Controlling recurrent infectious diseases is a vital yet complicated problem. A large portion of the controlling epidemic relies on patients visit clinics voluntarily. However, they may already transmit the disease to their contacts by the time they feel sick enough to visit the clinic, especially for conditions with a long incubation period. Therefore, active screening/case finding was deployed to provide a powerful yet expensive means to control disease spread in recent years. To make active screening success a given limit budget, one of the challenges that need to be addressed is that we do not know the exact state of each patient. Given the number of horizon and budget we have in each time step, we also need to plan our screening efficiently and screening the vital patients in time. Thus, we apply a reinforcement learning approach to solve active screening problems on the network SIS disease model. The first contribution of this work is that we identify three significant challenges in active screening problems: partially observable states, combinatorial action choice, high-dimensional state-action space. We further propose the corresponding solutions to overcome these challenges. Specifically, we resolve the issue of high-dimensional state-action space by encoding the actions and partially observable states into a lower dimension form, which is done by either manually, using domain expertise, or automatically, using the state of the art GCN approach. We show that our approach can scale up to large graphs and perform decently compared to other baselines of previous literature and current practice.
展开▼
