Challenges with Implementing Predictive Models for Inpatient Hypoglycemic Events in Clinical Decision Support

Area of Science:

• Clinical Informatics
• Predictive Analytics in Healthcare
• Patient Safety

Background:

• Inpatient hypoglycemia increases length of stay and mortality.
• Existing models aim to predict hypoglycemia risk in hospitalized patients.
• Developing effective predictive tools is crucial for improving patient outcomes.

Purpose of the Study:

• To identify barriers to implementing a developed predictive model for inpatient hypoglycemia.
• To inform the creation of a clinical decision support tool for hypoglycemia prediction.
• To evaluate the feasibility of using predictive analytics in clinical practice.

Main Methods:

• A logistic regression model was trained using data from diabetic patients receiving insulin.
• Data was collected from January 2020 to December 2021 at an academic medical center.
• The model predicted hypoglycemic events (glucose < 70 mg/dL) within 24 hours of borderline-low glucose measurements (70-90 mg/dL).

Main Results:

• The predictive model achieved an area under the curve (AUC) of 0.69 on the validation dataset.
• Despite model development, it was not implemented into clinical practice.
• Key factors influenced the decision against implementation.

Conclusions:

• Implementation was halted due to model predictiveness limitations and contextual factors.
• Predictive analytics may not always be feasible for clinical decision support.
• Challenges exist in predicting low-incidence events when key predictors lack structured EHR documentation.