Clinical Decision Support for Glycemic Management Reduces Hospital Length of Stay

Hyperglycemia and hypoglycemia are common among hospitalized patients, predisposing to poor outcomes. Hyperglycemia affects millions of hospitalized patients annually and is associated with an increased risk of infections and noninfectious complications (1,2), mortality (3,4), resource utilization, hospital length of stay (LOS), and costs of care (5). Approximately one in four hospitalized patients have a diagnosis of diabetes, and many more without diabetes develop hyperglycemia due to acute illness (3), making inpatient hyperglycemia a significant concern. Hyperglycemia incidence ranges from 32 to 38% and is higher in severely ill patients, rising to ∼80% after cardiac surgery (6,7). Hypoglycemia increases the odds of mortality and complications among hospitalized patients (8,9). To manage and prevent dysglycemia, standard-of-care recommendations provide comprehensive guidance for adequate and safe treatments (2,6,10).

Unfortunately, multiple barriers within health care systems obstruct adequate glycemic control. These barriers are related to hospital processes, glucose data gathering and interpretation, insulin administration choices, and gaps in health providers’ knowledge, attitudes, and clinical decision making (11–19). Addressing these obstacles requires adaptations within the systems of practice to correct deficiencies at different levels. Clinical decision support (CDS) is a potential resource to address gaps in glycemic management. CDS uses database-driven and person-specific health information intelligently categorized, managed, and presented in the context of practice to enhance health care processes (20). Multiple CDS modalities have been developed and used in hospitals for glycemic care, including computerized orders, glucose data networks and visualization, case finding tools, insulin calculators, electronic treatment protocols, intravenous and subcutaneous dosing algorithms, and alert systems. Some of these modalities have shown improvements in glycemic control (21–23), reduced clinical complications (24), lowered costs (24,25), and shortened hospital LOS (26). In particular, CDS using alerts to notify providers about abnormal or undesirable scenarios has the advantage of addressing problems in real time and immediately when required. While some studies have shown that alert-based CDS for inpatient glycemic care can improve some intermediary outcomes, such as care processes, providers’ proactivity, and glucose management decisions (27–32), evidence about the clinical and economic impact of alert-based CDS systems is lacking.

We previously reported the development and validation of our alert-based CDS tool integrated in the electronic medical record (EMR) that automatically detects hospital glycemic gaps in care (GIC) in real time (32). GlucAlert-CDS provides notifications to clinicians when glycemic GIC are identified and offers evidence-based recommendations for management of 1) hypoglycemia, 2) recurrent hyperglycemia in patients with type 1 and type 2 diabetes or stress hyperglycemia, and 3) inappropriate insulin use (e.g., sliding scale monotherapy in the context of recurrent hyperglycemia or any time in type 1 diabetes). The objective of implementing this management tool is to overcome common practice barriers that appear to limit glycemic control by recognizing predefined common data elements in the EMR using intelligent algorithms. The tool features the ability to collate glucose data, identify inadequate control, advise on insulin use, and promote proactive management through recommendations aligned with standards of care. In a study of 3,588 adult patients, the GlucAlert-CDS tool significantly reduced recurrent hyperglycemia in those with type 1 and type 2 diabetes by 10% and in those with stress hyperglycemia by 43%. The tool also reduced inappropriate insulin use in patients with type 1 and type 2 diabetes by 55 and 8%, respectively (32).

In the current study, we hypothesized that the GlucAlert-CDS tool, which is able to reduce glycemic GIC, can reduce hospital LOS among adult patients admitted to non–intensive care settings. We discuss the relevance of our findings from clinical and quality-of-care perspectives.

We integrated and validated a CDS tool in the EMR and implemented it in clinical practice across inpatient units at an academic medical center (32). During a 12-month hospital-wide pilot study, we used an interrupted time series design to activate and deactivate the CDS tool every 3 months from March 2018 to February 2019, starting with an active period. A GIC in the study is defined as the presence of a glycemic abnormality identified by the CDS tool. An alert is defined as a notification evoked in real time in the EMR about a GIC only during an active period and received by a provider caring for the patient. Notifications included evidence-based management recommendations. Monitoring of GIC occurred during active and inactive periods, enabling 6 months of data accrual for each period. GIC, such as glucose abnormalities and inappropriate insulin use, were detected using algorithms that followed rules and recognized common data elements. GIC were interrogated continuously, but notifications were only evoked during active periods. No detected gaps denotes that there had been no glycemic abnormality or inappropriate insulin use as defined by the CDS tool. Abnormal blood glucose (BG) categories and criteria were 1) severe hyperglycemia (BG ≥13.9 mmol/L [250 mg/dL] at least once) or recurrent hyperglycemia in patients with diabetes or stress hyperglycemia (BG ≥10.0 mmol/L [180 mg/dL] at least twice, 3 h apart) and 2) impending or established hypoglycemia (any BG 3.9–4.44 mmol/L [70–80 mg/dL] or ≤3.9 mmol/L [70 mg/dL], respectively). The inappropriate insulin use category was defined as the use of sliding scale monotherapy for recurrent hyperglycemia (BG ≥10.0 mmol/L [180 mg/dL] at least three times, 3 h apart) in any patient with 1) stress hyperglycemia in the absence of diabetes diagnosis, 2) type 2 diabetes, or 3) any time in patients with type 1 diabetes. These categories of GIC events represented the independent variables, and LOS represented the dependent variable (Table 2).

We included medical (medicine, family and community medicine, neurology, and medical subspecialties) and surgical (general surgery, surgical subspecialties, and gynecology) hospitalizations of patients aged ≥18 years in whom glycemic GIC were identified in any of the categories above. We correlated occurrence of GIC events and LOS during active and inactive periods of the CDS tool. Providers receiving notifications during active periods could have also cared for hospitalized patients during inactive periods, at which time they were not receiving alerts. Any providers from admitting services were potential recipients of notifications if GIC were detected among their patients. Before implementation, providers were informed about GlucAlert-CDS tool during division or department meetings and e-mail notifications. This study was approved by the Penn State College of Medicine institutional review board.

Analyses were performed using SAS 9.4 statistical software (SAS Institute, Cary, NC). Data are summarized according to patients’ demographic characteristics (Table 1) and admissions and LOS outcomes (Table 2). Overall data and comparisons included patients and admissions with no GIC events combined with patients and admissions who had GIC events. A separate analysis compared the LOS between study periods for only admissions with no GIC. All subsequent analyses included only patients and admissions with at least one GIC. Admissions with GIC crossing over periods were excluded from the analysis. Demographic characteristics of patients with at least one GIC event were compared between study periods using Wilcoxon rank sum or χ² tests. Table 2 shows LOS outcomes when the CDS tool was active and inactive for the entire sample of admissions with GIC events, including admissions in which the tool addressed hyperglycemia, hypoglycemia, inappropriate insulin use, and all gaps within clinical services. We also examined these subgroups individually within admissions with at least one GIC event. The hyperglycemia subgroup encompassed admissions among patients with diabetes and hyperglycemia or patients with stress hyperglycemia. In the hypoglycemia subgroup, we examined admissions with at least one hypoglycemic GIC evoked after an initial nonhypoglycemic GIC (this refers to an event of hypoglycemia after either hyperglycemia or inappropriate insulin use was detected) or at least one hypoglycemic GIC after an initial hypoglycemic GIC (this refers to events of hypoglycemia after previous hypoglycemia was detected). In the subgroup of insulin use, we examined inappropriately used insulin scales as monotherapy for type 1 diabetes, type 2 diabetes, and stress hyperglycemia. The clinical service subgroup included admissions within medical services or a majority of GIC within surgical services. The LOS was skewed and, therefore, log-transformed for analysis with a linear model for repeated measures. The mean estimates were back transformed by exponentiating the estimates to give geometric mean estimates, and a ratio of the geometric means was used to quantify the magnitude and direction of the differences. A mediation analysis was performed to determine whether the frequency of GIC mediated the effect of the CDS tool on the outcome of LOS. The direct effect, indirect effect, and percent mediated are reported (Table 3). For this analysis, we considered active and inactive periods of the CDS tool as the exposure variable and the number of GIC (events detected for dysglycemia or insulin misuse within clinical services) as the mediator, with LOS as the outcome. We invoked ordinal logistic regression using generalized estimating equations to compare the number of recurrences after an initial hypoglycemic GIC (1, 2, ≥3) between periods, using an odds ratio to quantify the magnitude and direction of the association (Table 4). The effect of the interaction between periods and the number of recurrences after an initial event (1, 2, ≥3) on LOS was also analyzed using the same repeated-measures model (Table 5). All P values in Table 2 were adjusted for multiple comparisons using the false discovery rate method.

Demographic characteristics of 3,482 unique patients, corresponding to 4,788 admissions during the 12-month pilot study are shown in Table 1. This sample showed male and non-Hispanic Caucasian predominance. Approximately two-thirds were admitted to medical services. The sample included 1,950 unique patients corresponding to 2,416 admissions during the CDS tool active period and 1,939 patients corresponding to 2,372 admissions during the inactive period. Groups shown in Table 2 detail the categories of GIC evoked by the CDS tool and the average LOS in hours for all adult hospitalizations. The population included 22,395 admissions corresponding to 16,499 unique patients during the study and comprised two groups. One group included 17,607 admissions in which the CDS tool did not identify GIC, signifying that no glycemic abnormalities were present according to the tool’s algorithm criteria. The other group consisted of 4,788 admissions (study sample) described above with at least one glycemic GIC detected, which was further allocated to the active period when providers received alert notifications and the inactive period when gaps were detected but no alerts were evoked in the patients’ records for providers. Admissions crossing over the active to inactive or inactive to active periods were removed from the analysis to avoid crossover bias. The average LOS of all admissions, with and without GIC during the active and inactive periods was 87.2 and 89.0 h, respectively (P = 0.053). The average LOS of admissions without GIC was 79.3 and 80.7 h during the active and inactive periods, respectively (P = 0.128). The average LOS among admissions with at least one glycemic GIC was 125.2 and 130.9 h during the active and inactive periods, respectively (P = 0.057). While no significant difference was found in LOS during the active and inactive periods (P = 0.413) between the group with at least one GIC (glycemic abnormality present), representing the study sample, compared with the group with and without GIC combined, representing the population, LOS was shorter by 5.7 and 1.8 h in those groups, respectively, when the tool was active. There was a nonsignificant 1.4-h LOS reduction among admissions with no GIC (no glycemic abnormality present). During the active period (when alerts were evoked for providers) the difference in LOS among admissions with GIC (125.2 h) versus admissions without (79.3 h) was 45.9 h (1.9 days). During the inactive period (when alerts were not evoked for providers), LOS of admissions with GIC (130.9 h) versus admissions without (80.7 h) was 50.2 h (2.1 days), as shown in Table 2. This denotes longer LOS when there were glycemic GIC and even longer when the tool was inactive, which is consistent with the evidence supporting longer LOS among patients with dysglycemia (5).

Within admissions with at least one GIC, average LOS during the active versus inactive period was reduced as follows: among all admissions, by 5.7 h (P = 0.57); among admissions of patients with diabetes with at least one hyperglycemic event, by 6.4 h (P = 0.054); among admissions of patients with at least one stress hyperglycemia event, by 31.0 h (P = 0.053); and among admissions of patients with diabetes with the majority of alerts occurring while in medical services, by 8.4 h (P = 0.039). The LOS difference among admissions with at least one hypoglycemic event (denoting recurrent hypoglycemia or impending hypoglycemia) after a previous hypoglycemic event and that received notifications by the CDS tool was evident in the unadjusted analysis (P = 0.025). In the adjusted analysis, which was used to assess the strength of the statistical significance when testing multiple hypotheses, this 29.1-h reduction in LOS (P = 0.074) was no longer significant. Average LOS reduction was not significantly different among the following admission groups: diabetes and the majority of GIC occurring while in surgical services (−2.4 h, P = 0.703); inappropriate insulin use as sliding scale monotherapy in type 1 diabetes (87.4 h, P = 0.192), type 2 diabetes (−7.5 h, P = 0.216), or stress hyperglycemia (−8.5 h, P = 0.260); admissions with one hypoglycemic or impending hypoglycemic event (1.7 h, P = 0.798); or admissions with one or more hypoglycemic event after a nonhypoglycemic event notification (nonhypoglycemic events are hyperglycemia and inappropriate insulin use) (−2.7 h, P = 0.798).

In Table 3, we show the effect of the active and inactive periods on LOS that was mediated by the frequency of glycemic GIC alerts. The mediation analysis aimed to assess the direct effect of the exposure and the indirect effect of the exposure through the mediator on the outcomes. This refers to the influence of providing notifications and recommendations through the active CDS tool (exposure) on glycemic management leading to LOS reductions (outcome). The direct effect of the CDS status was not statistically significant for all admissions with at least one GIC event (P = 0.468). However, the indirect effect (when GIC were present) through the mediation of the total number of GIC was statistically significant (P = 0.021). Similarly, among admissions of patients with diabetes in medical services, the direct effect of the CDS status was not statistically significant (P = 0.175), but the indirect effect of CDS status through the mediation of the total number of alerts was significant (P = 0.045). The percentage of admissions with GIC, in which the active CDS tool indirectly and significantly influenced LOS during the active period, was 69% for all admissions (P = 0.027) and 52% for those occurring in medical services (P = 0.019). Findings of all other subgroups are shown in Table 3.

To further understand the CDS tool’s influence on LOS reduction among admissions with recurrent hypoglycemia, we determined frequency and percentages of these recurrences (Table 4). The proportion of admissions with recurrent hypoglycemia (i.e., having two episodes) was larger when the tool was active versus inactive (57.9 vs. 54.4%). However, the proportion of admissions with multiple hypoglycemic events, having either three episodes (22.4 vs. 24.6%) or four or more episodes (19.7 vs. 21.0%), was smaller when the CDS tool was active versus inactive. To analyze the significance of recurrent hypoglycemia, we used ordinal logistic regression with the number of recurrences as the outcome variable comparing active and inactive periods. While the odds ratio was 0.88 (95% CI 0.64, 1.21), suggesting that the active period was less likely to allow multiple recurrences than the inactive period, the association was not statistically significant. Therefore, we conducted a post hoc analysis of this unique subsample with events of recurrent hypoglycemia, which demonstrated that our study only had 9% power to detect this particular subgroup’s difference, likely because of the small number of admissions (n = 592) with recurrent hypoglycemia. Table 4 shows the frequency of recurrent hypoglycemia during the study periods. It is anticipated that a larger sample size would show a significant difference.

Additionally, we analyzed differences in LOS between the active and inactive periods across admissions when patients experienced recurrent hypoglycemia, as shown in Table 5. The mean LOS when the CDS tool was active versus inactive for those experiencing two hypoglycemic events during the active period was reduced by 5.2 h (P = 0.655). Admissions with four or more hypoglycemic events while the tool was active had a 14.8-h reduction in LOS versus while the tool was inactive (P = 0.746). In these groups, the difference did not appear to be significant. However, for admissions with three hypoglycemic events, LOS was significantly reduced by 82.3 h (P = 0.006) during the active period. All three scenarios of recurrent hypoglycemia had shorter LOS when the CDS tool was actively alerting (about hypoglycemia, but statistical significance was only present for patients experiencing three hypoglycemic events.

We introduced our inpatient glycemic control GlucAlert-CDS tool and activated it intermittently every 3 months in the EMR for 6 active and 6 inactive months. The tool contributed to an LOS reduction averaging 5.7 h across all adult admissions with one or more hyperglycemic events. We propose that to the extent that hospital days are considered a limited resource, shorter LOS may result in cost benefits. Approximately 25% of hospitalization days in the U.S. are incurred by patients with diabetes (5). Our reduction in mean LOS by 5.7 h amounted to a 4.4% reduction in LOS (5.7 of 130.9 h as shown in Table 2) in the study sample. Given that the economic costs of diabetes grew by 26% from 2012 to 2017 because of increased prevalence and cost per person (5), it is increasingly important to also focus on inpatient glycemic control. Addressing glycemic abnormalities in the hospital may result in better resource utilization and shorter LOS (4,33). Additionally, decreasing LOS may mitigate the social impact of hospitalization by reducing time missed from family life, time lost from work, and potential out-of-pocket expenses (34,35). Our ongoing work envisions comprehensively assessing the economic impact of our GlucAlert-CDS tool in order to determine whether tangible cost benefits may be facilitated by this glucose management system.

In our study, LOS reductions averaged 6.4 h among patients with diabetes experiencing hyperglycemia and 31 h among patients with stress hyperglycemia. Our findings among patients with stress hyperglycemia suggest that hyperglycemia in this population may go overlooked, and the alert-based CDS system may be particularly effective in such groups, especially since poor outcomes have been previously observed in these patients (3,36). In a previous study, we reported that the CDS tool influenced control by significantly reducing recurrent hyperglycemia in patients with type 1 and 2 diabetes by 10% and in patients with stress hyperglycemia by 43% (32). In this study, such gains contributed to reduced LOS. Hyperglycemia in hospitalized patients with and without diabetes is associated with poor outcomes (1,2,6,36) and increased costs and use of resources, including transfers to intensive care and longer LOS (4,5,37).

Anticipating that GlucAlert-CDS could reduce gaps in glycemic care, and as a result shorten LOS, we performed a mediation analysis (Table 3). This analysis showed how much the effect of the exposure (CDS being active vs. inactive) directly and indirectly (when GIC were present) affected the outcome (LOS) through the mediator (the number of GIC) among all admissions with at least one GIC. The tool being active or inactive appeared to affect the number of GIC events, and the number of GIC events affected LOS. The CDS tool being active in the EMR, offering recommendations to providers, indirectly and significantly reduced LOS through its influence on the number of GIC events in some subgroups (Tables 3 and 5).

When GlucAlert-CDS was active, admissions with recurrent hypoglycemia had a statistically nonsignificant average LOS reduction of 29 h. Table 4 shows a lower percentage of recurrent hypoglycemia in the active period when three or more events occurred. Table 5 denotes that statistically significant LOS reduction of 82 h was observed in patients experiencing three hypoglycemic events. There was a substantial, yet nonsignificant, reduction in mean LOS by 5 and 15 h during the active versus inactive period when two or four or more hypoglycemic events occurred, respectively. In interpreting the nonsignificance of these findings, it is possible that two or fewer hypoglycemic events may not be as impactful and that patients with four or more hypoglycemic events may have had other issues contributing to LOS that GlucAlert-CDS could not help to overcome. Additionally, the small size of the subgroups may have influenced statistical significance. Some nonstatistically significant findings might be clinically significant and should be explored in more depth in larger studies.

In our study, a significant reduction in mean LOS of 8.4 h was attained among admissions to medical services, which accounted for 70% of the cohort. An 8-h reduction in LOS, representing nearly a day’s shift, is a desirable outcome because it may potentially affect hospital throughput and use of resources. Possibly, medical services were more attentive to recommendations for hyperglycemia, hypoglycemia, and insulin appropriateness than surgical services, hence influencing LOS. Future work will focus on differential effects of the GlucAlert-CDS tool across specialties and services.

Among patients with type 1 diabetes receiving insulin as sliding scale monotherapy, interestingly, there was a large (albeit nonsignificant) increase in LOS during the active period. Because our tool alerted about the inappropriateness of sliding scale monotherapy and the risk of diabetic ketoacidosis, providers may have taken extra time to obtain additional history or testing to ensure that no ketosis had occurred. We plan to investigate this further in our ongoing work.

Previous work has suggested that CDS helps to optimize diabetes care and glucose control. Benefits for health care processes are being recognized, and guidelines adopted as clinical and economic outcomes data are accumulating (38–42). Systematic reviews and meta-analysis data have shown that CDS aids decisions in ambulatory (43–46) and hospital-based diabetes care (21,22). CDS is increasingly aligned with recommendations from leading entities (2,6,47). As a clinical tool, it helps to achieve a key promise of health information technology, namely, in assisting quality improvement efforts (20). It is endorsed by the Office of the National Coordinator for Health Information Technology (20,38,48). Therefore, our findings of LOS reduction after implementing the GlucAlert-CDS tool have implications for quality of care, practice improvement, and health informatics applications for decision support. It may be assumed that clinical recovery, patient safety, care quality, resource utilization, and hospital costs can improve with better glycemic control, particularly if considering that dysglycemia among hospitalized patients prolongs LOS and predisposes to complications (3,4,36,49–51).

Our study has several limitations. First, we could not randomly assign patients or simultaneously match to a control group since the CDS tool in the EMR cannot be selectively allocated to patients or clinical units. We therefore pursued the interrupted time series design, which is considered one of the strongest quasi-experimental designs and can be an alternative to evaluate longitudinal effects of interventions when randomization is not possible (52,53). Second, providers may have deliberately or unintentionally dismissed notifications by the system. The design of GlucAlert-CDS is such that after a predetermined period, if criteria for the GIC remain present, notification with clinical recommendations reappear. Third, alert systems may result in fatigue from frequent notifications, excessive alerts may affect providers’ efficiency, workflow, and patient safety (54–56). Studies have suggested that programs must be designed to add value to patient care and minimize provider workload by having relevant notifications for practice (57,58). In our design, we built several attributes to avoid alert fatigue, including individualized support providing relevant alerts in real time; providing tailored, timely information without superseding clinician judgment; displaying daily glucose data assessments; using clinical criteria to evoke sound and current notifications without redundancy; and attending to acute or urgent scenarios. While we considered these features to be an incentive for adopting recommendations, this study did not include an assessment of perspectives of providers or direct analysis of decision making regarding insulin initiation or adjustments. We plan further work to examine the details of decision making with and without the alert system. Fourth, this program addressed glycemic GIC exclusively in the inpatient setting. It is unlikely that providers’ actions for discharge were influenced by the content of the glycemic management notifications or awareness of this study. Whether clinicians were more detailed about postdischarge diabetes care planning while receiving recommendations for inpatient diabetes care was not assessed. Fifth, a potential exists for knowledge carryover. Providers’ performance could have been influenced by an enhanced understanding of glycemic management during inactive periods, since they were exposed to recommendations in previous notifications. While we removed admissions crossing over study periods from the analysis, postalert performance assessment was beyond the scope of this pilot study. Finally, our data contained repeated measurements among patients who may have had multiple admissions. Our statistical analysis did not account for repeated measurements in the mediation analysis. However, it did account for multiple admissions per patient in all other analyses. Bind et al. (59) developed an approach for causal mediation analysis with longitudinal data through generalized mixed-effects models. We can apply this sophisticated mediation model in a variety of ways to our data; however, this is beyond the scope of this report. Our investigative team will pursue this analysis in a future project.

In conclusion, dysglycemia predisposes to poor outcomes and resource utilization among hospitalized patients. Adequate glucose management is hindered by barriers at the provider and system levels. CDS can aid in optimizing glycemic control and overcoming common limitations in management. In this report, we present the implementation of an alert-based CDS tool that reduced glycemic GIC and shortened hospital LOS. Our findings reveal a potential important outcome benefit of CDS for inpatient diabetes care and underscore the need for continuing to assess robust evidence related to the clinical and economic impact of CDS beyond process improvement and glucose management. Our findings are relevant to the domains of quality and efficiency of care and demonstrate a tangible outcome associated with a health informatics technology–enabled quality improvement program.

Funding. A.R.P.-L. received support from National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) career development grant K23DK107914-05 for this project and is supported by NIDDK grant R01DK130992-01. G.E.U. is partly supported by National Center for Advancing Translational Sciences Clinical and Translational Science Award 3UL1TR002378-05S2 and NIDDK grant 2P30DK111024-06.

Duality of Interest. G.E.U. has received unrestricted research support from AstraZeneca and Dexcom. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. A.R.P.-L. led the conduct of the project, including the study design; development and implementation of the alert-based CDS tool; data collection, analysis, and interpretation; and writing of the manuscript. P.H. advised on the study design and integrity and contributed to the interpretation of results and manuscript preparation, review, and edits. G.E.U. contributed to the study design, interpretation of findings, and manuscript review. E.B.L. contributed to the data management, analysis, and interpretation and manuscript preparation. F.T.Q. and L.W. contributed to the interpretation of findings and manuscript review. C.M.R. advised on the implementation of the alert-based CDS tool, review of findings, and manuscript review. C.J.D. contributed to implementation of the alert-based CDS tool, review of findings, and manuscript review. V.M.C. contributed to the study design, data analysis, interpretation of findings, and manuscript review. A.R.P.-L. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this study were presented in abstract form at the NIDDK Network of Minority Research Investigators virtual meeting, 28–30 April 2021, and at the 15th Advanced Technology and Treatment for Diabetes, Barcelona, Spain, 27–30 April 2022.