Developing Machine Learning Algorithms for Tracking Epidemics in Time and Space
The spread of an epidemic is a spatial-temporal process, which is the movement of infected cases in time and space. Thus, the temporal and spatial location of the case is the key to speculating on the transmission of infectious diseases. Prof. Tzia-Hung Wen of the Department of Geography, National Taiwan University has developed machine learning algorithms based on the same concept of typhoon track tracking to analyze the dynamics of epidemic clustering and spreading. The algorithm constructed by the lag time in the course of propagation and diffusion and the spatial distance between different cases can detect the sub-clusters in the same time and space and reveal the routes of transmission. The trend of dengue fever in southern Taiwan from 1998 to 2015 has proved the applicability of the algorithm. The research was published in international journals, including Scientific Reports 7：12565 and Annals of the American Association of Geographers 108（4）：1168-1186.
Prof. Wen suggested that the Department of Health should integrate epidemic reporting systems and algorithms as automatic early warning tools to explore the spread of dengue fever, measles, tuberculosis, and other diseases in time and space. Recently, the algorithm has been adopted by the National Mosquito-Borne Diseases Control Research Center of the National Health Research Institutes (NHRI) and the Department of Health of local government to track the outbreak of dengue fever this September.
Prof. Wen research team proposed the new classification of the evolution of disease transmission groups (Fig. 1). There are single and interaction patterns. The change of the transmission center and the area of the group are monitored in each pattern. Each change requires corresponding public health strategies. For example, if the epidemic in the transmission group is more prevalent than before, the area affected by the transmission may expand. The health authorities can thus focus on this area. If the group moves to another location, the center will move as well. The government can track the route to prevent it from spreading. Through combining the changes of these two characteristics, there are six different types of evolutionary mode for single pattern. In addition, to conclude the evolution of the interaction is to focus on the changes of the interaction between two or more transmission groups.
Fig. 2 shows that Prof. Wen’s method, MST-DBSCAN, is more intuitive than ST-KDE. MST-DBSCAN can prove the spreading of each hot-spot of ST-KDE. Hot-spot I is the earliest among the four. It is located in zone 1. Hot-spot II and III are in zone 4. Hot-spot IV is situated in zone 5 and it appeared later than II and III.