Beyond Mobility: Socioeconomic Context Shapes the Dynamics and Hidden States of COVID-19 Transmission.

We present an integrative framework to model and simulate the spatiotemporal progression of COVID-19 across urban communes using latent regime dynamics. Leveraging non-homogeneous Hidden Markov Models (nHMMs), we identify hidden epidemic states driven by internal and external mobility patterns and modulated by sociodemographic and structural covariates. Linear mixed-effects models reveal substantial inter-communal heterogeneity in transition dynamics, with urban quality and structural inequality critically shaping epidemic persistence and regression. To assess the generative capacity of the inferred latent structure, we implement two simulation strategies: (i) a hybrid Copula–Generalized Pareto Distribution model to capture non-Gaussian dependency and tail risks, and (ii) a Conditional Variational Autoencoder with LSTM (CVAE-LSTM), conditioned on latent states and commune-specific features. Severity-weighted data augmentation is introduced to mitigate underrepresentation of critical regimes, substantially improving the reproduction of epidemic peaks. Results confirm that latent states capture high-level patterns of disease propagation and that their inclusion enhances the fidelity of simulations across contrasting urban contexts. Our findings emphasize the limitations of mobility-only models and highlight the necessity of structurally-informed, phase-aware strategies to interpret and forecast infectious disease dynamics. The framework is generalizable to other epidemics and supports context-sensitive, subnational policy design in regions with marked spatial inequities.