SHM Converge Scientific Abstract ePoster Gallery-Machine Learning-Based Prediction of Hyperglycemia in Hospitalized Patients With Diabetes: An Eight-Year Study of Ehr Records Across 19 Hospital Sites

Machine Learning-Based Prediction of Hyperglycemia in Hospitalized Patients With Diabetes: An Eight-Year Study of Ehr Records Across 19 Hospital Sites

Back to course

Pdf Summary

This study used electronic health record data from 161,932 hospital admissions of adults with diabetes across 19 Mayo Clinic sites in Arizona, Minnesota, Wisconsin, and Florida from July 2017 to March 2024. The goal was to build a machine learning model to predict inpatient hyperglycemia and identify factors associated with it.<br /><br />Hyperglycemia was defined as at least one blood glucose value above 180 mg/dL during the admission. Researchers engineered 33 features from demographics, comorbidities, medications, lab values, and rurality of residence. They also captured temporal trends in labs such as HbA1c and creatinine using linear mixed-effects models. An XGBoost model was trained with an 80/20 train-test split, and SMOTE was used to address class imbalance. SHAP was applied to interpret the model.<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. Hyperglycemia occurred in 72.7% of admissions, and 16.2% had 30-day readmission.<br /><br />The model performed well, with overall accuracy of 0.79. For hyperglycemia, precision was 0.83, recall was 0.91, and F1 score was 0.87. The top predictors were HbA1c temporal trend, hospital steroid use, home long-acting insulin use, age-adjusted Elixhauser Comorbidity Index, and home short-acting insulin use.<br /><br />Higher risk was linked to rising HbA1c, steroid and antibiotic use during hospitalization, home insulin use, higher comorbidity burden, type 1 diabetes, rural residence, and night or weekend admission. Rising creatinine and hemoglobin trends, along with a higher diabetes composite score, were associated with fewer hyperglycemic episodes.<br /><br />The authors conclude that this is the largest multi-site study of hospital hyperglycemia to date and may help guide monitoring and resource allocation.

Asset Subtitle

Yuelei Fu

Meta Tag

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

Category Research

Concept Hyperglycemia

Concept Diabetes Mellitus

Concept Machine Learning

Concept Predictive Model

Concept XGBoost

Distinguished Non-Finalist

Presenter Organization Mayo Clinic

Presenting Author Yuelei Fu

Track Outcomes Research

Keywords

diabetes

hyperglycemia

machine learning

electronic health records

XGBoost

hospital admissions

HbA1c trend

steroid use

SHAP

readmission

Hyperglycemia

Diabetes Mellitus

Machine Learning

Predictive Model

XGBoost