Hotspots for Emerging Epidemics: Multi-Task and Transfer Learning over Mobility Networks

Emerging epidemics such as the COVID-19 entail economic and social challenges, which require immediate attention from policymakers. An essential building block towards the implementation of mitigation policies (e.g., lockdown and testing) is the identification of potential hotspots, defined as areas that contribute significantly to the spatial diffusion of infections. This work seeks to identify these hotspots for emerging epidemics through advanced analytical methodologies, i.e., a combination of long short-term memory (LSTM) model, multi-task learning, and transfer learning. To achieve this goal, we use data on COVID-19 infections and mobility over a network of locations to illustrate the proposed method. We first identify transmission hotspots by employing the LSTM model together with multi-task learning over a network of locations. Next, to illustrate the importance of these identified hotspots in deciding on lockdown policies, we compare the transmission hotspots-based policy with a pure infection load-based policy and show the hotspots-based policy leads to a 21% improvement in reducing the predicted new infections. Finally, we improve our hotspots identification with transfer learning from past influenza transmission data. We demonstrate that the inclusion of transfer learning reduces the mean absolute error in the infection prediction by 53.4% and consequently improves the hotspots identification. On a broader note, this paper proposes an advanced data-driven approach to identify transmission hotspots, which has considerable methodological and practical implications for the current and any future pandemic if one were to occur.