Evaluating Fairness and Algorithmic Bias in Ensemble-Based Diabetes Risk Prediction Models Across Diverse Socio-Demographic and Ethno-Racial Subpopulations

• Asher Noah

  covenant University

  Author

The integration of artificial intelligence into clinical risk prediction has demonstrated significant promise for early diabetes detection, yet mounting evidence reveals that algorithmic models can inadvertently perpetuate or exacerbate health disparities across socio-demographic and ethno-racial groups. This study addresses the critical gap in fairness evaluation within ensemble-based diabetes risk prediction by systematically assessing algorithmic bias across diverse subpopulations. Utilizing the PIMA Indian Diabetes dataset augmented with synthetic socio-demographic attributes, we implemented and evaluated multiple ensemble learning architectures including AdaBoost, Gradient Boosting, XGBoost, and a Stacking Ensemble, with fairness metrics including disparate impact, statistical parity difference, and equalized odds. The Stacking Ensemble achieved the highest overall predictive performance with 86.0% accuracy and 0.91 ROC-AUC ; however, significant disparities emerged across subpopulations, with false negative rates disproportionately affecting minority groups by margins of 12-18%. Fairness-aware optimization reduced these disparities by 42% while maintaining 83.4% accuracy. These findings demonstrate that high predictive performance does not guarantee equitable outcomes, underscoring the necessity of integrating fairness constraints throughout the model development lifecycle to ensure that AI-driven diabetes risk assessment serves all populations equitably.

• PDF

Copyright (c) 2026 Asher Noah (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.