Behavior-driven forecasts of neighborhood-level COVID-19 spread in New York City

The COVID-19 pandemic in New York City (NYC) was characterized by marked disparities in disease burdens across neighborhoods. Accurate neighborhood-level forecasts are critical for planning more equitable resource allocation to reduce health inequalities; however, such spatially high-resolution forecasts remain scarce in operational use. In this study, we analyze aggregated foot traffic data derived from mobile devices to measure the connectivity among 42 NYC neighborhoods driven by various human activities such as dining, shopping, and entertainment. Using real-world time-varying contact patterns in different place categories, we develop a parsimonious behavior-driven epidemic model that incorporates population mixing, indoor crowdedness, dwell time, and seasonality of virus transmissibility. We fit this model to neighborhood-level COVID-19 case data in NYC and further couple this model with a data assimilation algorithm to generate short-term forecasts of neighborhood-level COVID-19 cases in 2020. We find differential contact patterns and connectivity between neighborhoods driven by different human activities. The behavior-driven model supports accurate modeling of neighborhood-level SARS-CoV-2 transmission throughout 2020. In the best-fitting model, we estimate that the force of infection (FOI) in indoor settings increases sublinearly with crowdedness and dwell time. Retrospective forecasting demonstrates that this behavior-driven model generates improved short-term forecasts in NYC neighborhoods compared to several baseline models. Our findings indicate that aggregated foot-traffic data for routine human activities can support neighborhood-level COVID-19 forecasts in NYC. This behavior-driven model may be adapted for use with other respiratory pathogens sharing similar transmission routes.Author summary: A fundamental question in infectious disease modeling is whether the inclusion of more detailed processes results in more precise epidemic simulation, and to what extent system granularity is needed to inform real-world application of model outcomes. Here, we investigate the utility of foot traffic data, which capture mobility patterns during various human activities, in neighborhood-level disease modeling. Our results indicate that foot traffic data aggregated for place categories, combined with representation of the seasonality of SARS-CoV-2 transmissibility, can support forecasting of the heterogeneous COVID-19 spread across 42 NYC neighborhoods in 2020. We propose a parsimonious behavior-driven epidemic model that generates improved short-term forecasts at the neighborhood level. We also estimate that the force of infection (FOI) in indoor settings increases sublinearly with crowdedness and dwell time. Incorporating the differential FOIs in different place categories improves short-term COVID-19 forecasts. With proper modifications, this model could be applied to other respiratory diseases.