Nonlinear relationships between urban morphology, socioeconomic conditions, and infectious disease risk: evidence from COVID-19 in Tokyo

Abstract Urban infectious disease outbreaks pose critical challenges to public health in rapidly urbanizing cities. The COVID-19 pandemic provided a natural experiment to examine how area-level environmental and socioeconomic contexts are associated with infection risk over time. This study analyzed seven pandemic stages in Tokyo, Japan (2020–2022), across 53 municipalities using random forest-based interpretable machine learning models with SHapley Additive exPlanations and partial dependence plots diagnostics. Feature selection and nonlinear modeling identified key drivers among 49 candidate variables. The results revealed significant spatial clustering of COVID-19 infection rates, with persistent hotspots in central wards like Shinjuku, Minato, and Shibuya, while suburban regions of Western Tokyo maintained lower infection rates. Key built environment factors exhibited stage-specific, nonlinear, and threshold-like associations with infection rates, including road density (25 km km −2 ), FAR (0.25–0.3), and population density (400 and 600 people/km 2 ). These associations were predominantly positive beyond the identified thresholds, indicating elevated infection risk under higher density and connectivity conditions. Socio-demographic factors also showed temporal specificity: the percentage of foreigners displayed a threshold around 1.5%, and construction worker density emerged as a relevant correlate during the Omicron-dominated phase. Overall, the relative importance and marginal patterns of these associations varied across intervention stages, highlighting temporal instability in area-level risk correlates. Importantly, these findings are associational rather than causal, reflecting contextual exposure conditions at the municipal scale. From an environmental epidemiology perspective, the results suggest that stage-sensitive and spatially explicit interpretation of area-level indicators may enhance infectious disease surveillance, compared with approaches assuming temporally invariant or linear effects.