SPATIO-TEMPORAL GAUSSIAN PROCESS REGRESSION (STGPR) FOR RELATIVE RISK OF COVID-19 MODELING

• Main Author: Widyawati, Erni
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85833
• Institution: Institut Teknologi Bandung

The relative risk of an infectious disease, such as COVID-19, is a crucial aspect of disease mapping due to its significant role in public health. The spread of COVID-19 varies across both space and time. By understanding the relative risk, it becomes possible to predict the number of new COVID-19 cases, enabling early detection of areas with increasing risk. Consequently, preventive measures can be optimized, and resources can be efficiently allocated to minimize the impact of the disease. The relative risk of COVID-19 spread is represented by the intensity of a Poisson process, modeled using Spatio-Temporal Gaussian Process Regression (STGPR) through a logarithmic transformation. This model is constructed to integrate the spatial and temporal aspects of the available dataset while capturing the similarity between observations through the appropriate selection of kernels. Four STGPR models with different kernel structures are proposed: ARD RBF, ARD Matern 3/2, ARD RBF + ARD Matern 3/2, and ARD RBF x ARD Matern 3/2. Due to the Poisson-distributed nature of the observations y, direct inference is not feasible, necessitating the use of approximations, such as Laplace approximation. The best model is selected based on the lowest MAPE and WMAPE values, which indicate the model's performance in making predictions. From the implementation of the four models, the ARD RBF x ARD Matern 3/2 structure emerged as the best model.