Assessing Bias Within Diabetes Risk Prediction in Machine Learning Techniques

Abstract

As machine learning (ML) models increasingly influence clinical decision-making, concerns over fairness and bias have emerged, particularly in predictive healthcare tools. This study investigates racial disparities in diabetes risk prediction using Random Forest, XGBoost, and Linear Regression models trained on the 2021-2023 National Health and Nutrition Examination Survey (NHANES) dataset. Model performance was evaluated across key metrics—accuracy, sensitivity, specificity, F1 score, and ROC AUC—both overall and stratified by race and ethnicity. Initial results revealed strong aggregate performance (ROC AUC greater than 0.91) but significant disparities in recall and precision for underrepresented racial groups. After incorporating demographic stratification during preprocessing, prediction improved across minority groups, though overall model performance slightly declined. These results emphasize the trade-off between model performance and equity, and underscore the need for fairness-aware approaches in healthcare AI.

Ähnliche Arbeiten