there is a relatively low response rate. Also, we did not survey providers regarding their satisfaction with the service during the study period. Sixth, the study had limited power to detect clinically important differences in mortality and ICU transfers. Finally, this study is unable to compare this particular model of incorporating midlevel providers into general medical services with other models, only with traditional house staff services.

Future research should focus on determining the most effective and efficient ways to incorporate midlevel providers on academic general medicine services. One important question from the standpoint of house staff training is whether such services should be separate but equal, or should house staff gain experience during residency working with midlevel providers, since they are likely to encounter them in the future whether they stay in academics or not. Different models of care will likely have large implications for the quality and efficiency of patient care, house staff education and satisfaction, and physician assistant job satisfaction and turnover.

In summary, our study demonstrates that a geographically regionalized, multidisciplinary service staffed by hospitalists and physician assistants can be a safe alternative to house staff‐based services for the care of general medicine inpatients in an academic medical center.