The Problem
Systemic glucocorticosteroids, or steroids, are prescribed for a variety of indications and often for long durations in the hematology, oncology, and bone marrow transplant patient populations for their anti-inflammatory, anti-emetic, and tumor suppression properties. Unfortunately, exogenous steroids are frequently associated with a host of side effects including hyperglycemia which can occur in patients both with and without a history of diabetes. Hyperglycemia and other adverse effects may limit how long steroids may effectively be used and a patient’s overall prognosis. For example, a combination of cyclophosphamide, vincristine, doxorubicin, and dexamethasone known as hyper-CVAD alternating with rituximab, methotrexate, and cytarabine is used as part of induction therapy for acute lymphoblastic leukemia (ALL). In a 2004 study, researchers found that 103 of the 278 patients (37%) had hyperglycemia as defined as two or more random blood glucose levels ≥ 200 mg/dL. These patients had significantly shorter durations of complete response (24 vs 52 months, p=0.001) and median survival (29 vs 88 months, p<0.001) compared to those without hyperglycemia.Patients also had higher rates of infection (71.8% vs 56.0%, p=0.009) and sepsis (16.5% vs 8.0%, p=0.03). Although it is difficult to discern whether hyperglycemia is a direct cause of poor outcomes, this study provides convincing evidence that more should be done to promptly identify, prevent, and treat hyperglycemia.(1)
A protocol using insulin neutral protamine Hagedorn (NPH) in a small patient population over a 3-day period has been published, but it was primarily based on the authors’ clinical experience.(2) Because of the frequency of steroid induced hyperglycemia and its association with poor outcomes, Duke University Health System (DUHS) must develop a validated protocol to proactively prevent and treat steroid induced hyperglycemia.The protocol will also need to address the prevention of hypoglycemia or overcorrection of hyperglycemia. Insulin is designated as a high-alert medication by the Institute for Safe Medication Practices (ISMP) because it carries a heightened risk of causing significant patient harm when used inappropriately.(3) Therefore, the protocol must address both efficacy and safety concerns.
Our Solution
The purpose of this project is to develop a predictive model to identify patients at risk of steroid-induced hyperglycemia and optimize treatment approaches for these patients. The predictive model gathers risk factor data from adult patients receiving high-dose corticosteroids to identify those at risk of developing steroid-induced hyperglycemia. Once the model identifies patients, a clinical team will review the patients and determine if interventions in care are warranted. Once the model and workflow are tested, the next phase of the project is to enhance the model so that treatment approaches for patients at risk of developing steroid-induced hyperglycemia can be determined.
Systemic glucocorticosteroids are prescribed for a variety of indications and may be used at high doses and for long durations in some patient populations. Unfortunately, exogenous steroids are frequently associated with a host of side effects including hyperglycemia, increased risk for infection, Cushingoid symptoms, increased risk of cardiovascular events, hyperlipidemia, development of diabetes, and others. Steroid-induced hyperglycemia may occur in patients both with and without a history of diabetes. Due to the frequency of steroid-induced hyperglycemia and its association with poor outcomes, a tool that can standardize the approach of identifying and managing steroid-induced hyperglycemia is desired.
The development of a machine learning model and complementary workflow was achieved by the team. Various machine learning approaches were utilized to determine if a model could predict which patients receiving high dose corticosteroids would develop sustained hyperglycemia. A cohort of 11,995 inpatients seen at Duke University Hospital (DUH) between Oct 2014 and Aug 2018 was selected. Patients in this cohort were ≥ 18 years old and received high dose corticosteroids (≥ 20mg/day of prednisone equivalents). Thirty-two features were chosen based upon the expertise of an interdisciplinary team including an inpatient endocrinologist and pharmacist. The outcome of hyperglycemic events was defined as patients having two blood glucose values above 180 mg/dL within 12 hours of corticosteroid administration. The models were evaluated using k-fold cross-validation on data from DUH and validated on patient data sets from Duke Regional Hospital (DRH) and Duke Raleigh Hospital (DRAH). Model performance was evaluated using AUROCs and AUPRCs. interpretation by two independent cardiologists.
A clinical workflow was created in coordination with a multidisciplinary team that included endocrinologists and pharmacists. Twenty hours were spent on the creation and validation of this workflow. The purpose of the workflow was to facilitate the integration of the machine learning model in the inpatient environment. The pilot of the workflow and model integration is being deployed in a single unit of Duke University Hospital where many patients receive a high dose of corticosteroids.
The predictive approach of the model can impact the safe care of patients. Each morning, the model generates a report of patients at risk of developing steroid-induced hyperglycemia. Based on this report, the project team is in the process of validating the predictive model output via a pilot in the inpatient adult bone marrow transplant unit. A team that includes endocrinology fellows reviews the report and examines patient charts to determine whether intervention in the care of patients is needed to address the risk of hyperglycemia. Poster presentations were given at ADA and MLHC 2019 conferences. Once validation of the model and workflow is complete, the next steps for integrating this model into clinical care will be assessed.
The predictive approach of the model can impact the safe care of patients. Each morning, the model generates a report of patients at risk of developing steroid-induced hyperglycemia. Based on this report, the project team is in the process of validating the predictive model output via a pilot in the inpatient adult bone marrow transplant unit. A team that includes endocrinology fellows reviews the report and examines patient charts to determine whether intervention in the care of patients is needed to address the risk of hyperglycemia. Poster presentations were given at ADA and MLHC 2019 conferences. Once validation of the model and workflow is complete, the next steps for integrating this model into clinical care will be assessed.
References
- Weiser MA, Cabanillas ME, KonoplevaM, et al. Relation between the duration of remission and hyperglycemia during induction chemotherapy for acute lymphocytic leukemia with a hyperfractionated cyclophosphamide, vincristine, doxorubicin, and dexamethasone/methotrexate-cytarabine regimen. Cancer. 2004; 100: 1179-85.
- Grommesh B, Lausch MJ, Vannelli AJ, et al. Hospital insulin protocol aims for glucose control in glucocorticoid-induced hyperglycemia. Endocr Pract. 2016;22(2):180-9.
- Institute for Safe Medication Practices (ISMP). List of High-Alert Medications in Acute Care settings. Institute for Safe Medication Practices. 2014. Available at http://www.ismp.org/tools/institutionalhighAlert.asp. Accessed September 29, 2017.
