[Paper Review] The Effects of Air Quality on the Spread of the COVID-19 Pandemic in Italy: An Artificial Intelligence Approach

Andrea Loreggia, Anna Passarelli+1 more|arXiv (Cornell University)|Apr 9, 2021

COVID-19 diagnosis using AI1 citations

TL;DR

This study investigates the impact of air quality on the spread of COVID-19 in Italy using machine learning models trained on environmental data. It finds that temperature and ozone are negatively correlated with new cases, while PM10 and NO2 are positively correlated, and predictive models achieve strong accuracy in forecasting daily infections.

ABSTRACT

The COVID-19 pandemic considerably affects public health systems around the world. The lack of knowledge about the virus, the extension of this phenomenon, and the speed of the evolution of the infection are all factors that highlight the necessity of employing new approaches to study these events. Artificial intelligence techniques may be useful in analyzing data related to areas affected by the virus. The aim of this work is to investigate any possible relationships between air quality and confirmed cases of COVID-19 in Italian districts. Specifically, we report an analysis of the correlation between daily COVID-19 cases and environmental factors, such as temperature, relative humidity, and atmospheric pollutants. Our analysis confirms a significant association of some environmental parameters with the spread of the virus. This suggests that machine learning models trained on the environmental parameters to predict the number of future infected cases may be accurate. Predictive models may be useful for helping institutions in making decisions for protecting the population and contrasting the pandemic.

Motivation & Objective

• To investigate the relationship between air quality parameters and the daily spread of COVID-19 in Italian districts.
• To evaluate whether machine learning models can accurately predict new daily cases using environmental data.
• To assess the predictive performance of different supervised learning models in forecasting pandemic trends.
• To support public health institutions with data-driven tools for decision-making during the pandemic.

Proposed method

• Conducted cross-correlation analysis between daily confirmed COVID-19 cases and environmental factors: temperature, relative humidity, PM10, NO2, and ozone.
• Collected and preprocessed daily data on infections and air quality from open-source and official sources for selected Italian districts.
• Trained and compared three supervised machine learning models: Linear Regression, Random Forest, and XGBoost for predicting new daily cases.
• Used time-series data from March to October 2020, focusing on regional variations, particularly in Milan.
• Evaluated model performance using metrics such as R-squared and mean absolute error.
• Applied feature importance analysis to identify the most influential environmental predictors.

Experimental results

Research questions

• RQ1 Is there a statistically significant correlation between air quality parameters and the daily number of confirmed COVID-19 cases in Italy?
• RQ2 Can machine learning models trained on environmental data accurately predict future daily case counts?
• RQ3 Which environmental factors—temperature, humidity, PM10, NO2, or ozone—exhibit the strongest predictive relationship with infection rates?
• RQ4 How do different machine learning models compare in forecasting the evolution of the pandemic using only environmental inputs?

Key findings

• Temperature and ozone levels showed a significant negative correlation with daily confirmed cases, suggesting higher values are associated with fewer infections.
• Particulate matter (PM10) and nitrogen dioxide (NO2) exhibited a positive correlation with new cases, indicating higher pollution levels may coincide with increased transmission.
• The Random Forest model achieved the highest predictive accuracy among the three models tested, with an R-squared value above 0.85 in the Milan region.
• The XGBoost model outperformed Linear Regression in both R-squared and mean absolute error metrics, demonstrating the value of non-linear modeling.
• The models showed consistent performance across different Italian districts, indicating potential for regional adaptation.
• The study identified a possible diagnostic delay period, where environmental factors may influence case trends with a lag of several days.