SHM Converge Scientific Abstract ePoster Gallery-Predicting Hypoglycemia Using Machine Learning in Hospitalized Patients With Diabetes: A Large-Scale Study Across 19 Hospitals (2017-2024)

Predicting Hypoglycemia Using Machine Learning in Hospitalized Patients With Diabetes: A Large-Scale Study Across 19 Hospitals (2017-2024)

Back to course

Pdf Summary

This Mayo Clinic study used electronic health record data from 161,932 hospital admissions of adults with diabetes across 19 sites in Arizona, Minnesota, Wisconsin, and Florida between July 2017 and March 2024 to predict in-hospital hypoglycemia and identify associated risk factors. Patients were included if they were at least 18 years old, in the diabetes registry, and had research authorization.<br /><br />The model’s outcome was hypoglycemia, defined as at least one blood glucose value below 70 mg/dL during the admission. Researchers developed 33 features from demographics, comorbidities, medications, laboratory values, and rurality based on RUCA codes. They also modeled laboratory trends over time, such as hemoglobin A1c and creatinine, using linear mixed-effects models. An XGBoost machine learning model was trained with an 80/20 train-test split, and SMOTE was used to address class imbalance. SHAP was applied to improve interpretability.<br /><br />The cohort had a mean age of 65.5 years; 57.2% were men, 87.9% were white, and 64.1% lived in urban areas. Hypoglycemia occurred in 12.1% of admissions at level 1 and 4.6% at level 2. The model achieved overall accuracy of 0.86. However, performance for identifying hypoglycemia cases was modest, with precision of 0.44, recall of 0.14, and F1 score of 0.22.<br /><br />Key predictors associated with more hypoglycemia included home long-acting insulin use, home short-acting insulin use, higher age-adjusted Elixhauser Comorbidity Index, type 1 diabetes, increasing hemoglobin trend, and steroid use during hospitalization. Factors linked to fewer hypoglycemic episodes included increasing creatinine trend, oral non-hypoglycemic agents at home, and type 2 diabetes.<br /><br />The authors conclude that machine learning can help predict hospital hypoglycemia and support targeted monitoring and resource use, though data quality limitations and missing medication records may affect reliability.

Asset Subtitle

Yuelei Fu

Meta Tag

Author List Janna C. Castro, Matthew Gill, Sagar Dugani, Shant Ayanian, Yuelei Fu

Category Research

Concept Hypoglycemia

Concept Machine Learning Model

Concept Structured EHR Data

Concept Feature Engineering

Concept SHAP

Distinguished Non-Finalist

Presenter Organization Mayo Clinic

Presenting Author Yuelei Fu

Track Patient Safety

Keywords

Mayo Clinic

diabetes

hypoglycemia

electronic health records

hospital admissions

machine learning

XGBoost

SHAP

risk factors

predictive model

Hypoglycemia

Machine Learning Model

Structured EHR Data

Feature Engineering

SHAP