A High-Granularity, Machine Learning Informed Spatial Predictive Model for Epidemic Monitoring: The Case of COVID-19 in Lombardy Region, Italy

Featured Application: This work enables real-time and high-resolution short-term forecasting of epidemic trends at the municipal level by leveraging EMS dispatch data and spatial modeling. By predicting the expected evolution of disease-related calls one week in advance, it supports health authorities in anticipating emerging outbreaks and planning timely, localized interventions, even in contexts where official diagnoses may be delayed or incomplete. This study aimed at proposing a predictive model for real-time monitoring of epidemic dynamics at the municipal scale in Lombardy region, in northern Italy, leveraging Emergency Medical Services (EMS) dispatch data and Geographic Information Systems (GIS) methodologies. Unlike traditional epidemiological models that rely on official diagnoses and offer limited spatial granularity, our approach uses EMS call data (rapidly collected, geo-referenced, and unbiased by institutional delays) as an early proxy for outbreak detection. The model integrates spatial filtering and machine learning (random forest classifier) to categorize municipalities into five epidemic scenarios: from no diffusion to active spread with increasing trends. Developed in collaboration with the Lombardy EMS agency (AREU), the system is designed for operational applicability, emphasizing simplicity, speed, and interpretability. Despite the complexity of the phenomenon and the use of a five-class output, the model shows promising predictive capacity, particularly for identifying outbreak-free areas. Performance is affected by changing epidemic dynamics, such as those induced by widespread vaccination, yet remains informative for early warning. The framework supports health decision-makers with timely, localized insights, offering a scalable tool for epidemic preparedness and response.