A particle swarm optimization approach for predicting the number of COVID-19 deaths

The rapid spread of the COVID-19 pandemic has raised huge concerns about the prospect of a major health disaster that would result in a huge number of deaths. This anxiety was largely fueled by the fact that the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), responsible for the disease, was so far unknown, and therefore an accurate prediction of the number of deaths was particularly difficult. However, this prediction is of the utmost importance for public health authorities to make the most reliable decisions and establish the necessary precautions to protect people’s lives. In this paper, we present an approach for predicting the number of deaths from COVID-19. This approach requires modeling the number of infected cases using a generalized logistic function and using this function for inferring the number of deaths. An estimate of the parameters of the proposed model is obtained using a Particle Swarm Optimization algorithm (PSO) that requires iteratively solving a quadratic programming problem. In addition to the total number of deaths and number of infected cases, the model enables the estimation of the infection fatality rate (IFR). Furthermore, using some mild assumptions, we derive estimates of the number of active cases. The proposed approach was empirically assessed on official data provided by the State of Qatar. The results of our computational study show a good accuracy of the predicted number of deaths.

www.nature.com/scientificreports/ purely mathematical/algorithmic approach to answer the aforementioned questions. Our main goal is to derive a mathematical model to predict the number of deaths from COVID-19. Moreover, the model can be used to obtain estimates of IFR, number of infected cases and number of active cases. In addition to its simplicity, a special feature of the proposed approach is that it relies on parsimonious input data. More specifically, it requires the following inputs: (1) Daily number of deaths, and (2) Minimum and maximum time elapsed from infection to death (no distribution is required). Furthermore, to obtain estimates of the number of active cases, minimum and maximum time between infection and recovery are needed as well. The basic idea is to model the total number of infected cases using a generalized logistic function (also called Richards' curve) having four parameters. These parameters are estimated by minimizing the sum of the quadratic errors between the actual cumulative number of deaths and the predicted ones. An approximate solution to the optimization problem is obtained using a Particle Swarm Optimization algorithm (PSO). The effectiveness of the approach is assessed through experiments that were carried out on official data provided by the State of Qatar. The choice of this country is motivated in particular by the fact that since the detection of the first local COVID-19 positive case (on February 29, 2020), the epidemic has spread on such a large scale that Qatar has topped the world ranking for the rate of positivity in COVID-19 (which is measured as the ratio of positive cases to the total population).

Literature review
As per the study carried out in 4 , mathematical predictive models dealing with COVID-19 can be classified into two main categories: transmission dynamics models and phenomenological models. The main purpose of the first models category is to describe the mechanisms underlying an epidemic and understand its behaviour as well as the effects of human interventions. These models use a large number of parameters unlike the second models category. Many research studies were conducted to provide a transmission dynamic model for COVID- 19. In 5 , an SIR (susceptible-infected-recovered) epidemic model for COVID-19 spread in Indonesia with fuzzy parameter was constructed by taking into account the factors of vaccination, treatment, health protocols, and the virus-load. In 6 , a conceptual mathematical model on the transmission dynamics of COVID-19 between the frontliners and the general public was formulated by considering an SEIR (susceptible-exposed-infected-recovered) compartment model where a huge number of parameters is required (up to 23 parameters). According to the available infection data, authors in 7 elaborate a dynamical model based on the fractional derivatives of nonlinear equations that describe the outbreak of COVID-19. Such a model necessitates a very large number of parameters. An epidemiological model whose implementation is made in Excel spreadsheets is developed in 8 . When adjusting many parameters, it was shown that the model follows the evolution of COVID-19. A new SIR model was developed in 9 including direct and fomite transmission as well as household structure applied to COVID-19. The model involves many parameters such as: the infection rate related to direct transmission, the infection rate related to fomite transmission, the minimum contact radius and the average number of daily cross-household encounters between a susceptible and an infected person. In 10 , a fractional epidemic model in the Caputo sense taking into account quarantine, isolation, and environmental impacts was developed in order to track the dynamics of the COVID-19 pandemic. A Gaussian Process Regression model was proposed in 11 to predict the number of COVID-19 deaths. It was compared to Artificial Neural Network model. The study examines the impact of age, number of smokers and number of diabetic patients on increasing the number of deaths.
Phenomenological models which rely on a limited number of parameters, can afford a quick epidemic forecast efficiently 4 . In this context, a bimodal lognormal distribution function is developed in 4 for the prediction of the time distribution of COVID-19 deaths. Nevertheless, this model cannot be used for estimating the IFR nor the number of active cases. In 12 , a demographic scaling model for estimating the total number of COVID-19 infections was developed. This approach requires the following input data: the number of COVID-19-related deaths for the population of interest; and the age-specific infection fatality rates from a reference population, scaled to match the target population based on life tables. The study in 13 deals with the impact of air pollution during the COVID-19 pandemic on public health where a linear regression analysis was conducted. It was demonstrated that geo-environmental conditions can speed up the spread of COVID-19 in Italy and thus contribute to higher number of infected individuals and deaths.
Numerous research studies were conducted to investigate the influence of socioeconomic, demographic, climatological and environmental factors on the spread on COVID-19 14 . The idea in 14 consists in proposing an index quantifying the environmental risk of exposure of cities or regions to future epidemics of the COVID-19 and similar pandemics. Such an index helps policymakers in preventing future epidemics. In 15 , the study investigates correlation of the spread and decay durations with population density, temperature, and absolute humidity in four countries. In 16 , the study shows how the duration of lockdownd due to the COVID-19 pandemic, has an impact on the rates of infected people and deaths, as well as the growth of Gross Domestic Product (GDP) of nations. A comparative analysis of the first and second wave of COVID-19 was made in 17 . This study helps in establishing adequate decisions regarding the public health versus next COVID-19 waves or other pandemics. The approach of 18 analyzes the impact of socio-economic factors on the reduction of the COVID-19 fatality rate. It suggests that high GDP per capita, high healthcare spending and low levels of air pollution can reduce the fatality rate. An estimation study of the dispersion parameter was developed in 19 . This parameter helps in measuring the heterogeneity in COVID-19 transmission. The distribution of the number of secondary cases is considered as a negative binomial (NB) distribution with the dispersion parameter. In 20 , an estimation study of the excess of deaths in the United States due to the COVID-19 pandemic from March to May 2020 was conducted. The approach was based on data on all-cause deaths and pneumonia/influenza/COVID-19 deaths. An estimation study of age and sex-specific excess mortality risk and deaths due to COVID-19 was elaborated in 21  www.nature.com/scientificreports/ Unlike the aforementioned studies, our approach describes a novel phenomenological model that enables to derive an estimate of the number of positive cases (including the undetected cases) along with the IFR. Our work is in line with many mathematical approaches developed so far to model various epidemics such as severe acute respiratory syndrome (SARS) [22][23][24] , Ebola 25 , Middle East respiratory syndrome (MERS) 24,26 , etc. Indeed, the authors of 22 proposed a simple mathematical model to predict the number of cases and deaths of SARS. Unlike the algorithmic estimation methods, the developed model consisted in trying out several values of the parameters enabling the best fitting that matches the observed data. This modelization is considered simple to help the front line public health practitioners. In 25 , the study was about modeling the transmission dynamics of Ebola virus disease in Liberia and estimating the basic reproduction number when there is no control measures. It has been shown that the epidemic can be well contained if isolation and safe burial measures are taken. The provided results help in establishing guidance and allow the spread control of the virus. A compartmental model for the outbreak of MERS similar to SARS models was studied in 26 . A coupled system of non linear ordinary differential equations was considered. The study showed that the contact coefficients, the isolation rate constant and the transmission factor have a huge impact on the containment time and the outbreaks severity. In 23 , an SEIJR model (susceptible, exposed, infected, diagnosed, recovered) is considered for the study of SARS epidemic by including diffusion in the system in order to examine the spacial spread of the disease. The diffusion impact on the disease spread was investigated for different initial population distributions and interventions effect was analyzed. Authors in 27 modeled the interaction between the epidemic spreading and information diffusion by developing a two-layer network. The study showed that the knowledge diffusion has a significant impact on controlling not only rumor but also epidemic. The purpose of 28 consisted in developing a global threshold dynamics for several SVEIS (susceptible, vaccinated, exposed, infectious) epidemic models with temporary immunity, vaccination, latency and nonlinear incidence. The study revealed that reducing the outbreak peaks (such as: global seasonal influenza burden, measles cases surge and COVID-19 pandemic) can be efficient and ensured with timely vaccination more than other measures.
Since the first weeks of the COVID-19 epidemic, multiple research studies were elaborated aiming at predicting the new virus evolution and dynamics in order to guide public health authorities to take the most reliable decisions to protect people. A comparison study of the epidemics spread has been elaborated in 24 between the new virus COVID-19, SARS and MERS. A propagation growth model was established for the three epidemics and a nonlinear fitting was considered for the determination of their parameters. The obtained results confirmed that the COVID-19 presents the highest growth rate and the shortest multiplication cycle versus SARS and MERS. Authors in 29 investigated two models to predict the COVID-19 confirmed cases in the most affected countries: the autoregressive integrated moving average (ARIMA) and the least square support vector machine (LS-SVM). It has been concluded that the LS-SVM contributes to more accuracy than the ARIMA model. The provided predictions enable authorities to take the most convenient decisions and precautions. In 30 , the main idea consisted in predicting the number of positive cases of the COVID-19 pandemic in Italy with time evolution. The mathematical model is based on official data as well as on the use of a cumulative distribution function by referring to the Gauss error function with four parameters. The proposed approach provided very good fit with the China case when plotting the cumulative number of cases and the cumulative number of fatalities. The number of positive cases in Italy is then predicted by performing a number of fits of the available data. The fit flex point and the day in which a substantial attenuation of daily cases are both determined. The latter results were confirmed by performing Monte-Carlo simulations. Authors in 31 provided estimating models to characterize the dynamics of SARS-CoV-2 transmission in France. Delays from hospitalization to death and from hospitalization to ICU (Intensive Care Unit) are estimated by age range according to collected data. The infection fatality ratio is calculated from the Diamond Princess and the age distribution of the French hospitalized population. The risk of hospitalization, ICU admission and death are modeled and the impact of the lockdown has been taken into account to compute the transmission rate. Authors in 32 developed prediction models based on Genetic Programming to estimate confirmed cases (CC) and death cases (DC) of COVID-19 in three states of India as well as whole India. The gene expression programming model used simple linkage function and provided robust and reliable results (for CC and DC). It relied on daily situation reports of COVID-19 published by the Indian government. It was shown that proposed models are less sensitive to variables, based on experimental data and can be used to predict future cases. The expression trees simplicity introduced basic mathematical equations that can be implemented without time consuming. The main idea of 33 consisted in estimating the COVID-19 evolution in terms of cumulative number of infected cases, daily infected cases and the Fisher-Pry. The predictive model was based on the SIR model (SIR stands for Susceptible, Infected and Recovered) reduced to the Verhulst equation. It provided a reliable and accurate prediction of the epidemic for a whole country. According to the Fisher-Pry plots for different countries, the SIR-Verhulst modelization proved its efficiency in quantifying public health measures. That's why relying only on such a plot can be sufficient to reveal significant conclusions about the epidemic evolution in different countries.
In 34 , an estimation of the time between symptoms onset and outcome (death or recovery) is made based on individual-case data for infected cases dying due to COVID-19 in Hubei, Wuhan. This parameter was estimated as 17.8 days from onset-to-death and as 24.7 days from onset-to-recovery. Overall case fatality ratio (CFR) in China was estimated as 13.8% while the estimated overall infection fatality rate (IFR) was 0.66%. Authors provided detailed estimates of the two latter parameters for different age groups and showed higher ratios estimates for older persons (CFR: 13.4% and IFR: 7.8% for persons aged more than 80 years-old). In 35 , authors developed a modified SIR model to predict COVID-19 evolution in Italy. The idea consisted in considering the initial number of susceptible individuals as well as a proportionality between the detected number of positive cases and the actual/unknown number of infected individuals. The model parameters were determined for many Italian regions using a standard weighted least-squares optimization problem. The main idea in 36  www.nature.com/scientificreports/ and predicting the total infections and the day of the infection peak. The model provided reliable predictions for China, Italy and Spain. The impact of partial/total lockdown on the pandemic spread was studied through the curves interpretation and it was concluded that the epidemic lifespan is lockdown independent. Authors in 37 proposed a generalized SEIR (susceptible, exposed, infectious, recovered) model including insusceptible cases, exposed cases and quarantined cases in order to predict inflection point and possible ending time of the COVID-19 in 24 provinces in Mainland of China and 16 counties in Hubei province. The considered model involved several key parameters that were estimated as well such as: latent time and quarantine time. The model provided reliable results fitting the empirical public data. Authors in 38 , developed a new mathematical SEIR-HC [SEIR (susceptible, exposed, infectious, removed), HC (Hospital, Community)] model by considering two different social circles (individuals in hospital and community). A two-step iterative optimization approach is proposed to estimate the model parameters. It has been shown that quarantine and isolation are important measures that play a considerable role in reducing the epidemic spread. A mathematical model for COVID-19 was developed in 39 taking into consideration the impact of imported cases and different rates (isolating, diagnostic, recovery, mortality). The model is dynamic for the epidemic transmission process in terms of quantity of daily infected individuals. The results showed that the isolation of infected cases and the strict control of input cases contribute in stopping the epidemic spread. Authors in 40 developed a SEIQR (susceptible, exposed, infected, quarantined, recovered) mathematical model to study the COVID-19 spread. The model is based on nonlinear system of coupled ordinary differential equations. A study of the model in terms of consistency, stability and convergence was carried out and showed its validity. A two-parameter SIR mathematical model is investigated in 41 to study the SARS-Cov-2 epidemic. The elaboration showed that the SIR model is restricted as it can be applied reliably on small populations to provide consistent results.

Optimization approach for COVID-19 deaths estimation
In the sequel, we shall use the following notation. The total number of cases I(t) is assumed to follow a generalized logistic function. The generalized logistic function is an S-shaped function that has been originally used for modeling growth curves. In contrast to the well-known logistic function (Verhulst function), the generalized logistic function is used for modeling situations where the growth curve is asymmetrical about its inflection point. Several variants of this function have been investigated so far. In this paper, we consider the following form: where (K, Q, , ν) are four parameters allowing a high flexibility of the sigmoid curve. During the last years, the generalized logistic model has been widely used in epidemiology for deriving predictions of the spread of epidemics including SARS 42,43 and dengue fever 44,45 . Recently, several authors used it for deriving prediction models for the COVID-19 outbreak [46][47][48] . The computation of the optimal parameters of a generalized logistic function is a challenging nonlinear least squares problem that has been addressed by many authors 49 .
Remark To our knowledge, all of the previously published papers dealing with the use of the generalized logistic function to predict the spread of an epidemic address the problem of fitting positive case data. However, in the  www.nature.com/scientificreports/ case of the COVID-19 outbreak, focusing on positive cases is largely questionable. Indeed, as the number of cases infected with COVID-19 is generally much higher than the number of positive cases diagnosed, the practical relevance of the approach then becomes questionable. In contrast, in this paper, we shall build a model for predicting the total number of cases (including both detected and undetected ones).

The number of new cases on day t is denoted by N(t) and is deduced as follows:
Consequently, the number of new deaths on day t is modeled as the the sum of a linear combination of the number of new cases of days t − τ max , t − τ max + 1, . . . , t − τ min . Hence, we set: where a j is the expected rate of individuals that die j days after being infected ( j = τ min , . . . , τ max ).
Accordingly, the cumulative number of deaths can be expressed as follows: Similarly to the daily deaths, the number of infected people that recover on day t is modeled as the the sum of a linear combination of the number of new cases of days t − r max , t − r max + 1, . . . , t − r min . Thus, we define: where b j represents the rate of individuals that recover j days after being infected ( j = r min , . . . , r max ).
Clearly, a j ( b j ) can be seen as the probability that an individual will die (recover) j days after being infected. Therefore, we have: In the sequel, we shall make the assumption that the distribution of the b j is unimodal. We denote the mode by j * . This assumption is supported by the findings of 34 who showed that the best fit of the onset-to-recovery distribution is a gamma distribution.
Therefore, the number of active cases on day t is denoted by A(t) and is given by: To estimate the cumulative number of deaths, we use the following regression model: where ǫ t is a normally distributed random error with mean zero. Using (8), we get:    www.nature.com/scientificreports/ Given a quadruplet (K, Q, , ν) , we define Z(K, Q, , ν) as the optimal value of the following Quadratic Programming (QP) problem: Clearly, (QP) aims at finding the values of the a j 's, b j 's and ǫ t 's that minimize the sum of quadratic errors between the reported cumulative number of deaths and the predicted ones. Constraints (17)  To find the parameters of the generalized logistic function, we formulate a Global Optimization problem (GO) that requires finding the quadruplet that minimizes Z(.). Hence, the problem reads as follows: Finding a global minimum of Problem (GO) poses very hard challenges. Instead, we derive an approximate solution by using a Particle Swarm Optimization algorithm (PSO). Since its introduction about two decades ago by Kennedy and Eberhart 50 , this approach has been extensively used for solving hard continuous global optimization problems (See 51-53 and the references therein). PSO algorithm was applied in several research studies to deal with COVID-19 prediction. In 54 , a susceptible-exposed-infected-quarantined-recovered-dead (SEIQRD) model was presented to predict the COVID-19 spread in Italy. The model parameters are estimated using the PSO algorithm. In 55 , a novel PSO-BLS (particle swarm optimization broad learning system) was developed to predict the dynamics of COVID-19. The proposed model shows higher accuracy and stability versus deep learning models when predicting the number of active infected cases and removed cases. Four phenomenological epidemic models as well as SIR model were investigated in 56 in order to predict the cumulative number of confirmed cases due to COVID-19 in Saudi Arabia and the possible end-date of the pandemic. The idea was based on an optimization approach using the PSO algorithm.
Once problem (GO) is solved, an IFR estimate is computed as the sum of the infection-to-death probabilities:

Numerical results
We consider the following settings: 1. The minimum/maximum time from onset of symptoms to death is 5 days/21 days 34 .
2. The minimum/maximum time from onset of symptoms to recovery is 4 days/34 days 34 . The mode of this distribution is 22. 3. The mean incubation time (IT) of COVID-19 is about 6 days 57 . Hence, we set the minimum/maximum time from infection to death to 11 days/27 days, the minimum/maximum time from infection to recovery to 10 days/40 days, and the mode of the infection-to-recovery distribution to 28 days. 4. The empirical data of cumulative deaths are provided in 58,59 . The date t = 1 corresponds to February 29, 2020, which is the date when the first positive case was detected in Qatar.
The proposed approach was implemented via MATLAB R2019b (we used the PSO subroutine embedded therein), and run on a machine with an Intel i5-2430M processor and 6 GB of RAM.  www.nature.com/scientificreports/ To assess the performance of the proposed approach, we run it on two input sets that were obtained by considering the number of deaths that were reported during the first 120 and 180 days of the outbreak, respectively. The obtained generalized logistic curves as well as the curve of the reported cumulative number of deaths (up to Day 300) are displayed in Fig. 1. The parameters of the derived curves are displayed in Table 1. In this Table, we also report: the values of the objective functions, the values of the root-mean-square errors (RMSE), and the CPU times.
We performed a Kolmogorov-Smirnov test for checking the hypothesis that the prediction errors are normally distributed. The results are displayed in Table 2. We observe that for both scenarios the null hypothesis (that is, the errors are normally distributed) cannot be rejected at the 5% significance level. Furthermore, we performed a hypothesis test for checking whether the mean error is zero. We found that, in both cases, it is rejected at the 5% significance level, but it cannot be rejected at the 1% significance level (see the the corresponding confidence intervals in Table 3).   www.nature.com/scientificreports/ To get a quantitative measure of the quality of the predictions, we computed for each prediction scenario (that is, T = 120 and 180, respectively), the relative deviations, which are defined as follows: for t = T + 1, . . . , 215 , where 215 corresponds to the last day for which data is available (September 30, 2020). We found that the first scenario (T = 120) exhibits a positive mean deviation of 7.22% , which indicates that the estimates are on the average slightly overestimated. However, the estimates are fairly robust since the maximum absolute deviation is 14.88% . On the other hand, the second scenario (T = 180) exhibits a negative mean deviation of −3.85% , and a remarkable maximum absolute deviation of 6.64% only.
Using (25), the first scenario (T = 120) yields an estimate value of the IFR that is equal to 0.11%. On the other hand, using much more available data, the second scenario (T = 180) yields an estimate value that is equal to 0.068%. To the best of our knowledge, there is no published estimate of the IFR in Qatar, and therefore it is infeasible at this stage to get a rigorous assessment of the accuracy of our estimates. Nevertheless, it is possible to get some insights by comparing these values with those estimates that were recently published in the literature and that are based on data from antibody-prevalence studies. In particular, the authors in 60 published the results of a nation-wide study in Spain that tested for antibodies in more than 61, 000 randomly selected individuals. This study provides estimates of the range of the age-specific IFR for severe acute respiratory syndrome coronavirus 2. The results of this study provide evidence that the IFR is nearly zero for anyone who is under forty and increases to 7.2% for people in their eighties (see Table 4).
We used Table 4 for deriving a "rough" estimate of overall IFR for Qatar. Toward this end, we considered the present distribution of the population of Qatar by age group 61 . This distribution is displayed in Table 5. Combining the data in Tables 4 and 5, we derived an overall IFR for Qatar that is equal to 0.086%, which is pretty close    www.nature.com/scientificreports/ (through RT-PCR tests) largely underestimates the actual spread of the outbreak. Also, we observe that in both cases, the estimate values have generally increased between the first prediction (Day 120) and the second one (Day 180). In particular, the values of parameter K, which corresponds to the asymptotic limit of I(t) was first estimated to 195,000 and then to 296,000. This latter estimate suggests that by the end of the outbreak about 10.5% of the population of Qatar would have been infected by the SARS-Cov-2. At this point it is worth mentioning that sampling-based methods for estimating seroprevalence of anti-SARS-CoV-2 antibodies in a population have been used elsewhere by many authors. In particular, a study that was carried out on a large sample of randomly selected adult patients receiving dialysis in the US found evidence that 9.3% of the US adult population has been infected so far 3 . On the other hand, a second study 63 that was carried out in the region of Geneva, Switzerland (which was heavily affected by the first wave of the outbreak), found a seroprevalence ranging on different weeks between 4.8 and 10.9%. Using (10) and (13), we derived estimates of the number of active cases. The obtained curves are displayed in Fig. 4. Here again, we observe that the number of detected active cases are much smaller than the predicted values, and that the estimate values of Day 180 are significantly larger than were obtained on Day 120.

Conclusion
Keeping in mind that the proposed approach is intended to be used primarily for forecasting of the number of deaths and estimation of the death rate from infection, we believe that it offers remarkable advantages over similar approaches, but also has some limitations. It is a glaring fact that the vast majority of the modeling approaches that have been proposed so far for modeling and forecasting the spread of COVID-19 rely heavily on data and parameters that are difficult to know precisely. Indeed, these models require information on many parameters that pertain to mobility and behavior of the population, the reproduction number R 0 , the current number of infections, the infection fatality ratio, as well as the fraction of asymptomatic (contagious) individuals, to quote just a few. Often, a small variation in any of these input data and the model can provide forecasts that are off by orders of magnitude. In contrast, the proposed approach is based solely on the number of deaths, which is both much more precise and easier to obtain. At this point, it should be noted that the number of observed deaths is rarely used in the literature to predict the number of deaths. However, whenever this number is used, other parameters are also used. For example, the authors of 64 have developed an interesting SIR model to estimate the number of new COVID-19 infections, as well as the number of people likely to die. However, unlike our approach, they also rely on the infection fatality rate (IFR) for which accurate estimate is still unknown. Here we emphasize that, in our approach, the IFR is an output of the model calculations and not an input. On the other hand, an obvious limitation of our approach is that it assumes that the spread of the epidemic is a stationary process. That is, we assume that no new viral variants have appeared, nor that new epidemic control measures or vaccination campaigns have been implemented. Furthermore, a second limitation is that it cannot be useful for performing a what-if-analysis. Which means that it cannot help answer questions such as "what will be the impact of social distancing?" or "what will be the impact of the school closures?" www.nature.com/scientificreports/ We anticipate that if the proposed approach is run on a periodic basis (as new data accumulates), then it will prove useful for public health authorities for making informed decisions and thus mitigating the impact of the outbreak.
Finally, we used in this study (for the sake of convenience) the PSO code that is embedded within MATLAB. However, it is likely that some recent and much sophisticated PSO variants (including those presented in [51][52][53] ) might prove much effective and thus would probably yield better estimate values. This is a first future research direction that we recommend investigating. Furthermore, we recommend as a future research direction to use the proposed approach for deriving estimates of the IFR for different countries and to develop a regression analysis for investigating the influence of socioeconomic, demographic, climatological and environmental factors on the spread on COVID-19.