Predicting hypoglycemia in critically Ill patients using machine learning and electronic health records

Area of Science:

• Critical Care Medicine
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Hypoglycemia is a frequent complication in critically ill patients, linked to adverse outcomes.
• Effective prediction of hypoglycemia is crucial for timely intervention in intensive care units (ICUs).

Purpose of the Study:

• To develop and validate a machine learning model for predicting hypoglycemia in ICU patients.
• To utilize electronic health record (EHR) data for early identification of hypoglycemia risk.

Main Methods:

• Trained and tested machine learning algorithms on a large, multicenter ICU EHR dataset (eICU Collaborative Research Database).
• Utilized 44 features including demographics, labs, medications, and vitals to predict hypoglycemic events (blood glucose < 72 mg/dL).
• Employed the eXtreme gradient boosting (XGBoost) model, using data from the first two hours of ICU stay.

Main Results:

• The XGBoost model achieved an AUROC of 0.85, with 0.76 sensitivity and 0.76 specificity.
• The model demonstrated strong discrimination and calibration for predicting hypoglycemia.
• Data from 61,575 patients across 199 hospitals were analyzed.

Conclusions:

• The developed machine learning model shows significant potential for predicting hypoglycemia in ICU patients.
• Further prospective trials are needed to assess the clinical utility of this model in preventing hypoglycemia.