During an outbreak, rapid responses are required to contain the spread of the epidemic. In response to this response, a localized approach has several advantages, limiting the required resources and reducing the impact of large-scale interventions. However, training a separate machine learning (ML) model at local scale is often infeasible due to limited data availability. Data is also difficult to centralize due to its high sensitivity and privacy constraints. This study considers a regionalized strategy based on German counties and communities managed by relevant local health authorities (LHAs). We propose a privacy-preserving predictive method that can assist public health experts and decision-makers in not opposing the availability of detailed situational data to protect privacy. ML methods using federated learning (FL) train shared models without centrally managing the raw data. Considering counties, communities, or LHAs as clients, we study an FL framework with client-level differential privacy (DP) to find a balance between utility and privacy. We train a shared multilayer perceptron on a sliding window of recent case counts to predict future case counts, but the clients only exchange norm-clipped updates and the server aggregates updates with DP noise. We evaluated our approach to county-level COVID-19 data (Omicron) in two phases: November 2020 and March 2022. As expected, privacy is very strict (\(\Valepsilon \le 0.5\)) produces unstable and unusable predictions. Moderate strength, but still a privacy-protecting level (\(\valepsilon = 2\)), the DP model is very close to the non-DP model. \(R^2\about 0.93\) (vs. 0.96) and mean absolute percentage error (MAPE) \(\approx. 26\%\) November 2020. \(R^2\about 0.85\) (vs. 0.90) and MAPE \(\approx. 24\%\) Overall, our results support the feasibility of privacy-preserving cooperation between health authorities in regional forecasting. During the COVID-19 assessment phase, client-level DP-FL provided useful county-level predictions with formal privacy guarantees based on a defined threat model. Still, appropriate privacy budgets should be reevaluated depending on other infectious disease stages and applications.
