Development of health geomatics solutions for data science in emergency medical services

The objective of the proposed study is the development and implementation of health-GIS frameworks as evidence-based and data-driven support to policy makers in the field of Emergency Medical Services, identifying applications, developing relevant data processing models, and validating them with dedicated metrics. Specifically, from a methodological point of view, the focus is on spatially explicit models for regression or classification tasks, addressed with a supervised learning approach by exploiting machine learning algorithms. With a common background in the data source, mainly related to calls to the health emergency number from citizens and intervention of ambulances, the research is articulated in a series of study-cases, relevant to two fields of application, having as focus the territory of Lombardy region in Italy: the management of Public Access Defibrillation (PAD), and the disease surveillance for the pandemic of COVID-19. The first study-case is aimed at identifying and evaluating an optimized distribution of publicly accessible Automated External Defibrillators over the territory of the city of Milan, the largest city of Lombardy (1.37 million residents), crossing the spatial coverage of the current distribution with a geographic risk map to estimate the demand over the territory. The proposed strategy for AEDs deployment resulted more effective compared to the existing distribution, increasing the spatial coverage of out-of-hospital cardiac arrests from 41.77% to 73.33%. The second study-case is the first one relevant to disease surveillance of COVID-19, whose aim is to identify the beginning of anomalous trends (change in the data morphology) in emergency calls and EMS ambulances dispatches and reconstruct COVID-19 spatiotemporal evolution in Lombardy, by applying a dedicated signal processing method. Local starting days anticipated the official diagnoses and casualties, thus demonstrating how these parameters could be effectively used as early indicators for the spatiotemporal evolution of the epidemic on a certain territory. The third study-case furtherly elaborates the previous results by implementing two different models for disease surveillance: the former aiming at the generation of local early alerts, while the latter for its prediction on limited geographical areas. Both methods are based on supervised learning, trained on time-series of calls and ambulances dispatches on limited areas, up to single municipalities. To evaluate the significance of the early alert model, 5 classes of alert were computed based on the results of the machine learning, and the number of ambulances dispatches in the following 7 days was analyzed, thus revealing statistically significant differences across the classes. The predictive model was evaluated retrospectively, reaching a performance of 78.9% accuracy in predicting the curve evolution in single municipalities. Finally, a fourth study-case explores the effect of air pollutants in the first spread of COVID-19 in Lombardy, by inspecting the correlation between exposure to pollutants and the velocity of diffusion (estimated from ambulances dispatches). The correlation was separately performed on different clusters of municipalities, classified according to land-use and socioeconomic indicators, thus only comparing territories with similar anthropic characteristics. As a result, significant exponential correlations were found for ammonia (NH3) in both prevalently agricultural (R2 = 0.565) and mildly urbanized (R2 = 0.688) areas, which is a novel result compared to scientific literature. The implemented models did not propose brand new computational techniques, but multiple novelties were instead introduced from the point of view of case-specific problem solving, thus inferring new knowledge about a series of real world issues. As a matter of fact, despite a common background in the main data source exploited, the implementation of health-GIS frameworks was consistently different across all the phases of the addressed study cases. Yet, the implemented frameworks were in all study cases successful in providing relevant knowledge, showing the possibility to be applied in supporting the definition of evidence-based policies. Both the spatial component (of data and models) and the data-science approach resulted to be an added value for the resolution of the addressed tasks, showing that the proposed methodologies can and should be implemented in the real-world scenario to address the above mentioned or similar problems.