key: cord-0799982-13itcjqb
authors: Magazzino, Cosimo; Mele, Marco; Schneider, Nicolas
title: The relationship between air pollution and COVID-19-related deaths: An application to three French cities
date: 2020-09-12
journal: Appl Energy
DOI: 10.1016/j.apenergy.2020.115835
sha: 6b0bd7171a7fcab74d1596659d90a85e9adda234
doc_id: 799982
cord_uid: 13itcjqb

Being heavily dependent to oil products (mainly gasoline and diesel), the French transport sector is the main emitter of Particulate Matter (PMs) whose critical levels induce harmful health effects for urban inhabitants. We selected three major French cities (Paris, Lyon, and Marseille) to investigate the relationship between the Coronavirus Disease 19 (COVID-19) outbreak and air pollution. Using Artificial Neural Networks (ANNs) experiments, we have determined the concentration of PM(2.5) and PM(10) linked to COVID-19-related deaths. Our focus is on the potential effects of Particulate Matter (PM) in spreading the epidemic. The underlying hypothesis is that a pre-determined particulate concentration can foster COVID-19 and make the respiratory system more susceptible to this infection. The empirical strategy used an innovative Machine Learning (ML) methodology. In particular, through the so-called cutting technique in ANNs, we found new threshold levels of PM(2.5) and PM(10) connected to COVID-19: 17.4 µg/m(3) (PM(2.5)) and 29.6 µg/m(3) (PM(10)) for Paris; 15.6 µg/m(3) (PM(2.5)) and 20.6 µg/m(3) (PM(10)) for Lyon; 14.3 µg/m(3) (PM(2.5)) and 22.04 µg/m(3) (PM(10)) for Marseille. Interestingly, all the threshold values identified by the ANNs are higher than the limits imposed by the European Parliament. Finally, a Causal Direction from Dependency (D2C) algorithm is applied to check the consistency of our findings.

As one of the most urgent threats on the planet, the World Health Organization (WHO) declared as a global health emergency on March 12 th 2020 (WHO, 2020a) . At the world level and as of April 20 th 2020, the WHO reported 157,847 confirmed deaths and 2,314,621 confirmed cases (WHO, 2020b) . This virus spreads through three channels: saliva, nasal discharge, or airborne particles (Mitra et al., 2020) . Even though most people infected recover without relying on advanced treatment, elderly and sensitive people present an important risk to develop serious and deadly illness. The most effective way to prevent and reduce transmission worldwide is by washing hands and using alcohol-based hand sanitizer frequently. Indeed, experts are unanimous on the necessity to control and lower contact among the population, not only to protect uncontaminated individuals, but also to isolate the bearer of the virus. Most of the governments have understood that the quarantine strategy is necessary to keep the pathogen dynamic under control (Mitra et al., 2020) .

As the epicentre of the pandemic, China was the first country to completely shut-down commercial activities, restrict domestic and international travel, and impose a containment system on the population . Following that, similar policy measures were adopted by numerous countries, notably France. Beginning on March 17 th 2020, the national containment measure decided by the French President, Emmanuel Macron, involved: closing of schools, colleges and universities; shutting down non-essential companies and sending workers home, restriction of public transport operations; forbidding gatherings and meetings in public spaces; requiring individuals to remain in their residence, aside from going out for necessities (Elysée, 2020) . This state of sanitary emergency is not without major consequences.

It is expected to seriously impact the world economy 1 , without sparing the French economic growth. Numerous industries are impacted, especially tourism, catering, apparel, culture, aeronautic, and automotive sectors. Although there is no possible comparison, looking at the COVID-19-related deaths recording gives a broad picture of the health consequences of that virus in France. Just before the mandatory containment measures (i.e., March 12 th 2020), the WHO (2020c) reported 48 confirmed deaths and 2,269 confirmed cases. The Situation Report on COVID-19 published by the WHO (2020b) indicated 22,821 confirmed deaths and 123,279 confirmed cases, as of April 27 th 2020.

Economic activity is a well-known factor contributing to global environmental pollution because fuel combustion directly releases Greenhouse Gas (GHG) emissions into the atmosphere (Intergovernmental Panel on Climate Change (IPCC), 2007). In the most densely populated cities, air pollution recorded high concentrations measures up to the point where the physical health of the population is now directly threatened (Van Mierlo et al., 2006; Anjum, 2020) . Nowadays, primary pollutants 2 (methane, CH 4 ; carbon monoxide, CO; nitrogen dioxide, NO 2 ; sulphur oxides, SO x ) and secondary pollutants (nitrogen dioxide, NO 2 ; ozone, O 3 ; sulphur trioxide, SO 3 ) display systematically significant levels in urban areas. But another environmental concern is drawing the attention from researchers: the critical concentration of small size components (PM 10 and PM 2.5 ) 3 in major cities. Factories and house heating are important driver of these particles. However, in major European cities, transport is widely recognised to be the main source of this local pollution inducing specific health effects for urban inhabitants (Pant and Harrison, 2013) . Far from being a local concern, mega-cities in the developing world are also facing identical challenges.

The origin of PMs emissions come from road traffic as it relies almost exclusively on oil products (gasoline and diesel) (Lozhkina et al., 2016) . In urbanized areas, these particles have recorded a constant growth over the past decades with detrimental impacts on human health (Proost and Van Dender, 2012; Rissler et al., 2012) . Tightly linked to this view, it raises the question of the traffic congestion in big cities: beside inducing longer transport durations, it increases fuel consumption and carbon emissions 4 , while not allowing efficient travels (Barth and Boriboonsomsin, 2008; Franceschetti et al., 2013) . Consequently, much attention has been focused on planning urban areas capable of properly lowering traffic congestion (i.e., public transport, incentives for bike users and private vehicles restrictions) as well as introducing lowcarbon energy sources with direct applications to transport mobility (Van Mierlo et al., 2006; Liaquat et al., 2010; Zhang et al., 2010) . Nonetheless, cleaner alternative fuels insignificantly contribute to the energy supply, and appear domestically unavailable in most of the urban areas.

2 Primary pollutants refer to any type of pollutant that are directly emitted from a single source into the air including many sources including vehicles, coal-fired power plants, natural gas power plants, and biomass burning. They differ from secondary pollutants which are instead formed when two or more primary pollutants react with each other in the atmosphere. 3 PM 10 and PM 2.5 are ultrafine harmful particles pollution whose diameter is inferior to 0.1 and 0.025 mm, respectively. 4 As an illustration, the International Road Transport Union (IRU, 2012) estimated that more than 100 billion litres of wasted fuel (corresponding to 250 billion tonnes of CO 2 equivalent) were attributed to traffic congestion in the American cities for the year 2004. Far from being unique, European and (especially) French cities also experience important congestion issues. Accordingly, transport needs on fossil fuels have not decreased over time 5 . Thus, climate change issues are still burning as oil products dependence for road traffic remains critical (Proost and Van Dender, 2012) . Dealing with direct exhaust congestion and pulmonary health diseases, the municipalities are now questioning the nature and the way fossil energy is consumed by road transport. To get a broad overview on the needs of petroleum-based fuels for combustion, one must look at the growth trend in car ownership. In 2010, 700 million light duty vehicles, automobiles, light trucks, SUVs (Sport Utility Vehicles) and minivans, were on roadways in the world. By 2030, these numbers are projected to increase 1.3 billion, and by 2050, it may reach 2 billion vehicles (Balat and Balat, 2009) . Controlling for the population size, car ownership rates reached 0.48 cars per capita in Europe in 2010, which is strongly higher than China (with 0.03 cars per capita) but remains lower than the US (0.76 cars per capita) for the same year (European Union Publication Office (EUPO), 2012). Being the core of PMs emissions in major French cities, the consumption of imported oil products for transport recorded a drastic rise since 1970's. Thus, one key sustainability challenge emerges: how can we allow efficient transportation services within cities with acceptable environmental externalities? To address the PMs problem at its major source, the French government committed to the ambitious goal of enhancing the share of renewable energy in the transport sector to 15% by 2030, in which biofuels will take a major share (Baudry et al., 2018) 6 .

However, an important limit remains: when world agricultural prices rose (as it was the case in the aftermath of the 2008 crisis), biofuels are blamed for the threat they can potentially create for food security, and are often pushed to the background (Doumax et al., 2014) . Meanwhile, France is experiencing a burning debate relative to the taxation of fossil fuels (Doumax et al., 2014) . Indeed, diesel benefits from a lower tax rate compared to gasoline, explaining why diesel engines are predominant in the French car fleet. Facing the need to internalize its environmental and health externalities, the French Cour des Comptes (2012) suggested the creation of new incentives towards diesel consumption reduction. Presented as controversial, this project has nonetheless induced large-scale demonstrations across the territory, worsening the public acceptance of future fuel tax reforms. 5 According to British Petroleum (BP, 2012) , the share of biofuels in world liquid fuel is expected to not exceed 4% by 2030. 6 Earlier, the 2009 Renewable Energy Directive (RED) had set the share of biofuels in the energy supply mix for road transportation should reach 10% by 2020 for each member state (European Parliament and Council, 2009). One crucial concern associated with the emission of PMs is that it induces important health risks 7 for the population including respiratory, infections, asthma, chronic obstructive pulmonary disease, and lung cancer (Kim et al., 2018; Nunez, 2019) . Because of their small sizes and light weight, it is said that these fine particles tend to longer subsist in the air than larger ones and can penetrate deep into the lungs and the circulatory system (Wang et al., 2019) . This is in line with Boldo et al. (2006) who performed a huge Health Impact Assessment (HIA) on 23 European cities and estimated that 16,926 premature deaths from all causes could be prevented if long-term exposure to PM 2.5 levels were reduced to 15 μg/m 3 in each city.

Therefore, solving the poor air quality issue has become a key target for mayors of large cities.

Recently, this issue has been growing for French cities. The Paris city council banned cars from using the roads near the Seine river in 2016. In Marseille, parking costs for non-residents increased when the city recorded pollution peaks in 2019. In Lyon, the council adopted a motion in 2020 to allow only less polluting vehicles to circulate when pollution peaks are confirmed.

However, we must admit that none of these measures are comparable to the COVID-19 related shut-down. Interestingly, one unexpected externality of the COVID-19 lockdown is the significant lowering in both primary and secondary pollutant emissions, raising questions about the well-established relationship among human activities and air quality. Indeed, major French cities have experienced remarkable drops in air pollution, notably for PMs and NO 2 emissions as these pollutants come mainly from traffic. NO 2 level has drastically fallen in March 2020 when compared to the same period in 2019.

In addition, several investigations applied on various cases confirmed that the public and domestic transportation's restriction has resulted in obvious reductions of fuel combustion. Just before the mandatory containment measures (i.e., March 12 th 2020), the WHO (2020b) reported 48 confirmed deaths and 2,269 confirmed cases. The Situation Report n. 98 on COVID-19 published by the WHO (2020c) indicated 22,821 confirmed deaths and 123,279 confirmed cases, as of April 27 th 2020. Looking at the city level, confirmed deaths recorded a dramatic increase over the March 18 th -April 27 th period: from 14 to 1,387 in Paris; from 0 to 481 in Lyon; from 4 to 381 in Marseille (French National Public Health Agency, 2020) . If the restriction of public and domestic transportation has resulted in obvious reductions of fuel combustion, several questions remain unanswered. Since there are no specific vaccines for COVID-19 for the moment, the ongoing COVID-19 crisis is currently far from being over. 7 Air pollution contributed to 9% of deaths worldwide for the year 2017 which is equivalent to 7 million premature deaths (Anjum, 2020) . Consequently, air pollution has been recognized as one of the world's important deaths drivers after high blood pressure, smoking and high blood sugar (Ritchie and Roser, 2019) .

Since it clearly appears that there are remarkable differences in terms of the rate of COVID-19 spread in the world, it would be relevant to assess the potential influence of atmospheric pollution as a contributing factor to COVID-19 mortality (Contini and Costabile, 2020) . As a response, a few seminal studies have been recently conducted on various cases with multiple pollutants types Conticini et al. (2020) and Putrino et al. (2020) on Italy). These empirical investigations confirmed the existence of a significant association between air pollution and COVID-19 cases or mortality, making poor air quality an additional co-factor of COVID-19 lethality. This finding is in line with the scientific literature highlighting that the exposure to air pollution matters for the spread of various viral infections (Chen et al., 2010; Ye et al., 2016; Chen et al., 2017; Peng et al., 2020) . The underlying hypothesis is that a pre-determined particulate concentration can foster COVID-19 and make the respiratory system more susceptible to this infection. In fact, airborne particles could serve as carrier of pathogens, making the viral infection spread more harmful (Setti et al., 2020) .

Fernandes (2020) provided two scenarios. The first hypothesizes that GDP growth takes a hit, ranging from 3-5% depending on the country. In the second one, GDP can fall as much as 10%.

The economic costs of a recession are unequally distributed. Le Quéré et al. (2020) calculated that, at their peak, CO 2 emissions in individual countries decreased by -26% on average during the COVID-19 forced confinement.

An in-depth review of the literature highlights that no study has been performed on the French case so far, even though French cities experienced a dramatic COVID-19 outbreak. In fact, confirmed deaths recorded a huge increase over the March 18 th -April 27 th period: from 14 to 1,387 in Paris; from 0 to 481 in Lyon; from 4 to 381 in Marseille (French National Public Health Agency, 2020) . To the best of our knowledge, no estimated threshold of PM 2.5 and PM 10 connected to COVID-19 fatality was found. Knowing that the ongoing COVID-19 crisis is currently far from being over, the complexity of such topic requires urgent investigations. Being a fruitful research direction, demonstrating the air pollution-COVID-19 relationship could partially explain the efficiency of national lockdown measures. In addition, identifying the atmospheric co-factors (i.e., the polluting emissions linked to fossil energy combustion by transport) enabling the COVID-19 virus to spread across the urban population would help policymakers to control its epidemic diffusion with more efficiency.

This research seeks to contribute to the literature by bringing three novelty aspects. First, this paper fills the gap in the literature and proposes the first empirical assessment on the relationship between air pollution and COVID-19-related-deaths in France. To do so, we collected different sets of unique data on Particulate Matter (PM 10 and PM 2.5 ) and COVID-19related-deaths over the largest and most recent available period (from March 18 th to April 27 th 2020 representing 41 consecutive days of observation) for three major French cities (Paris, Lyon and Marseille). Second, following the empirical strategy employed by an emerging air pollution-virus epidemic literature (Alimadadi et al., 2020; Barstugan et al., 2020; Punn et al., 2020; Randhawa et al., 2020; Tuli et al., 2020) , this study applies Artificial Neural Networks (ANNs) experiments and used a Machine Learning (ML) approach. Then, to check the consistency of our results, we perform a D2C (Causal Direction from Dependency) algorithm capable of predicting the existence of a direct causal link between two variables in a multivariate setting. Third and overall, this paper represents the first empirical estimation of threshold levels of PM 2.5 and PM 10 connected to COVID-19. Hence, our empirical findings are the only ones currently able to significantly demonstrate the concentration amount of PM 10 and PM 2.5 capable of generating the adverse effect of COVID-19. Bringing high information value for policy purposes, this study is believed to create new research opportunities connected to environmental and health issues. This paper investigates the relationship between air pollution and COVID-19 diffusion for three major French cities (Paris, Lyon and Marseille). We focus on two pollutants (PM 10 and PM 2.5 ) over the most available and recent period for all cities: from March 18 th to April 27 th 2020.

The rest of this paper is organized as follows: Section 2 gives a review of the literature.

Section 3 describes the data and the methodology employed. Section 4 presents the empirical results. Section 5 shows results' interpretation and discussion. Section 6 suggests concluding remarks and policy recommendations.

Being at the core of air pollution in major cities, the consumption of oil products (mostly gasoline and diesel) for road transport elicits important environmental and health externalities.

Because of its recent nature, the literature tackling the COVID-19-air pollution relationship is very seminal.

A first strand of studies assessed the impact of COVID-19 on environmental pollution. A second group of research considered air pollution as a determining factor of COVID-19 lethality.

This sub-section aims at presenting research works performed at different scales (international, national/provincial, and city level). First, Anjum (2020) made a broad overview on the international COVID-19 situation and its link with air pollution for major countries (China, India, France, Italy and the USA). Second, we present some recent empirical research performed by and and focusing on Chinese provinces.

Third, we outline the contribution of Mitra et al. (2020) on the specific case of Kolkata city (India).

Anjum (2020) built a global assessment on COVID-19 restrictions and air pollution enhancements. Starting with an overview on the relationship between air pollution and respiratory diseases, the author compiled major public data reporting drastic reductions of air pollution for countries affected by COVID-19 virus spread. With a special focus on major cities in China, Lombardy (Italy), France, the USA, and India, Anjum (2020) suggested that the temporary nationwide lockdowns have first resulted in obvious significant reductions in air pollutions. However, at the end of the COVID-19 crisis, one cannot omit that restoring the normal situation may reverse air pollution trends.

Since Wuhan announced lockdown on January 23 th 2020, a major part of human an economic activity has been prohibited. However, severe air pollution events continued to occur. The recent studies from Huang et al. (2020) and aim at explaining why severe air pollution were not avoided in China. To do so, they both estimated emissions reduction due to COVID-19 outbreak. analysed the variations in primary and secondary pollution emissions during the COVID-19 lockdown and underlined the link between these pollutants. Using a chemical transport modelling, they showed that haze events during the COVID-19 lockdown were driven by a global reduction of secondary pollution emissions. They Quality (CMAQ) model over the period from January 01 to February 12 th 2020. Focusing on Chinese provinces, the authors found evidence that PM 2.5 concentrations decreased by 20% over this period. They engaged in further analysis in including the meteorological factor and suggested that the unfavourable weather conditions influenced the simulation results. Both studies highlight an interesting point: even though the COVID-19 lockdown produced large primary and secondary emissions reduction, this are temporary enhancements and it would not avoid severe air pollution degradation on the long-run in China. Therefore, there is a large room for improvements.

Focusing on the city of Kolkata (India), Mitra et al. (2020) compared the atmospheric CO 2 levels between April 2020 (lockdown phase) and April 2019 (pre-COVID-19 phase). Using data taken from 12 different locations, the authors observed significant variation of CO 2 levels between periods but no change between sites. Thus, as industries and transports represent the main determinants of CO 2 emissions, Mitra et al. (2020) interpreted this result as the direct lockdown effect due to COVID-19.

This sub-section displays the relevant studies tackling the atmospheric determinants of population's vulnerability to COVID-19.

Wu et al. (2020) Pansini and Fornacca (2020) investigated the geographical character of COVID-19 and its correlation with several annual satellite and ground indexes of air quality in China, Iran, Italy, Spain, France, Germany, UK, and the USA. They registered more viral infections in the areas afflicted by high PM 2.5 and NO 2 values. Higher mortality was also correlated with relatively poor air quality. Travaglio et al. (2020) showed that levels of multiple poor air quality markers, including nitrogen oxides (NOx) and sulphur dioxide (SO 2 ), are associated with increased numbers of COVID-19-related deaths throughout England, after an adjustment for population density.

Working on the Italian case, Setti et al. (2020) highlighted the association between the number of PM 10 daily limit value exceedances, registered in the period 10 th February-29 th February 2020, and the number of COVID-19 infected persons updated at 3 rd March 2020.

Empirical findings confirmed the existence of a direct relationship between the number of persons infected by COVID-19 and the PM 10 concentration levels in specific areas of Italy.

Focusing on Northern Italy, Conticini et al. (2020) examined the correlation between the atmospheric pollution and the high level of COVID-19 lethality. Results provided evidence that people living in an area with high levels of pollutants are more susceptible to develop chronic respiratory conditions and suitable to infective agents. Furthermore, air pollution was suggested to partially explain the differences in mortality in this region.

Nonetheless, these studies unanimously call for a careful interpretation of the results. As mentioned in Contini and Costabile (2020) , misleading estimations are plausible mainly due to the different strategies used for counting deaths related to COVID-19 in the world. Table 1 summarizes the main information of this seminal literature. Accordingly, our paper finds its empirical contribution by providing pioneer threshold levels estimations of PM 2.5 and PM 10 connected to COVID-19 on major French cities (Paris, Marseille, and Lyon). Another novelty aspect is methodological. We refer to the emerging literature on the relationship among air pollution and virus epidemic diffusion (Alimadadi et al., 2020; Barstugan et al., 2020; Punn et al., 2020; Randhawa et al., 2020; and Tuli et al., 2020) to rely on a ML methodology. Given the complexity of our research question, we apply ANNs experiments with univariate input and output. Then, to check the deep consistency of our results, a D2C Causal Direction from Dependency algorithm capable of predicting the existence of a direct causal link between two variables in a multivariate setting is performed. Though their limitations, our results are thought to represent a fruitful research direction whose direct implications would help policymakers.

The consumption of oil products (mainly gasoline and diesel) for road transport is at the source of an important concentration of PM in urban areas. Our focus is on the potential effects of these harmful particles in spreading the current COVID-19 epidemic. The underlying hypothesis is that a pre-determined particulate concentration can foster COVID-19 and make the respiratory system more susceptible to this infection. Hence, to empirically assess this relationship, we rely on daily data at city level. 

After collecting the data and storing it in a database from which a well-defined structure can be extracted, we proceed with the search for the optimal ML algorithm. These techniques are subsequently adopted using the measurements collected in the previous phase. Our goal is to be able to train a machine so that it learns to perform certain operations in order to obtain the desired results without explicitly providing it with the necessary algorithm. The machine must be able to determine the links present in a dataset by developing a specific behaviour. This learned behaviour can then be used to make timely forecasts and estimates on new data.

The particular complexity and non-linearity of the relationships associated with air pollution suggest the use of a more flexible tool than usual. Boznar et al. (1993) , Cigizoglu and Kisi (2005) , Pasero and Mesin (2010) Magazzino (2020) . We use a series of neurons grouped in layers. Each layer receives inputs from the previous layer and supplies the outputs to the next layer. The initial layer, connected to the input data, is called the input layer, whilst the final layer, which provides the network output, is called the output layer. Among them, there can be various hidden layers, which can increase the complexity and potential of the network.

In our NN experiments, we define non-linear activation functions for each neuron. This function is applied to the output of the neuron before it is passed onto the next neuron. Usually, the same activation function is applied to neurons of the same layer. These functions are chosen according to needs, and allow us to model a non-linearity within the network. By superimposing a series of non-linear layers, we obtain an extremely sophisticated model capable of understanding even very complicated relationships. During the training phase, to correct the weights, we use a back-propagation technique. Starting from the output nodes, we use the various gradients for each layer, updating the weights accordingly, until reaching the input layer. Therefore, keeping in mind the characteristics of the available data and the NN estimation technologies, we outline a path that hopefully produces tangible results. The main objective is to define a model capable of carrying out predictive analysis of the quantity of PM 10 and PM 2.5 that assists the spread of COVID-19. To carry out the experiments on the three selected French cities, we build ANNs, which are defined through a series of experimental tests. The training of the network takes place by providing it with a series of precise indications regarding the type of algorithm and recommended training. In particular, the difficulty in framing the best architecture of the NN is to find the balance between "Scaling Model" and "Unscaling Method".

We therefore choose the combinations that minimizes errors concerning the number of layers that are activated by hyperbolic activation functions. Next, we carry out tests to optimize the algorithm that the machine has chosen to use. Through these tests, we analyze how long it takes (epochs) for the predictive error of estimate to decrease until it reaches a value close to zero.

Our analysis proceeds through the performance of the "selections order" about the neurons used by the NN and we verify the distribution of data on the predictive regression line. Finally, we use the so-called cutting technique. In other words, we try to estimate a precise point in the neural transmission from the inputs (PM 10 and PM 2.5 levels) to the target (number of deaths in the three French cities due to . This identified point will describe the predictive concentrations of particulate matter, which leads to an increase in the number of deaths due to COVID-19. 

In this section, we show the results obtained for the French cities of Paris, Lyon, Marseille. The ANNs architecture used for each city is always the same (Figure 1 ). We expanded the data to allow the machine to be able to use a more efficient combination of information. We generated the logarithm of each variable (ln) and the first difference (d). Therefore, the inputs are PM 2.5 , PM 10 as well as their transforms. The target is represented by the Deaths variable, which represents the number of deaths in cities due to COVID-19.

Source: our processing on command strings.

Since the architecture of the NN is the same for all the cities, the descriptive neural statistics are described below.

As shown by the variables bars chart, the variables used are 9, of which six represent the input process, and one is the generated target. The instances pie chart reveals that the instances of the ML process were equal to 41. Those representing the training are 25 (60.9%). This result underlines how, compared to a choice of n projects, our model chose 25 out of 41 potential models. These are the ones that best suit the target. The result confirms that our choice is appropriate. The selection requests were eight (19%). The instances, therefore, selected the best possible NN process concerning the generated target. This result allows us to continue the processing. As for the testing instances, they were eight (19%). This value represents the result of the choice of numerous training models. Since it is the same and never less than the selection instances, this reinforces the previous findings. Finally, the number of unused instances is zero (0%). This result also confirms the effectiveness of our model. No anomalous values (which would have invalidated the results) were generated.

To facilitate the reading of the results obtained and ensure their replication by the scientific community, we have inserted the tests on NN in the Appendix. Below we show our estimation experiment on the concentration level of PM 2.5 and PM 10 connected to the deaths during the COVID-19 epidemic in the three French cities. The system used is the "Plot Directional

Output". It is very useful to see how the outputs vary as a function of a single input, when all the others are fixed. This can be seen as the cut of the NN model along some input direction and through some reference point. We want to remind all readers that before proceeding with this experiment, we performed the NN through the "Perform Order Selection" and the "Perform Input Selection". Paris. This test represents an answer to a question. The algorithm asks the machine, for each variable, its weight on another variable (Deaths). The question has an answer in a range of 0-1, which is then processed in percentage terms. Compared to PM 2.5 , PM 10 represents the variable whose threshold value is more strongly correlated with deaths from COVID-19. We can say that concentrations of 29.6 µg/m 3 of PM 10 would be ideal for the spread of COVID-19. In addition, the adverse health effects caused by this type of particulate matter would aggravate COVID-19 disease 5.

-Lyon 1415 The results on the relationship between PM 2.5 concentration and deaths from COVID-19 are interesting ( Figure 4) . We can see that the highest number of deaths is reached at 15.6 µg/m 3 .

Since the recorded concentration of PM 2.5 on 27 April 2020 was 8.5 µg/m 3 , our result is predictive. In other words, it is necessary to record values lower than those obtained by us to mitigate the number of Coronavirus deaths compared to the amount of PM 2.5 . In fact, if a PM 2.5 level had been maintained at 15.6 µg/m 3 , it is likely that the number of deaths would have been greater than the recorded total of 481. In Figure 5 , we analyze the effect compared to PM 10 . Therefore, we use an algorithm in ML that can generate causal effects between the variables.

Following Pearl (1988 Pearl ( , 2000 Pearl ( , 2001 Pearl ( , 2015 , we use a D2C algorithm on Proportion-based causality using the Oryx 2.0.8 software and the same dataset used for ANNs analysis. However, since an algorithm in ML needs many variables (remembering that the data is not interpreted as a time series), we will carry out mathematical transformations. Thus, in addition to the logarithmic transformation and the first difference already carried out in the previous analysis,

we generate the square of the selected variables, and the first difference (calculated in logarithmic terms).

We perform a Machine Learning analysis, following the process shown in Figure 8 . Starting from the dataset used for ANNs analysis, we increase the variables through mathematical transformations to obtain a large dataset necessary for our D2C algorithm.

Subsequently, the causality model is processed, and we analyze those variables significant for us. Once the D2C commands are imported into the Oryx software, the analysis generates the causalities mentioned above typical of a ML process. Finally, we carry out the ML test (AC and AUPRC) to verify the correctness of the algorithm.

In Tables 3-5 , we show the causality and significance tests to determine the relationship between the variables object of the study on Paris, Lyon, and Marseille. In the model, n filtered factors were used. The self-learning machine worked in the following way. It started from a set of commands with functionality still to be preset. Subsequently, ten classifiers were trained and tested to achieve the predictive causal link between variables. These ten classifiers worked through a binary calculation sequence, alternating the values [0] with [1]. As we can see from the results, the algorithm worked by performing on average over 18,500

repetitions for each combination of causality between our variables. The closing percentage of the calculation, within the average of the repeats, has always been higher than 80%. Hence, our algorithm has ever completed each cycle for each pair of variables. The value of the Average Causality is uniform for all the pairs. As regards the significance of the results of predictive causality, we parameterized the AUPRC. It was divided into True or False with respect to a P-Value lower or higher than 5%. We obtained that only a causal relationship is significant within the AUPRC analysis. It is attributable to a unidirectional causality running from PM 2.5 to Deaths, PM 10 to Deaths, dPM 2.5 to dDeaths, and dPM 10 to dDeaths. These results clearly confirm the choice of the variables of the ANNs model. Besides, we have to underline that the causality predictive relationship is also present in the variation over time between the particulates and the COVID-19 deaths. This result would confirm the hypothesis that the threshold values found in the ANNs could influence the continuation of the pandemic and deaths. Although there may be many causes of deaths, the link of predictive causality has recorded a solid relationship between pollutants and deaths from COVID-19. We believe that, at least for a value higher than 80% (percentage in the tables), high levels of fine particulate matter have caused the aggravation of COVID-19 patients, generating death. As we can see from Table 7 , all our threshold values are lower than those of the EU. These findings are important in a COVID-19 pandemic situation. EU limit values for particulate matter (PM 10 and PM 2.5 ) are excessively high. They are, on average, more significant than our threshold value for a value of 14.4 µg/m 3 (PM 10 ) and 9.2 µg/m 3 (PM 2.5 ). This result suggests a EU economic policy capable of reducing the limit values of emissions from fine particles. These limit values should respect our threshold value.

The relationship between our results and deaths from COVID-19 can be interpreted in the following way. The most likely explanation is that the levels of PM 10 and PM 2.5 found in our study generate an inflammatory response in the lungs. However, key molecular events in response to PM exposure are involved in altering the homeostasis of cardiovascular physiology.

COVID-19 would seem to support a similar mechanism, inducing the rapid onset of a state of inflammation, with an equally rapid increase in inflammatory cytokines, comparable to that caused by short-term exposure to PM. Another interpretation of our results would confirm the hypothesis that particulate matter acts as a "carrier" in transporting the virus which coagulates on the surface of the particles over a longer distance. Particulates are at least a dozen times larger in diameter than the virus. This hypothesis, already advanced in the literature for some time on specific cases, would imply that the spread of the virus is facilitated, not by smog in general, but by fine particulates. Dominici et al. (2020) in a pre-review, found a correlation between fine dust pollution and coronavirus mortality. The increase of just one microgram per cubic meter of PM 2.5 would correspond to a 15% increase in the mortality rate due to the SARS-COVID-19 virus. According to the authors, the results obtained are statistically significant and robust, with a confidence interval of 95%. In a working paper, Becchetti et al. (2020) showed a link between the COVID-19 lethality index and air quality, regarding predisposition to pulmonary pathologies. Ogen (2020) analysed the relationship between NO 2 and COVID-19.

He reviewed data from the ESA Sentinel 5P satellite and mapped the distribution of nitrogen dioxide in Europe in the months leading up to the pandemic. The results showed that 78% of the deaths from COVID-19 were concentrated in five areas located in northern Italy and central Spain. In these areas, there was a very high level of nitrogen dioxide.

The result of our study is different from those mentioned above. We have found a precise quantity of PM 2.5 and PM 10 that can increase the probability of death in a COVID-19 context.

These three French cities could serve as a study sample. In particular, we can say that all the cities in the world that have a population density similar to these three French cities, must keep the level of PM 2.5 and PM 10 below the threshold values that we found. Milan in February recorded PM 10 values above our threshold level (Table 8) Italian cities. Atmospheric particulate matter may, therefore, have played a carrier role in COVID-19 (Chen et al., 2010; Ye et al., 2016; Chen et al., 2017; Peng et al., 2020) . The positive correlation between PM 10 and the number of deaths could also be due to another reason.

Continuous exposure to fine particles causes sever inflammation of lung tissue. The angiotensin II converting enzyme (ACE-2) is involved in this inflammation process. This enzyme is also the key receptor through which the COVID-19 is able to enter into human cells.

Road transport displays important fossil fuel needs whose combustion is a leading driver of Particulate Matter (PM) concentrations in high populated cities worldwide. Far from being spared, French cities have been experiencing critical air pollution levels (notably PM 10 and PM 2.5 ) for the last two decades, inducing chronic adverse health effects for urban inhabitants.

While taking a dramatic dimension, several municipalities are calling for an urgent introduction of low-carbon fuels for road transport in complement with a more efficient regulation of urban travels within cities. The purpose of this paper aims is to analyse the potential effects of these harmful particles in spreading the current COVID-19 epidemic in France. The underlying hypothesis is that a pre-determined particulate concentration can foster COVID-19 and make the respiratory system more susceptible to this infection.

This study investigated the empirical relationship between particulate matter concentrations (PM 10 and PM 2.5 ) and deaths from COVID-19 in three French cities (Paris, Marseille and Lyon).

To do so, we collected and averaged the concentrations measures given by operating environmental monitoring stations for each city and each pollutant. Then, we merged it with the daily reports on COVID-19-related-deaths provided at city level. Through the use of an experiment in ML with ANNs, we estimated the threshold value of PM 10 and PM 2.5 , beyond which the number of deaths in the presence of COVID-19 would increase. Then, we checked the consistency of our results using a D2C algorithm capable of predicting the existence of a direct causal link between two variables in a multivariate setting. The study takes into account the adverse health effects of the particulate objects of the study. Even in the absence of a pandemic situation, high concentrations of PM 10 and PM 2.5 generate adverse effects and danger to human health. These tiny particles can be inhaled, reaching the deepest part of the human respiratory system (Dagher et al., 2005; Dominici et al., 2006; Arbex et al., 2009) . They can be inhaled and penetrate deep into the lungs and the circulatory system, travelling into the blood and reaching the cells (Wang et al., 2019) . Numerous scientific studies have shown that particle exposure can induce various health effects including heart or lung disease, non-fatal heart attacks, irregular heartbeat, aggravated asthma, decreased lung function, and increased respiratory symptoms such as irritation of the airways, coughing, or difficulty breathing (Atkinson et al., 2010; Meister et al., 2012; Correia et al., 2013; Fang et al., 2013; Cadelis et al., 2014) . Thus, reducing the exposure to high PMs concentrations levels could prevent thousands of premature deaths (Boldo et al., 2006; Kim et al., 2015) . Meanwhile, the most common symptoms of COVID-19 illness are fever, cough, and fatigue, while other symptoms include sputum production, headache, haemoptysis, diarrhoea, dyspnoea, and lymphopenia (Carlos et al., 2020; Ren et al., 2020; Rothan and Byrareddy, 2020; . By comparing these symptoms, we notice that PMs and COVID-19 induce both adverse respiratory effects causing deaths in the worst cases. But it should be mentioned that unlike PMs, the COVID-19 infection may rapidly induce deaths or severe complications.

Linking air pollution and COVID-19 fatality refers to the empirical assessment of a crucial hypothesis: a pre-determined particulate concentration could foster COVID-19 and make the respiratory system more susceptible to this infection. As potential carriers of pathogens, airborne particles could make the viral infection spread more harmful (Setti et al., 2020) .

Meaningful for future environmental and health policies, such empirical analysis is believed to bring additional findings regarding the atmospheric co-factors of COVID-19 lethality. The finer fractions could filter even deeper into our body by travelling into the blood and reaching the cells. We have followed some research that affirms a correlation between air pollution and the spread of COVID-19. Starting from these hypotheses, we wanted to verify the possibility of determining precise values of PM 10 and PM 2.5 which correspond to the optimal value for the diffusion of the coronavirus. We found that if the signal from the neural network (from input to output) is cut to a precise amount, there is a reduction in the number of coronavirus deaths in the three French cities studied. This result suggests that there are certain conditions which increase the likelihood of the spread and aggravation of the disease. The three cities taken as a statistical sample in this study have different population densities. We found that threshold values of PM 2.5 and PM 10 were different among Paris, Lyon, and Marseille. In particular, the new threshold levels of PM 2.5 and PM 10 connected to COVID19 are: 17.4 µg/m 3 (PM 2.5 ) and

29.6 µg/m 3 (PM 10 ) for Paris; 15.6 µg/m 3 (PM 2.5 ) and 20.6 µg/m 3 (PM 10 ) for Lyon; 14.3 µg/m 3 (PM 2.5 ) and 22.04 µg/m 3 (PM 10 ) for Marseille. These findings have been corroborated by the D2C algorithm whose results confirmed the direct relationship between air pollution and COVID-19-related deaths. It is interesting to note that all the threshold values identified by the ANNs are higher than the limits imposed by the European Parliament. From the point of view of the European economic policy, raising environmental standards may require a quantification and balance between the interest in protecting public health and promoting economic growth.

In line with previous studies, our results showed evidence of a direct relationship between air pollution and COVID-19 mortality in France, confirming previous research regarding environmental factors involved in viral infection spread. This could partially explain the efficiency of national lockdown measures and provide useful implications for the prevention of this virus. We believe, therefore, that this result can be replicated in any city that has a population density similar to one of the cities that we studied. With its limits, our study underlines the necessity to recommend further environmental intervention policies limiting the concentrations of PM 10 and PM 2.5 below the thresholds that we found. Pending further scientific confirmation, our threshold value could be considered a possible indirect indicator of the virulence of the COVID-19 epidemic. Another significant point drawn from this analysis is that it could bring useful information in the event of a second wave of the pandemic. Policymakers are expected to keep PM 10 and PM 2.5 concentrations in line with our threshold until a cure or vaccine is available. To address this problem at its major source, a particular attention should be payed to the consumption of oil products (gasoline and diesel) for road transport in urban areas. Adequate urban planning is suggested to solve the chronic traffic congestion experienced by French cities. In addition, promoting low-carbon (biofuels), active (walking and cycling) and collective (bus and subway) transport modes for short-distance travels may lower the fleet's size of motor vehicles in French cities, reducing thus its petrol needs and harmful PMs emissions.

Artificial intelligence and machine learning to fight COVID-19

Good in The Worst: COVID-19 Restrictions and Ease in Global Air Pollution

Urban air pollution and chronic obstructive pulmonary disease-related emergency department visits

Urban ambient particle metrics and health: a time-series analysis

Recent trends in global production and utilization of bio-ethanol fuel

Air pollution sharply falls worldwide on COVID-19

Coronavirus (COVID-19) Classification using CT Images by

Real-world carbon dioxide impacts of traffic congestion

Can microalgae biodiesel contribute to achieve the sustainability objectives in the transport sector in France by 2030? A comparison between first, second and third generation biofuels though a range-based Multi-Actor Multi-Criteria Analysis

Understanding the heterogeneity of adverse COVID-19 outcomes: the role of poor quality of air and lockdown decisions

Apheis: Health impact assessment of long-term exposure to PM 2

A neural network based method for short term predictions of ambient SO 2 concentrations in highly polluted industrial areas of complex terrain

British Petroleum (BP)

Short-term effects of the particulate pollutants contained in Saharan dust on the visits of children to the emergency department due to asthmatic conditions in Guadeloupe (French Archipelago of the Caribbean)

Is short-term exposure to ambient fine particles associated with measles incidence in China? A multi-city study

Ambient influenza and avian influenza virus during dust storm days and background days

Flow Prediction by two Back Propagation Techniques Using k-fold Partitioning of Neural Network Training Data

Can atmospheric pollution be considered a cofactor in extremely high level of SARS-CoV-2 lethality in Northern Italy

Does Air Pollution Influence COVID-19 Outbreaks

The effect of air pollution control on life expectancy in the United States: an analysis of 545 US counties for the period

Pro-inflammatory effects of Dunkerque city air pollution particulate matter 2.5 in human epithelial lung cells (L132) in culture

Fine particulate air pollution and hospital admission for cardiovascular and respiratory diseases

Exposure to air pollution and COVID-19 mortality in the United States: A nationwide cross-sectional study

Biofuels, tax policies and oil prices in France: Insights from a dynamic CGE model

Adresse aux Français du Président de la République Emmanuel Macron

European Union publication office (EUPO), 2012. EU Transport in Figures

Air pollution and associated human mortality: the role of air pollutant emissions, climate change and methane concentration increases from the preindustrial period to present

Economic Effects of Coronavirus Outbreak (COVID-19) on the World Economy

The timedependent pollution-routing problem

French Federation of Certified Associations for the air quality Monitoring

Estimation de la Population au 1 er janvier 2020. Séries par région, département, sexe et âge de 1975 à 2020

COVID-19, France et Monde. GEODES: suivre l'évolution de l'épidémie de COVID-19 en France

The Ultimate Guide to Understanding Air Quality Data in 2020

Clinical features of patients infected with 2019 novel coronavirus in Wuhan

Contribution of Working Group III to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change

Air pollutants and early origins of respiratory diseases

A review on the human health impact of airborne particulate matter

Temporary reduction in daily global CO2 emissions during the COVID-19 forced confinement

Potential emissions reduction in road transport sector using biofuel in developing countries

Motor transport related harmful PM2. 5 and PM10: from onroad measurements to the modelling of air pollution by neural network approach on street and urban level

The relationship between municipal solid waste and greenhouse gas emissions: Evidence from Switzerland

Estimated short-term effects of coarse particles on daily mortality in

A Machine Learning Analysis of the Relationship among Iron and Steel Industries, Air Pollution, and Economic Growth in China

Impact of COVID-19 related shutdown on atmospheric carbon dioxide level in the city of Kolkata

Climate 101: Air Pollution

Assessing nitrogen dioxide (NO2) levels as a contributing factor to the coronavirus (COVID-19) fatality rate. Science of The Total Environment

Higher virulence of COVID-19 in the air-polluted regions of eight severely affected countries

Estimation of the contribution of road traffic emissions to particulate matter concentrations from field measurements: a review

Artificial Neural Networks for Pollution Forecast

Probabilistic Reasoning in Intelligent Systems

Causality: Models, Reasoning, and Inference

Direct and indirect effects

Causes of effects and effects of causes

The effects of air pollution and meteorological factors on measles cases in Lanzhou

Energy and environment challenges in the transport sector

COVID-19 Epidemic Analysis using Machine Learning and Deep Learning Algorithms

Italy: knowledge, management of patients and clinical experience of Italian dentists during the spread of contagion

Machine learning using intrinsic genomic signatures for rapid classification of novel pathogens: COVID-19 case study

Identification of a novel coronavirus causing severe pneumonia in human: a descriptive study

Experimental determination of deposition of diesel exhaust particles in the human respiratory tract

Air Pollution. Our World in Data, last revised in November

The epidemiology and pathogenesis of coronavirus disease (COVID-19) outbreak

Evaluation of the potential relationship between Particulate Matter (PM) pollution and COVID-19 infection spread in Italy

Links between air pollution and COVID-19 in England

Which energy source for road transport in the future? A comparison of battery, hybrid and fuel cell vehicles

Severe air pollution events not avoided by reduced anthropogenic activities during COVID-19 outbreak. Resources, Conservation and Recycling

The impacts of urbanization on fine particulate matter (PM 2.5 ) concentrations: Empirical evidence from 135 countries worldwide

Updated understanding of the outbreak of 2019 novel coronavirus (2019-nCoV) in Wuhan

Director-General's opening remarks at the media briefing on COVID-19

World Health Organization (WHO), 2020b. Coronavirus disease 2019 (COVID-19)

Situation Report -52

?sfvrsn=e2bfc9c0_4> World Health Organization (WHO), 2020c. Coronavirus disease 2019

Situation Report -98

Exposure to air pollution and COVID-19 mortality in the United States

Haze is a risk factor contributing to the rapid spread of respiratory syncytial virus in children

Association between short-term exposure to air pollution and COVID-19 infection: Evidence from China

A deep learning and image-based model for air quality estimation

Derivation of Backpropagation in Convolutional Neural Network (CNN)

Experimental investigation of regulated and unregulated emissions from a diesel engine fuelled with Euro V diesel fuel and fumigation methanol

Conceptualization, Methodology, Software

Visualization, Investigation, Validation, Writing-Reviewing and Editing

Data curation, Writing-Original draft preparation