The future of spatial epidemiology in the AI era: enhancing machine learning approaches with explicit spatial structure

Spatial epidemiology, defined as the study of spatial patterns in disease burdens or health outcomes, aims to estimate disease risk or incidence by identifying geographical risk factors and populations at risk (Morrison et al., 2024). Research in spatial epidemiology relies on both conventional approaches and Machine- Learning (ML) algorithms to explore geographic patterns of diseases and identify influential factors (Pfeiffer & Stevens, 2015). Traditional spatial techniques, including spatial autocorrelation using global Moran’s I, Geary’s C (Amgalan et al., 2022), and Ripley’s K Function (Kan et al., 2022), Local Indicators of Spatial Association (LISA) (Sansuk et al., 2023), hotspot analysis by Getis-Ord Gi* (Lun et al., 2022), spatial lag models (Rey & Franklin, 2022), and Geographically Weighted Regression (GWR) (Kiani et al., 2024) are designed to explicitly incorporate the spatial structure of data into spatial modelling, often referred to as spatially aware models (Reich et al., 2021). Beyond these models, several other spatially aware approaches that have been widely applied in epidemiological studies include but are not limited to Bayesian spatial models that account for spatial uncertainty in disease mapping, such as Bayesian Hierarchical models, Conditional Autoregressive (CAR), and Besage, York, and Mollie’ (BYM) models (Louzada et al., 2021). Bayesian methods are statistically rigorous techniques that assume neighboring regions share similar values. Kulldorff’s Spatial Scan Statistic is another traditional spatial technique that uses a moving circular window to extract significant disease clusters (Tango, 2021). Moreover, geostatistical models such as Kriging and Inverse Distance Weighting (IDW) allow for continuous spatial interpolation of health data (Nayak et al., 2021). [...]