Differentially private federated learning for localized control of infectious disease dynamics

Clinical Snapshot

Key Findings

• P Value: Not reported — study uses predictive performance metrics rather than hypothesis testing framework

• Effect Size: At ε = 10 (moderately strong privacy): R² = 0.95 (November 2020) and R² = 0.88 (March 2022 Omicron). Non-DP baseline: R² = 0.96 and R² = 0.90 respectively. Absolute utility loss from DP: ΔR² ≈ 0.01–0.02

• Primary Outcome: County-level COVID-19 case count forecasting accuracy using a differentially private federated multilayer perceptron, evaluated across two epidemic phases

• Nnt Or Sensitivity: Not applicable in traditional clinical sense. Relevant operational metric: at ε = 10, MAPE values are reported but partially obscured in abstract rendering. At very strict privacy (ε = 1), forecasts are described as 'unstable and unusable', representing a functional sensitivity threshold for the privacy-utility trade-off

• Confidence Interval: Not reported — a significant methodological limitation given the stochastic nature of differential privacy noise

Clinical Application

Conceptually feasible for jurisdictions with existing digital surveillance infrastructure and technical capacity for federated computation. Practical barriers include: heterogeneous IT infrastructure across local health authorities, requirement for coordinated server architecture, staff technical capacity, and ongoing calibration of privacy budgets as epidemic conditions evolve. The framework requires prospective piloting before operational deployment. Directly relevant to the Australian public health context where state and territory health departments operate under distinct data governance frameworks and the Privacy Act 1988 (Cth) constrains cross-jurisdictional health data sharing. The Australian Centre for Disease Control (ACDC), established in 2024, is actively developing national communicable disease surveillance architecture where privacy-preserving federated approaches could enable collaboration between Commonwealth, state, and territory health authorities without requiring centralisation of notifiable disease data. The My Health Record system and state-based notifiable disease surveillance systems (e.g., NNDSS) operate in siloed environments where federated learning could bridge data utility gaps. RACGP and AHPRA frameworks emphasise patient data privacy, making DP-FL technically aligned with Australian regulatory expectations. PBS and TGA considerations are not directly applicable to this surveillance methodology study, though downstream applications informing antiviral stockpiling or vaccine deployment decisions would engage TGA regulatory frameworks. Local health authorities, public health units, and epidemic surveillance agencies operating within federated administrative structures where data sharing across jurisdictions is constrained by privacy legislation or data sovereignty requirements

Abstract

In times of epidemics, swift reaction is necessary to mitigate epidemic spreading. For this reaction, localized approaches have several advantages, limiting necessary resources and reducing the impact of interventions on a larger scale. However, training a separate machine learning (ML) model on a local scale is often not feasible due to limited available data. Centralizing the data is also challenging because of its high sensitivity and privacy constraints. In this study, we consider a localized strategy based on the German counties and communities managed by the related local health authorities (LHA). For the preservation of privacy to not oppose the availability of detailed situational data, we propose a privacy-preserving forecasting method that can assist public health experts and decision makers. ML methods with federated learning (FL) train a shared model without centralizing raw data. Considering the counties, communities or LHAs as clients and finding a balance between utility and privacy, we study a FL framework with client-level differential privacy (DP). We train a shared multilayer perceptron on sliding windows of recent case counts to forecast the number of cases in the future, while clients exchange only norm-clipped updates and the server aggregates updates with DP noise. We evaluate the approach on COVID-19 data on county-level during two phases: November 2020 and March 2022 (Omicron). As expected, very strict privacy ([Formula: see text]) yields unstable, unusable forecasts. At a moderately strong but still privacy-preserving level ([Formula: see text]), the DP model closely approaches the non-DP model: [Formula: see text] (vs. 0.96) and mean absolute percentage error (MAPE) [Formula: see text] in November 2020; [Formula: see text] (vs. 0.90) and MAPE [Formula: see text] in March 2022. Overall, our results support the feasibility of privacy-preserving collaboration among health authorities for local forecasting. In the evaluated COVID-19 phases, client-level DP-FL delivered useful county-level predictions with formal privacy guarantees under the stated threat model. The appropriate privacy budget should nevertheless be re-evaluated for other epidemic phases and applications.

References

1. 1.Kerkouche, R., Zunker, H., Fritz, M., & Kühn, M. J. (2026). Differentially private federated learning for localized control of infectious disease dynamics. Scientific Reports. https://doi.org/10.1136/bmjopen-2021-055630