Associating COVID-19 prevalence and built environment design: An explainable machine learning approach

Stay-at-home orders were globally adopted as one of the most important nonpharmaceutical interventions (NPIs) during the recent global pandemic. In a high-rise high-density context of Hong Kong, inter-building airborne transmissions were reported, especially in public housing. The role of residential building design in infection dynamics is under-studied. To unravel how architectural and urban design was linked to airborne virus transmission during the pandemic, we fitted explainable machine learning (EML) models associating COVID-19 prevalence with architectural design controlling for other built environment (BE) factors including socio-demographics, road information, land use, and points of interest (POIs). 284 public housing that underwent restriction-testing declaration (RTD) during the peak period of the pandemic's fifth wave were our sample. An additional 35 RTD-issued private housing blocks were used for an initial comparison of infection prevalence across public and private housing. Our findings show a significant differential in prevalence over different design forms, with "8-" and "L-" shaped buildings appearing to be more susceptible, with a significantly greater percentage of infections than "X-" and "Y-" shaped structures. The percentage of vacant land, public residential within a 500-m buffer, and the proportion of children ages under 14 ​at small tertiary planning unit level (STPU) were the three most influential co-variates in our model. Among specific architectural design features, the number of floors, radial layouts, and building corners were the most significantly associated with COVID-19 prevalence, followed by building average flat (apartment) size and shape factor. The study indicates that public housing residents were more at risk during this wave of the pandemic, which needs further investigation. Using machine learning, we provide insights into how to manage the design of high density neighbourhoods for resilience against airborne disease vectors. • A comprehensive explanation machine learning framework to explore the association between architectural design and building-level COVID-19 prevalence. • A rigid feature selection approach to select the most representative feature subset to improve the machine learning performance and interpretability. • “X-” and “Y-” shaped buildings were more resistant to COVD-19 transmission while “L-” and “8-” shaped buildings were the most vulnerable. • The proportion of vacant land within 500-m buffer of the building and proportion of children under 14 were the most significant factors and were both positively associated with COVID-19 prevalence.