Privacy-Preserving and Fair Federated Data Mining for Early Prediction of Health Crises: An Empirical Study on Baghdad Hospitals

Abstract

Early health crises, such as sepsis, cardiac decompensation, and respiratory failure, kill patients who might otherwise survive, and the problem is worse in resource-limited settings like Iraq, where data infrastructure is fragile, and populations are routinely underrepresented in predictive models. The study built FedFairMine to address this directly. It is a federated data mining framework that keeps raw patient records on local hospital servers while still training a shared predictive model. Three design choices drive it: gradient-based feature extraction runs entirely on-device, so no raw data leaves the hospital; a fairness-weighted aggregation step explicitly corrects for performance gaps across age, gender, and socioeconomic groups; and a differential privacy layer tuned specifically for Iraqi electronic health records. The study tested the system across six Baghdad teaching hospitals — Al-Kindi, Ghazi Al-Hariri, Ibn Sina, Medical City, Al-Yarmouk, and Al-Kadhimiya — using 47,312 de-identified patient episodes from 2021 to 2024. FedFairMine reached an F1-score of 0.912, beating FedAvg, FedProx, and Scaffold by 8.3%, 5.1%, and 3.7%. Demographic disparity dropped to an equalized odds difference of 0.021, a 76% improvement over centralized baselines. The results show that predictive accuracy, demographic fairness, and patient privacy are not competing goals; they can be achieved together.