Leveraging GeoAI and agent-based modeling to address infectious diseases : a case study of COVID-19 in Hong Kong

Abstract

The COVID-19 pandemic has damaged the global economy and public health. While various predictive models have been devised by researchers worldwide, a notable gap remains in establishing long-term and community-oriented models for policymaking. To fill this gap, this study demonstrates an approach in developing several community-oriented models to aid policymakers in deploying anti-epidemic measures for short-term and long-term purposes. The upward trajectory of urbanization heightens the susceptibility to infectious diseases. This trend is attributed to the concentration of populations in certain areas, resulting in elevated population densities and heightened rates of close contact, thereby amplifying the risk of infection. Community-oriented models are emerging as a prevailing approach to scrutinize the epidemic dynamics across different communities within urban settings. Hong Kong is chosen as the focal point in this study due to its diverse communities characterized by high population density. This study aims to achieve three research objectives: RO1) To examine the geographical characteristics of the high-risk location where the confirmed patients always visited during COVID-19 Omicron; RO2) To illustrate the spatiotemporal diffusion patterns of COVID-19 Omicron and the special characteristics of its trajectory. RO3) To project the epidemic dynamics under the scenarios with various controversial interventions, including Compulsory Universal Test and lockdown, by using 3-dimenstional Agent-Based Modeling. For RO1, Self-Organizing Map (SOM) was developed to identify the characteristics of high-risk communities, including sociodemographic, built environment, and air quality factors. For the analysis in sociodemographic factors and built environment factors, a Kruskal-Wallis post-hoc Dunn’s test was conducted on the clustering results. The test results can highlight the characteristics of high-risk communities. For the analysis of air quality factors with temporal features, SOM was employed with Dynamic Time Warping distance measure. Given the numerous clusters generated, hierarchical agglomerative clustering was then employed to classify these initial clusters. For RO2, SOM was conducted to categorize the temporal virus diffusion pattern. Incorporate with Sammon projection, the outcomes of SOM transferred to directed topological structure. Special diffusion patterns (e.g., loop and repeated pattern) can be identified from that structure. For RO3, a three-dimensional Agent-Based Model was developed to simulate the Omicron transmission. Four scenarios were created: Baseline: Following the schedule of realistic measures; S1: Based on baseline, Compulsory Universal Test (CUT) implemented from 26th March to 3rd April; S2: Based on S1, a citywide lockdown from 1st March to 7th March; S3: Based on S2, the schedule of CUT changed to 8th March to 14th March. The validation of baseline scenarios was conducted. For short-term prevention, implementing lockdown measures proves most effective in reducing infection and fatality rates. Even if the optimal timing for lockdown initiation is missed, CUT can still help avert subsequent waves. Moreover, the government can prioritize lockdowns in areas experiencing persistent outbreaks, followed by those with sporadic outbreaks. For long-term prevention, decreasing CO, NO/NOX and O3 concentration is necessary. Moreover, improving land use planning to ensure ample job opportunities and essential amenities within residential areas can help decrease reliance on specific neighborhoods, mitigating overcrowding and reducing time spent in transit.