Exploring Machine Learning Methods to Predict Glycohemoglobin (HbA1c) and Classification with Non-Invasive Physical and Demographic Measurements

Diabetes mellitus (DM) is a growing global health challenge, with glycohemoglobin, or HbA1c, serving as a key biomarker for its diagnosis and management. However, HbA1c testing typically requires blood collection and laboratory resources, which may be inaccessible in low-resource settings. This study investigates the potential of machine learning models to predict HbA1c levels and classify elevated HbA1c (≥6.5%) using non-invasive physical and demographic measurements from the National Health and Nutrition Examination Survey (NHANES) 2021–2023 dataset. Following pre-processing, a total of 4,735 observations were split into a 70% training set (N=3,315) and a 30% testing set (N=1,420) through stratified sampling. Multiple machine learning models were developed to predict the continuous outcome of HbA1c and to classify elevated HbA1c levels. For numerical prediction, models including linear regression with ordinary least squares (OLS), LASSO regression, random forest, and boosted regression demonstrated similar predictive performance, with test RMSEs ranging from 0.9339% to 0.9457% and R² values between 0.1629 and 0.1835. These results suggest that while the selected features were consistent across models, they explained only a small fraction of the variance in HbA1c levels, limiting their clinical utility. For classification, standard logistic regression, LASSO logistic regression, random forest, boosted trees, and support vector machines were evaluated. The standard logistic regression, LASSO logistic regression, and boosted trees achieved the highest AUC-ROC scores (0.7728 to 0.8239), with logistic regression slightly outperforming the others based on Cohen’s Kappa. However, the Cohen’s Kappa between 0.2487 to 0.2836 indicated only fair classification agreement beyond chance and the features included in this analysis alone may not be sufficient for reliable HbA1c classification in clinical practice.

Author

• Zhai, Yang

Language

• en

Date created

• 2025

Type of Work

• Master's Thesis

License

• All rights reserved

Rights statement

• In Copyright

GW Unit

• Public Health

Degree

• M.S.

Advisor

• Elmi, Angelo

Committee Member(s)

• Ciarleglio, Adam

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/th83m024s