Osagie, Efosa ORCID: https://orcid.org/0009-0004-3462-7175, Shemi, Ayo-Ogbor and Balasundaram, Rebecca (2026) A Fairness-Aware Machine Learning Framework for Sexual and Reproductive Health: Evaluating Algorithmic Bias Across Models. Journal of Data Science and Intelligent Systems.

Preview

Text Accepted Manuscript - A Fairness-Aware Machine Learning Framework for Sexual and Reproductive Health_ Evaluating Algorithmic Bias Across Models.pdf - Accepted Version Download (590kB) | Preview

Abstract

Advances in computational infrastructure and the widespread adoption of Electronic Health Record (EHR) systems have accelerated the integration of Artificial Intelligence (AI) and Machine Learning (ML) into Sexual and Reproductive Health (SRH) services. These technologies enhance diagnostic accuracy, support clinical decision‑making, and enable predictive analytics using diverse healthcare data. However, biases within training datasets can produce unfair outcomes, particularly for underrepresented groups. This study proposes a fairness‑aware ML framework designed to detect and mitigate algorithmic bias in SRH services.The framework is evaluated using two open‑source datasets: a large SRH dataset from England (2014–2015) containing 2,126,413 records, and the PCOS dataset covering the top 75 countries, enabling assessment of generalisability and intersectional fairness. It integrates pre‑processing, in‑processing, and post‑processing techniques, including model‑specific and group‑specific thresholding. Results show that on the SRH England dataset, Logistic Regression (LR) achieved near‑optimal parity fairness with minimal performance loss, improving Disparate Impact from 0.99 to 1.00 while maintaining 0.66 accuracy. Random Forest (RF) and Gradient Boosting (GB) exhibited larger fairness shifts, with Disparate Impact decreasing from 0.94 to 0.66 (RF) and 0.93 to 0.77 (GB), though accuracy remained stable. On the PCOS dataset, LR reduced bias with only a 1.96% accuracy drop, while GB improved performance but saw fairness decline, with Disparate Impact falling from 1.08 to 0.57. RF improved fairness but experienced a 28% accuracy reduction. Overall, the findings show that fairness‑aware ML can substantially reduce bias, though equity–performance trade‑offs vary across models and datasets.

Item Type:	Article
Status:	Published
DOI:	10.47852/bonviewJDSIS62027678
Subjects:	Q Science > QA Mathematics Q Science > QA Mathematics > QA75 Electronic computers. Computer science R Medicine > RA Public aspects of medicine R Medicine > RA Public aspects of medicine > RA0421 Public health. Hygiene. Preventive Medicine
School/Department:	London Campus
URI:	https://ray.yorksj.ac.uk/id/eprint/14755

University Staff: Request a correction | RaY Editors: Update this record