A federated learning-based privacy-preserving prediction model for cross-organizational data governance

Cross-organizational data collaboration plays a critical role in advancing public health prediction, particularly for diseases that require coordinated multinational surveillance such as Neglected Tropical Diseases. However, three intertwined challenges severely constrain such collaboration: rigid data sovereignty barriers imposed by international privacy regulations that prohibit centralization of raw clinical records, residual privacy leakage risks even when only model gradients are shared, and pronounced statistical heterogeneity arising from substantial disparities in disease prevalence, healthcare infrastructure, and socioeconomic conditions across participating regions. Addressing these compounded challenges in a single unified framework remains an open problem. To this end, this paper proposes FedPrivGov, a federated learning-based privacy-preserving prediction framework tailored for cross-organizational data governance. FedPrivGov organically integrates Federated Averaging, a calibrated differential privacy noise-injection mechanism, and a secure model-update aggregation protocol, enabling multi-organizational collaborative modeling without any raw data leaving local nodes. Adopting Neglected Tropical Disease infection-outcome prediction as the application scenario, we construct FedNTD-v1, a cross-border distributed dataset encompassing six countries, six disease categories, and 1,650 multi-dimensional records, organized into 324 federated client nodes that faithfully simulate a real-world non-independent and identically distributed environment. Comprehensive experiments demonstrate that FedPrivGov substantially outperforms the local-only training baseline, raising global accuracy from 74.3% to 81.2% and area under the ROC curve from 0.761 to 0.833, while incurring only approximately one percentage point of accuracy loss under the standard privacy budget. This study provides a deployable technical pathway for privacy-preserving collaborative modeling of cross-organizational sensitive data, offering important reference value for public health data governance.