PreprintPDF Available

A simple but complex enough θ-SIR type model to be used with COVID-19 real data. Application to the case of Italy

Authors:
ResearchGate Logo

This preprint is featured on the COVID-19 research community page

View COVID-19 community

Abstract and Figures

* Published version: https://doi.org/10.1016/j.physd.2020.132839 ____________________________________________________________________________________ *Java code and data are available at: https://github.com/momat-ucm/T-SIR-T ____________________________________________________________________________________ * Summary: Since the start of the COVID-19 pandemic in China many models have appeared in the literature, trying to simulate its dynamics. Focusing on modeling the biological and sociological mechanisms which influence the disease spread, the basic reference example is the SIR model. However, it is too simple to be able to model those mechanisms (including the three main type of control measures: social distancing, contact tracing and sanitary measures) to fit real data and to simulate possible future scenarios. A question, then, arises: how much and how do we need to complexify a SIR model? We develop a θ-SEIHQRD model, which may be the simplest one satisfying the mentioned requirements for arbitrary territories and can be simplified in particular cases. We show its very good performance in the Italian case and study different future scenarios.
Content may be subject to copyright.
A simple but complex enough θ-SIR type model to be used with
COVID-19 real data. Application to the case of Italy
A.M. Ramosa, M.R. Ferrándezb, M. Vela-Péreza, A.B. Kubika, B. Ivorraa
aMOMAT Research Group
Interdisciplinary Mathematics Institute
Complutense University of Madrid
bDepartment of Computer Science,
University of Almería
emails: angel@mat.ucm.es; mrferrandez@ual.es; maria.vela@ucm.es; akubik@ucm.es; ivorra@ucm.es
23 December 2020
(update of a previous version published on 23 July 2020)
DOI link: http://www.doi.org/10.13140/RG.2.2.32466.17601
Accepted for publication in Physica D: Nonlinear Phenomena
DOI link: https://doi.org/10.1016/j.physd.2020.132839 1
Abstract
Since the start of the COVID-19 pandemic in China many models have appeared in the literature, trying to
simulate its dynamics. Focusing on modeling the biological and sociological mechanisms which influence the disease
spread, the basic reference example is the SIR model. However, it is too simple to be able to model those mechanisms
(including the three main type of control measures: social distancing, contact tracing and health system measures)
to fit real data and to simulate possible future scenarios. A question, then, arises: how much and how do we
need to complexify a SIR model? We develop a θ-SEIHQRD model, which may be the simplest one satisfying the
mentioned requirements for arbitrary territories and can be simplified in particular cases. We show its very good
performance in the Italian case and study different future scenarios.
Keywords: Mathematical model, θ-SEIQHRD model, COVID-19, Coronavirus, SARS-CoV-2, Pandemic, Numerical
simulation, Parameter estimation, basic reproduction number, effective reproduction number
1Citation: A.M Ramos, M.R. Ferrández, M. Vela-Pérez, A.B. Kubik and B. Ivorra. A simple but complex enough θ-SIR type model to
be used with COVID-19 real data. Application to the case of Italy. Physica D: Nonlinear Phenomena. Available online (132839). 2021.
DOI: https://doi.org/10.1016/j.physd.2020.132839
1
Data and code availability
All the data and all the model simulation and figure codes used in this manuscript are available at
https://www.ucm.es/momat/software-momat and at https://github.com/momat-ucm/T-SIR-T.
The core model was written in Java and the figures were obtained using MATLAB.
1 Introduction
The outbreak of novel coronavirus disease 2019 (COVID-19) continues to spread rapidly around the world [60]. Orig-
inated in Wuhan in central China, where it was already active at the beginning of December 2019, reached multiple
countries in merely a month until it became a pandemic as of 11 March 2020 [63] when the disease was confirmed in
more than 118,000 cases reported globally in 114 countries.
Although some authors (see, e.g., [37]) propose SEIR type models with little variations, COVID-19 is a disease
caused by a new virus and needs a model taking into account its known specific characteristics. In particular, it would
be convenient to develop a model which satisfies the following: a) includes the biological and sociological mechanisms
influencing the disease spread, b) its parameters have a clear interpretation in terms of those mechanisms (such as a
contact rate, fatality ratios, duration of infection, etc.), and c) incorporates
the effect of undetected infected people (see [39]), being able to show the dependence of the impact of COVID-19
on the percentage of detected cases over the real total infected cases,
the effect of different infectiousness conditions of hospitalized people (differentiating those with mild and severe
conditions that will recover from those who will finally die),
the estimation of the needs of beds in hospitals (which is one of the major problems for policy makers addressing
COVID-19),
the effect of different control measures.
Some of the models that appeared at the beginning of the pandemic, especially for the study of the first wave (see
[19, 33, 37, 39]), do not incorporate all the special characteristics (undetected cases, undetected deaths, quarantine,
social distancing, etc.) of the COVID-19 pandemic. Some of them are very complex in terms of the number of
compartments and connections between them, trying to simulate many details that could probably be neglected.
Furthermore, after developing a model, it would be desirable to check how it fits real data and to use it in order to
study future scenarios with different control measures (here, we consider social distancing, contact tracing and health
system measures).
This paper does not take into account the possibility of resusceptibility as it is done, for example, in [45]. This
is due to the fact that the number of confirmed reinfections is still not relevant enough compared with the total
figures of the pandemic. According to [48], only 27 cases of reinfection were documented worldwide when it was
written. Furthermore, the webpage https://bnonews.com/index.php/2020/08/covid-19-reinfection-tracker/ reports
an updated list of reinfections, with a total of 31 cases, at the time we wrote this article.
In [20], a disease contact rate (β)and R0with different values before and after lockdown of countries are considered
and applied in a SEIR type model for the confirmed cases tested with several countries. The model presented here
allows to model not only confirmed cases but also deaths, hospitalized and recovered by considering different values
of βthat depend on the particular control measures for every time period.
Concerning the particular case of Italy, some studies have dealt with its particular situation, such as [49], [10], [18]
and [4]. It is important to note that all the models presented in these works only serve the particular case of Italy,
and are not general models applicable to other countries.
The paper [49] is an initial work published in mid-March 2020, which highlights the importance of this pandemic
and the serious consequences it can have in countries like Italy where data show rapid progress in infections, which
can cause a health system collapse due to the lack of beds in hospitals and ICU posts. The data are fitted using
2
an exponential type model, which is not realistic for long-term modeling, although it allows a wake-up call to the
country’s health authorities.
In [10], the distribution in terms of mobility is also studied through a network of the different Italian regions. It
is confirmed the effectiveness at a regional level of the national lockdown strategy and proposed coordinated regional
interventions to prevent future national lockdowns. However, asymptomatic patients are not considered.
In [18], a geographical model is presented to describe the number of daily cases, hospitalized and deaths within the
Italian regions for the first month of the disease and study the effects of drastic measures for transmission containment.
Authors concentrate on estimating the effects of severe progressive restrictions posed to human mobility and human-
to-human contacts in Italy. As a novelty, a presymptom compartment is considered. The proposed model is extended
and further developed in [4], where the effect of mobility between regions and provinces within Italy is also considered,
taking into account the control measures at all times to describe the variable of new hospitalizations. The number of
people who have to be isolated in each region is calculated to avoid a drastic growth in the number of infected. This
model is very interesting, because it incorporates the variable of mobility between regions, but only the number of
new hospitalizations is described, without studying the number of deaths since they focus on the implications of the
relaxation of confinement measures.
An interesting model appearing in [19], tested for the case of Italy, has some similarities with the one presented
here, but there are some important differences to be remarked. First of all, the model we propose is a general one
that may be used with any country or region, just considering the correspondent data, whereas the model in [19]
seems to be defined for the particular case of the first wave in Italy, although it could be adapted to other territories.
Secondly, our model considers the compartment for people in quarantine (Q, detected asymptomatic or with mild
symptoms, infectious people) that is not the same as the proposed compartment for diagnosed (D, asymptomatic
infected, detected) in the former model, which is even connected to the possibility of dying. Moreover, the general role
of the undetected people in our model is completely different and some of them can even die (those who are in IDu).
In this paper, we develop a general deterministic epidemiological model able to simulate the biological and sociolog-
ical COVID-19 mechanisms in different territories, to fit real data, and to simulate different future scenarios of control
measures. The choice of using a deterministic model instead of a stochastic one is done as a first approach, since such
kind of models present some advantages, such as: a low computational complexity allowing a better calibration of the
model parameters or the possibility of using the theory of ordinary differential equations for suitably analyzing and
interpreting the model. Furthermore, according to [12], deterministic models should be the first tool to be used when
modeling a new problem. The authors of that work also note that the stochastic models are not suitable when it is
difficult or impossible to determine the probability distribution, are difficult to analyze and require more data for the
calibration of the model. We note that, although we use a deterministic model, we must deal with some probabilistic
processes. In those cases, random values are substituted by their expected values.
We look for the simplest model, having the minimum complexity to take into account all the main non-negligible
processes of this pandemic. When simulating the dynamics of the COVID-19 disease in a territory, our model can be
used in its full general form or in simplified particular versions of it, depending on the way the territory deals with
the pandemic. Of course, the model may be modified by introducing age structure, stochasticity, levels of uncertainty,
mobility between territories, spatial structure (see, e.g., [35]), etc. in order to focus in special features of the pandemic
or to refine some results.
The reference epidemiological model is the SIR one, which divides the population of a territory in three compart-
ments: Susceptible (S), Infectious (I) and Recovered (R). For the COVID-19 pandemic these three compartments
are not enough. First of all, the incubation period for the new coronavirus SARS-CoV-2 is between 5 and 6 days (see
[64]), which is important to incorporate in the model by means of a compartment for Exposed (E) people. Secondly,
estimation of needs of beds in hospital is important for policy makers, which implies the convenience of suitable
compartments for people in Hospitals (H), following a compartment (I) for an initial stage of the infectious period.
Moreover, the disease contact rates inside and outside hospitals are usually different. Thirdly, one of the main data
reported each day is the number of Deaths (D), which deserves another compartment and makes convenient to split
Hin two branches, one for people in hospital that will recover (HR) and the other one for people in hospital that will
die (HD). Therefore, we need, at least, a SEIHRD model.
Furthermore, one of the characteristics of COVID-19 is the existence of undetected infectious cases, either because
they are asymptomatic (see [39]) or because, even if they have symptoms, the corresponding country/territory does
3
not have enough tests to check the infectiousness or the health system is saturated. In order to be able to consider the
impact of this COVID-19 characteristic, two different branches should distinguish between detected and undetected
cases and deaths. This can be done in terms of the instantaneous ratio θ(t)of new detected infectious cases per unit time
over the total (detected and undetected) new infectious cases per unit time, generating the corresponding θSEIHRD
model. As done previously with people in hospital, a distinction can be done between the undetected infectious people
(compartment Iu) that will recover (compartment Ru) and the undetected infectious people (compartment IDu) that
will die (compartment Du). Finally, some territories (like Italy) use a quarantine period (typically at home) for people
leaving the hospital and for detected asymptomatic (or with mild symptoms) infectious people, which makes necessary
a new compartment (Q), which completes the full θSEIHQRD general model. A diagram summarizing the model is
given in Figure 1.
A detailed description and discussion of the model, its parameters and outputs is provided in Sections 2 and 3. We
point out that all the parameters in the model have well-understood epidemiological interpretations such as average
duration of incubation period, contact rates, etc. In Section 2, we also show some of the outputs of the model and
how to compute them. Particular outputs of interest for epidemiologists are the computation of the associated basic
reproduction number and the evolution of the effective reproduction number.
HD
γωγI
(θ−ω)γ(1-p) I
γDu
I
E
γ
I
(1−θ−ω )γ
uγ
21
I
β + β + ...
E
Detected Cases
Undetected Cases
+ Control measures (θ−ω)γ
Iu
u
R
Iu
τ τ
I
u
ω γ I
Du D
Iu
Q
HR
γ
d
R
γ
HR
HD
D
Q
I
p
S E
Figure 1: Diagram summarizing the model for COVID-19 given by system (1).
Once we have developed the model, the question is how to use it with real cases. Let us suppose we want to use
it to study the dynamics of the COVID-19 in a particular territory. Three steps can be followed:
1. The first thing is to decide if all the compartments are necessary, or only some of them. For instance, in
China and Senegal all infected people are treated in hospitals (or other kind of health facilities) and, therefore,
compartment Qis not necessary [33]. However, other territories, like Italy or Spain, use the quarantine at home
for infected people with very mild symptoms or not symptoms at all. Other possibilities of simplifying the model
are also possible, for particular territories.
4
2. After deciding the compartments that will be used with a particular territory, the following step is to choose
the parameters involved in the model. To do that we use values given in the literature (as the duration of
the incubation period), parametrization of functions involved in the model and a multiobjective technique, to
identify parameters. In Section 3 we explain how to do it. This is, actually, part of the validation process,
consisting in comparing the results of the simulations (run with the chosen parameters) with the available data
about the impact of the pandemic in the territory under study. Typically, a good fitting of simulations outputs
to official temporal series (of cases, deaths, people in hospital, etc.) is necessary.
3. Finally, if the validation process is successful, we can simulate different future scenarios to study the efficiency
of some control measures (being increased or relaxed), the impact of the undetected cases in the dynamics of
the pandemic, etc.
In Section 4, we illustrate those three steps with the particular case of Italy.
2 Mathematical formulation of the model
In this section we give a detailed description of the θSEIHQRD model that we propose for the COVID-19. Then,
we described some outputs of the model that will be used in the numerical experiments performed in Section 4.
2.1 The mathematical model
According to the known characteristics of the COVID-19 pandemic, we assume that each person is in one of the
following compartments (see [2, 47, 61, 57]):
Susceptible (denoted by S): The person is not infected by the disease pathogen.
Exposed (denoted by E): The person is in the incubation period after being infected by the disease pathogen,
and has no clinical signs. This stage is also called presymptomatic period. According to the available literature
(see, e.g., [2]), there exist some evidences that a presymptomatic person could infect other people but with a
lower probability than people in the infectious compartments. After the incubation period, the person passes to
the Infectious compartment I.
Infectious (denoted by I): After the incubation period, it is the first compartment of the infectious period,
where nobody is expected to be detected yet. The person has finished the incubation period and may infect
other people (even if he/she is asymptomatic, see, e.g., [61]). After this period, people in this compartment can
be, either taken in charge by health authorities (and we classify them as hospitalized or quarantine, according
to the severity of their symptoms), or not detected by authorities and continue as infectious (but in other
compartment, Iuor IDu).
Infectious but undetected that will survive the disease (denoted by Iu): After being in the compartment I, the
person can still infect other people but is not detected and is not reported by authorities. We assume that
only asymptomatic infected people or those with low or medium symptoms can reach this compartment, not the
people who will die. After this period, people in this compartment survive the disease and pass to the Recovered
compartment Ru.
Infectious but undetected that will die (denoted by IDu): After being in the compartment I, the person can still
infect other people, have clinical signs but not be detected and reported by authorities. We assume that a person
in this state is part of a population that is highly vulnerable to the COVID-19 (such as people in retirement
homes) and will die quickly without being treated in hospital. After being in this state, a person passes to the
compartment Du.
We remark that vulnerable people died outside hospitals during the first wave of the pandemic in some countries
(such as Italy or Spain; see [32, 30, 24, 14, 41]). Indeed, due to the saturation of the whole health system (mainly
5
because of lack of beds and/or healthcare workers), this vulnerable population did not receive adequate medical
cares and quickly died before being hospitalized. In some of those countries (such as Italy, see [32, 30]), a few
a-posteriori estimations of the number of people dead due to COVID-19 and not detected have been reported by
authorities or other organizations and, currently, efforts are made to avoid reproducing this extreme situation
(see [41]). Regarding those few data and for the sake of simplicity of the model, we have decided to model people
who have suffered this situation by using the specific compartment IDu. We note that the results reported later
in Section 4 indicate that this choice is adequate.
Quarantine (denoted by Q): The person is in quarantine at home or some special facilities (different from
hospitals) during a fixed period.
For example, according to [9, 29], the isolation is performed in one room, always wearing mask at home, etc. to
reduce the risk of transmission to cohabiting people. Of course, infection could occur between people in the same
home, but this is usually before entering the quarantine period (i.e., while being in the compartment I) and the
model takes that into account. Furthermore, according to [19], transmission between people in quarantine at
home (thus, aware of their infection) and other person (including cohabiting persons) is severely limited.
Therefore, in this work we assume that a person in compartment Qcannot infect other people (although other
possibilities are also easy to include, as explained below) and at the end of this state passes to the compartment
Rd. The use of this compartment in the model depends on the strategy of each territory (some of them do not
use this control strategy).
Hospitalized that will recover (denoted by HR): The person is in hospital and (according to [19, 51, 58]) can still
infect other people (in particular healthcare workers). At the end of this state, a person passes to the compartment
Q(if implemented in the territory under study; otherwise the person passes directly to the compartment Rd).
Hospitalized that will die (denoted by HD): The person is hospitalized and can still infect other people. At the
end of this state, a person passes to the compartment D.
Dead by COVID-19 (denoted by D) after being at hospital.
Dead by COVID-19 but undetected (denoted by Du): The person has not survived the disease and the death
was not classified by authorities as caused by COVID-19.
Recovered after being previously detected as infectious (denoted by Rd): The person was previously detected as
infectious, survived the disease, passed the quarantine period (if implemented in the territory under study), is
no longer infectious and has developed a natural immunity to the virus. Although some cases of reinfection are
known, following [47], they seem to be isolated cases; thus, despite the fact that the evolution of this immunity
is unknown for a longer period of time, we make the natural immunity assumption. However, it can be modified
if new findings regarding this topic are found.
Recovered after being previously infectious but undetected (denoted by Ru): The person was not previously
detected as infectious, survived the disease, is no longer infectious and has developed a natural immunity to the
virus (with similar assumptions as above).
We remark that, in this work, we work with two compartments, namely HRand HDdescribed previously, to model
the dynamics of the number of people in hospital. One may note that a more detailed classification of hospitalization
could be considered (for instance, according to the severity of the symptoms, the availability of medical material
or beds, etc.) to give more recommendations to decision makers. This could be done by introducing additional
compartments to the model. However, this would require the use of detailed data regarding people in hospital which,
to the best of our knowledge, are not publicly available for most of the countries (in particular, Italy). Furthermore,
adding more compartments will complexify deeply the model, making difficult the parameter estimation process. For
those reasons, we decided to use two compartments. We highlight the fact that the results reported in Section 4 (see
Figure 2) indicate that this choice is reasonable as the model estimates quite well the reported data related to people
in hospital (which are, for the case of Italy, the number of people in hospital, without any specific classification).
6
The authorities may apply various control measures in order to control the COVID-19 spread (see [6]):
Isolation: Infected people are isolated from contact with other people. Only health professionals are in contact
with them. However, infection of those professionals also occurs (see [64]). Isolated patients receive an adequate
medical treatment that reduces the COVID-19 fatality rate.
Lockdown areas: Movement of people in areas with a high number of infected people may be restricted and
controlled, to avoid the spread of the disease.
Contact tracing: The ob jective of tracing is to identify potential infectious contacts which may have infected a
person or spread COVID-19 to other people. Increase the number of tests in order to increase the percentage of
detected infected people.
Increase of health system resources: The number of operational beds and personal available to detect and treat
affected people is increased, producing a decrease in the length of the infectious period for the compartment I.
The model is used to evaluate the spread of a human disease within a considered territory during a fixed time
interval.
At the beginning of the simulation, the model parameters are set by the user (as shown below). We can start
our simulation at any initial time t00, considering that only susceptible people live in the territories that are free
of the disease, whereas the number of people in compartments S,E,I,Iu,IDu,HR,HD,Rd,Ru,Dand Duof the
infected territories are set to their corresponding values. Then, for the time interval [t0, t0+Tmax], with Tmax IN
being the maximum number of simulation days, the model is applied. If at the end of a simulation day tthere is less
than 1 person in each compartment E,I,Iu,IDu,HRand HDin all the territory, the simulation is stopped. Else,
the simulation is stopped when t=t0+Tmax. Furthermore, the control measures are also implemented and they can
be activated or deactivated, when starting the model, in order to quantify their effectiveness to reduce the magnitude
and duration of the COVID-19 pandemic in the territory under study.
A diagram summarizing the main structure of our full model is presented in Figure 1.
We assume that people in a territory are characterized to be in one of the compartments described above: S,E,I,
Iu,IDu,HR,HD,Q,Rd,Ru,Duor D. For the sake of simplicity we assume that, at each time, the population inside
a territory is homogeneously distributed. Thus, the spatial distribution of the epidemic inside a territory is omitted
(it can be taken into account by dividing some territories into a set of smaller regions with similar characteristics).
Since the natural natality and mortality (not from COVID-19) do not seem to be important factors for COVID-19
(at least for relatively short periods of time), they are omitted in the model (actually, they can be readily included, if
necessary).
In this work we will only consider the within–country/territory disease spread of one territory where, starting from
suitable values of t0, the COVID-19 pandemic is already spreading by its own, with a relatively negligible dependence
on the movement of people in/out of this territory. The between–country/territory disease spread may be also taken
into account in the model following the approach developed in [35] with the Be-CoDiS model.
There are some studies about the relevance of humidity and temperature for the spread of COVID-19. In [40], it is
shown that the observed patterns of COVID-19 are not completely consistent with the hypothesis that high absolute
humidity may limit the survival and transmission of this virus. Furthermore, in [56] it has been found that the lower
is the temperature, the greater is the survival period of the SARS-CoV-2 outside the host. Since there is no clear
scientific evidence of the effect of the humidity and the temperature on the SARS-CoV-2, we have not included these
two factors in our model (this would need to be revised in case of appearing new findings regarding this topic).
Under those assumptions, the evolution of the compartments mentioned above is modeled by the following system
7
of ordinary differential equations:
˙
S=S
NβEE+βII+βIuIu+βIDuIDu+βHRHR+βHDHD,
˙
E=S
NβEE+βII+βIuIu+βIDuIDu+βHRHR+βHDHDγEE+τ1τ2,
˙
I=γEEγII,
˙
Iu= (1 θωu)γIIγIuIu,
˙
IDu=ωuγIIγIDuIDu,
˙
HR=p(θω)γIIγHRHR,
˙
HD=ωγIIγHDHD,
˙
Q= (1 p) (θω)γII+γHRHRγQQ,
˙
Rd=γQQ,
˙
Ru=γIuIu,
˙
Du=γIDuIDu.
˙
D=γHDHD.
(1)
In this system we have used, as stated above, that people in compartment Qcannot infect other people. To include
the possibility of infection, we would only need to add the term βQQinside the parenthesis of the first two equations.
In (1):
Nis the number of people in the territory before the start of the pandemic.
ω(t)[0,1] is the instantaneous infection detected fatality ratio (iIdFR) in the territory, at time t: the proportion
of new detected people (after being in compartment I) that will die of COVID-19 after 1
γHD(t)days, per unit
time, compared to the total number of new infectious people (detected or undetected), per unit time. More
precisely,
ω(t) =
˙
D(t+1
γHD(t))
˙cm(t) + ˙cm,u(t)=
γHD(t+1
γHD(t))HD(t+1
γHD(t))
γI(t)I(t),
where cm(t)(respectively, cm,u(t)) is the model cumulative number of COVID-19 detected (respectively, unde-
tected) cases at day t(see more details about them below, when describing the outputs of the model).
ωu(t)[0,1] is the instantaneous infection undetected fatality ratio (iIuFR) in the territory, at time t: the
proportion of new undetected people (after being in compartment I) that will die of COVID-19 after 1
γIDu(t)
days, per unit time, compared to the total number of new infectious people (detected or undetected), per unit
time. More precisely,
ωu(t) =
˙
Du(t+1
γIDu(t))
˙cm(t) + ˙cm,u(t).
8
We point out that ω(t)+ωu(t)[0,1] is the instantaneous infection fatality ratio (iIFR), at time t: the proportion
of new (detected or undetected) people (after being in compartment I) that will die of COVID-19, per unit time,
compared to the total number of new infectious people (detected or undetected), per unit time. This value is
closely related to the IFR defined below (see the Results section). Actually, if it is assumed that the number of
total deaths is proportional to the number of total cases, then IFR is, essentially, ω+ωu.
θ(t)[ω(t),1ωu(t)] [0,1] is the proportion of the number of new infectious people that are detected and
documented by the authorities (after being in compartment I), per unit time, compared to the total number of
new infectious people (detected or undetected), per unit time, at time t. More precisely,
θ(t) = ˙cm(t)
˙cm(t) + ˙cm,u(t).
We point out that
ω(t)
θ(t)=
˙
D(t+1
γHD(t))
˙cm(t)[0,1]
is the instantaneous case fatality ratio (iCFR), at time t: the proportion of the number of new detected people
(after being in compartment I) that will die of COVID-19 after 1
γHD(t)days, per unit time, compared to the
number of new detected people (after being in compartment I), per unit time.
This value is closely related to the Case Fatality Ratio (CFR) defined as the expected value of the probability
of dying for a person who is infected and detected, i.e. the number of detected deaths due to the virus over the
number of detected infections. Actually, if it is assumed that the number of detected deaths is proportional to
the number of detected cases, then CFR is, essentially, ω
θ.
βE, βI, βQ, βIu, βIDu, βHR, βHD[0,)are the disease contact rates (day1) of a person in the respective
compartments. They may change with time, due to the control measures.
γI(t)(0,+)is the transition rate (day1) from compartment Ito compartments Iu,IDu,Q,HRor HD,
at time t. Similarly, γE, γIu(t), γIDu(t), γQ(t), γHR(t),and γHD(t)(0,+)denote the transition rates (day1)
from compartments E,Iu,IDu,Q,HRand HDto compartments I,Ru,Du,Rd,Qand D, respectively, at time
t. For some particular territories, some compartments could be excluded from the model.
pis the ratio of the number of new detected infected people that will survive the disease and are hospitalized
(they go to compartment HR) over the total number of new detected infected people that will survive the disease
(and may also do quarantine, compartment Q, without going to hospital).
τ1(t)and τ2(t)are the people infected that arrive/leave from/to other territories, per day. Both can be modeled
following the between-country spread part of the Be-CoDiS model described in [35]. For the sake of simplicity
we have included those terms only in the second equation, corresponding to compartment E, but they could be
included in the equations for other compartments too. When a discrete sequence of imported/exported infections
is known it can be included in the numerical resolution of the system by simple addition/removal to/from the
suitable compartment, at the corresponding discrete time.
System (1) is completed with initial data S(t0),E(t0),I(t0),Iu(t0),IDu(t0),Q(t0),HR(t0),HD(t0),Rd(t0),
Ru(t0),Du(t0)and D(t0)given in [0,). These values must be non-negative and S(t0) + E(t0) + I(t0) + Iu(t0) +
IDu(t0) + Q(t0) + HR(t0) + HD(t0) + Rd(t0) + Ru(t0) + Du(t0) + D(t0) = N. It is standard (see, e.g., [34], [36]) to
prove that S(t),E(t),I(t),Iu(t),IDu(t),Q(t),HR(t),HD(t),Rd(t),Ru(t),Du(t)and D(t)are non-negative and
S(t) + E(t) + I(t) + Iu(t) + IDu(t) + Q(t) + HR(t) + HD(t) + Rd(t) + Ru(t) + Du(t) + D(t) = N, for all tt0.
We remark that system (1) is an epidemic-type model (as described, e.g., in [22]), which is not trying to capture
endemic behaviors. Actually, there is no incoming flux in the Susceptible compartment; therefore, the fate of every
individual will be either staying susceptible, recovering or dying. This means there are no endemic equilibria in our
system; on the contrary, there exists a manifold S+Rd+Ru+Du+D=Ncontaining infinite disease-free equilibrium
points.
9
We point out that the 8th, 9th, 10th, 11th and 12th equations of system (1) are not coupled with the other
equations. Thus, we can solve the first seven equations of that system and the solution of the last five ones can be
computed as follows: Qis the solution of the 8th equation of system (1), which is a linear ordinary differential equation
(we do not include here the easy corresponding explicit solution, for the sake of simplicity) and
D(t) = D(t0) + Zt
t0
γHD(s)HD(s)ds,
Du(t) = Du(t0) + Zt
t0
γIDu(s)IDu(s)ds,
Rd(t) = Rd(t0) + Zt
t0
γQ(s)Q(s)ds,
Ru(t) = Ru(t0) + Zt
t0
γIu(s)Iu(s)ds.
It is important to remark that, if for some territory the model does not include the compartment Q, then p= 1,
˙
Rd=γHRHRand
Rd(t) = Rd(t0) + Zt
t0
γHR(s)HR(s)ds.
For the numerical simulations presented in this paper, the first eight equations of system (1) were numerically
solved with the classic 4th order Runge–Kutta method with four stages (RK4) and with four hours as time step (which
was tested to see that it is suitable to get stable results).
2.2 Outputs of the model
Here, we present some of the outputs that can be obtained from the mathematical model, used to analyze the results
of the simulations:
cm(t): The model cumulative number of COVID-19 detected cases at day t, which is given by
cm(t) = HR(t) + HD(t) + Q(t) + Rd(t) + D(t),
and can be also computed as follows:
cm(t) = cm(t0) + Zt
t0
d
dtHR(s) + HD(s) + Q(s) + Rd(s) + D(s)ds=cm(t0) + Zt
t0
θ(s)γI(s)I(s)ds.
cm,n(t): The number of new cases per day at day t, which is given by
cm,n(t) = θ(t)γI(t)I(t).
hm(t): The model cumulative number of COVID-19 detected infected healthcare workers at day t, which is given
by
hm(t) = hm(t0) + Zt
t0
S(s)
NβHR(s)HR(s) + βHD(s)HD(s)ds.
cm,u(t): The model cumulative number of COVID-19 undetected cases at day t, which is given by
cm,u(t) = E(t) + I(t) + Iu(t) + IDu(t) + Ru(t) + Du(t),
10
and can be also computed as follows:
cm,u(t) = E(t) + I(t) + Iu(t0) + IDu(t0) + Ru(t0) + Du(t0) + Zt
t0
d
dtIu(s) + IDu(s) + Ru(s) + Du(s)ds
=E(t) + I(t) + Iu(t0) + IDu(t0) + Ru(t0) + Du(t0) + Zt
t0
(1 θ(s))γI(s)I(s)ds.
dm(t): The model cumulative number of detected deaths (due to COVID-19), at day t, which is given by D(t).
dm,u(t): The model cumulative number of undetected deaths, at day t, which is given by Du(t).
Re: The effective reproduction number of COVID-19. It is defined as the number of cases that one infected person
generates, on average, over the course of its infectious period [5]. Part of the population can be already infected
and/or control measures may have been implemented. It depends on the considered territory and changes during
the spread of the disease. Furthermore, if at time t=0 no control measures have been applied, Re(0) = R0, where
R0is the basic reproduction number of COVID-19, which is defined as the number of cases that one infected
person generates on average over the course of its infectious period, in an otherwise uninfected population and
without special control measures. Typically, the spread of the disease slows down when Re(t)<1. Taking into
account this definition, for cases with constant transition rates we propose the following non-classical estimation:
Re(t) = Rt+1
γE
tβE(τ)S(τ)
Ndτ+Rt+1
γE+1
γI
t+1
γE
βI(τ)S(τ)
Ndτ
+1θ(t+1
γE+1
γI)ωu(t+1
γE+1
γI)Rt+1
γE+1
γI+1
γIu
t+1
γE+1
γI
βIu(τ)S(τ)
Ndτ
+ωu(t+1
γE+1
γI)Rt+1
γE+1
γI+1
γIDu
t+1
γE+1
γI
βIDu(τ)S(τ)
Ndτ
+p(t+1
γE+1
γI)θ(t+1
γE+1
γI)ω(t+1
γE+1
γI)Rt+1
γE+1
γI+1
γHR
t+1
γE+1
γI
βHR(τ)S(τ)
Ndτ
+ω(t+1
γE+1
γI)Rt+1
γE+1
γI+1
γHD
t+1
γE+1
γI
βHD(τ)S(τ)
Ndτ.
(2)
In this formula, we estimate the mean number of secondary cases produced by one person infected at time t,
during its whole infectiousness period, taking into account all the possible evolutions of its health status.
Composite rectangular, trapezoidal or Simpson rules can be used to obtain an estimation of (2).
We remark that other estimations for R0can be also considered. For instance, one can use the so-called "Next
generation Matrix" method to approximate its value (see [54]). However, as pointed in [7], this approximation
does not correspond accurately to the biological definition of R0(e.g., parameters are assumed to be constant
and the change of the proportion of susceptible people during the infectious time of a person is omitted). We
remark that Reshould be used with caution by policymakers, without relying only on that value (see [1]).
Note that we can estimate the particular contribution of each infectious compartment (i.e. E,I,Iu,IDu,HR
and HD) to the value of Reand R0by computing only their respective integrals in (2).
Hos(t): The number of people in hospital is estimated as follows:
Hos(t) = HD(t) + HR(t).
This function can help to estimate and plan the number of clinical beds needed to treat all the COVID-19 cases.
Hosn(t): The number of new hospitalized people per day at day t, which is given by
Hosn(t) = [p(θ(t)ω(t)) + ω(t)] γI(t)I(t).
11
MHos: The maximum number of hospitalized people at the same time (in territory i) during the time interval
[t0, T ]. It is computed as:
MHos = max
t[t0,T ]Hos(t).
This number can help to estimate and plan the number of clinical beds needed to treat all the COVID-19 cases.
ΓE(t)and ΓIu(t): The number of people infected during the time interval [t0, T ], by contact with people in
compartment Eand Iu, respectively. They are given by:
ΓE(t) = Zt
t0
βE(s)E(s)S(s)
Nds,
ΓIu(t) = Zt
t0
βIu(s)Iu(s)S(s)
Nds.
We point out that those outputs and others that can be computed similarly can be compared with reported data
when available.
3 Model parameter estimation for COVID-19
Here, we present the methodology we used to estimate the parameters of the model described in Section 2.1. We detail
each kind of parameter according to its category.
3.1 Parametrization of some of the functions involved in the model
The model involves several functions. We now detail how some of these functions can be written, using empirical
assumptions, in terms of constant parameters, which will have to be later identified for each territory under study
(we will do it below for the case of Italy). Of course, other alternatives may be also used, taking into account the
particularities of each territory.
-Disease contact rates: We consider that
βE(t) = mE(t)βE,0, βI(t) = mI(t)βI ,0, βIu(t) = mIu(t)βIu,0(θ(t)), βIDu(t) = mIDu(t)βIDu,0,
βHR(t) = mHR(t)βHR,0, βHD(t) = mHD(t)βHD,0.
Here βE,0, βI ,0, βIu,0, βIDu,0, βHR,0, βHD,0are the respective disease contact rates when no control measures are taken
and mE, mI, mIu, mIDu, mHR, mHDare functions simulating de control measures, as described below.
Regarding βIu,0, we may consider
βIu,0(θ) =
βI,0, if θ[0, ω],
nonincreasing , if θ(ω, 1),
βI,0, if θ= 1,
where ωis a fatality rate value that is explained below and βIand βIare suitable lower and upper bounds, respectively.
For the sake of simplicity, here we take βI,0=βI,0.
Furthermore, following the idea proposed in [33, 35, 39], we assume that there exists a relationship between the
contact rates βI,0,βE,0,βIDu,0and βI ,0.
More precisely, following [33], we consider that people in compartments E,Iuand IDuare less infectious than
people in compartment I(due to their probably lower virus load or isolation measures, see [2, 39, 61]). Thus, we
consider that βE,0=CEβI ,0and βI,0=CuβI ,0and βIDu,0=CIDu,0βI,0, where CE, Cu, CIDu[0,1].
12
A similar approach can be followed for βHRand βHDassuming, for instance, that βHR,0=βHD,0=CHβI,0.
Alternatively (as we do here for the case of Italy) we can compute good estimates for βHRand βHD,0using historical
data, as shown below.
-Control measures: We may distinguish three different types of control measures: Social distancing, contact tracing
and health system measures. Each of them may be included in the model by using different functions, as explained
below.
Functions mE(t),mI(t),mIu(t)[0,1] (%) represent the reduction of some of the disease contact rates, at time
t, due to the efficiency of the control measures corresponding to social distancing applied to the compartments E,I
and Iu, respectively (see, e.g., [46]). Here, for the sake of simplicity, we consider that those three functions are equal
(of course, other alternatives are also possible) and given by (see [33])
mE(t) = mIu(t) = mI(t) = m(t) =
(m0m1) exp κ1(tt0)+m1t[t0, λ1],
(m(λ1)m2) exp κ2(tλ1)+m2t(λ1, λ2],
.
.
.
(m(λq1)mq) exp κq(tλq1)+mqt(λq1,).
We may also define in a similar way functions m(t) and m(s) related with contact tracing and health system
control measures, respectively, which include similar parameters denoted by m(t)
i, m(s)
i, κ(t)
i, κ(s)
i, q(t) and q(s). Here,
q, q(t) , q(s) Nare the number of changes of control strategies.
Here, for every i∈ {0, ..., q},j∈ {0, ..., q(s) }and k∈ {0, ..., q(t) }, constants mi, m(s)
j, m(t)
k[0,1] measure the
intensity of the control measures (lower value implies higher intensity), κi, κ(s)
j, κ(t)
t[0,+)(day1) simulate the
efficiency of the control measures (higher value implies faster effectiveness of the control measures) and λi, λ(s)
j, λ(t)
k
[t0,]denote days of changes of each control strategy.
The values of λi, λ(s)
jare typically taken from the literature (using the dates when the territories implement special
control measures). Some of the values of mi, m(s)
jcan be also known. The rest of the parameters need to be identified,
as explained below. When doing this, to ease the search process, we use new parameters ciinstead of mi, as follows:
If we want mi+1 mi, then we look for a suitable parameter ci+1 [0,1] such that mi+1 =ci+1mi.
If we want mi+1 mi, then we look for a suitable parameter ci+1 [0,1] such that mi+1 =mi+ci+1(1 mi).
A similar approach can be followed by using new parameters c(s)
iinstead of m(s)
i.
Function θis related with the contact tracing measures, since the higher those measures are at time t, the closer
θ(t)is to 1. The other function involved with the contact tracing measures is m(t), which is used below for estimating
the transition rates. The function m(s) is used to compute the fatality ratios and to define mHR(t) = mHD(t) =
m(s)(t)[0,1], representing the reduction of the disease contact rates βHR,0, βHD,0, at time t, due to the efficiency
of the health system measures applied to the corresponding compartments. We note that this approach is not taken
into account when using the alternative mentioned above to compute estimates for βHRand βHDusing historical data
(which will be shown below).
-Fatality ratios: As observed in other epidemics (see, e.g. [35]), ω(t)can be affected by the application of health
system measures. Thus, following [33], we propose to consider
ω(t) = m(s)(t)ω+ (1 m(s) (t))ω,
where ω[0,1] is the iIdFR when no health system measures are applied (i.e. m(s) = 1); and ω[0,1] is the iIdFR
when the implemented health system measures are fully applied (i.e. m(s) = 0). We note that ωω, thus, we consider
13
ω=ω+δω, with δω[0,1ω]. Alternatively (as we do here for the case of Italy), these constants can be estimated
from values of the IFR in the literature (see, e.g., [3, 53]).
Additionally, we assume that
ωu(t) =
ωu,0, if t=t0
ωu,1, if t=tu,1
.
.
.
ωu,n, if t=tu,n
constant after the last value tu,n
linear otherwise,
where nN∪ {0}and the corresponding values ti,ωu,iare parameters to be chosen for each territory, depending on
its particular circumstances.
-Transition rates:
According to [39], the transition rate γEfrom Eto Idepends only on the disease and, therefore, is considered
constant.
Some other transition rates could depend on time. For instance, according to [39, 38], depending on the studied
territory, the value of γI(t)can be increased due to the application of contact tracing control measures (i.e.
people with symptoms are detected earlier). As a consequence, the values of γIu(t),γIDu(t),γQ(t),γHR(t)and
γHD(t)can be decreased (e.g. people with symptoms stay under observation during more time). In order to
parametrize these functions, we denote by dI,dIu,dIDu,dQ,dHRand dHDthe initial duration (given as a suitable
average) in days of a person in compartment I,Iu,IDu,Q,HRand HD, respectively, without the application of
special contact tracing measures. Then, following [35], we may use the following functions:
γI(t) = 1
dIg(t)(day1), γIu(t) = 1
dIu+g(t)(day1), γIDu(t) = 1
dIDu+g(t)(day1),
γQ(t) = 1
dQ+g(t)(day1), γHR(t) = 1
dHR+g(t)(day1), γHD(t) = 1
dHD+g(t)(day1),
where g(t) = dg(1 m(t)(t)) represents the decrease of the duration of dIdue to the application of contact
tracing measures, at time t; and dgis the maximum number of days that dIcan be decreased.
We remark that, for some territories, it can be suitable to consider constant (i.e. not dependent on time) all the
transition rates, which corresponds to the case with dg= 0.
-Ratio p: We describe it by using the following function depending on the value of θω(we remark that other
alternatives are also possible):
p=
p0(θ0ω0)
θω, if θωθ0ω0,
11p0
θ0ω0
(θω), if θωθ0ω0.
Here p0=p(tθ0)and ω0=ω(tθ0), where tθ0and θ0are described below (when we show how to estimate θ(t)). Indeed,
we assume that, if θω > θ0ω0all new detected infected people, when comparing to the case of θ0and ω0, will
go to state Q, as they should present light symptoms. Furthermore, we use a linear interpolation between the case
θω= 0 (with p= 1, i.e., only persons with severe symptoms are detected) and θω=θ0ω0(with p=p0). More
complex formulas could be also considered, but we use the previous one, for the sake of simplicity.
We remark that one of the functions involved in the model, namely θ(t), has not been parameterized above. We will
show below how to deal with it.
14
3.2 Estimation of parameters and non-parameterized functions involved in the model
Once the model is built, if we want to use it for a particular country or territory, we need to identify suitable model
parameters and functions for the case under study. Some of them can be found in the literature. However, due to lack
of information regarding the behavior of the SARS-CoV-2 and implicit uncertainty, others need to be estimated.
To do that, we need as much information as possible about the dynamics of the pandemic in the studied territory.
In particular, we may use the following temporal series of the territory under study, if available:
cr(t): The reported cumulative number of COVID-19 infected, cases at time t.
dr(t): The reported cumulative number of COVID-19 deaths, at time t.
hr(t): The reported cumulative number of COVID-19 infected healthcare workers, at time t.
τ1,r(t): The reported number of known imported cases on day t. More precisely, infected people arriving from
an external territory. Usually they are in the incubation period, but, in some cases, they are moving between
hospitals of different countries or territories (this last possibility is not included in the equations, for the sake of
simplicity, but it can be easily done).
τ2,r(t): The reported number of known exported cases on day t. More precisely, infected people going to an
external territory. Usually, they are in the incubation period, but sometimes they are moving between hospitals
of different countries or territories (this last possibility is not included in the equations, for the sake of simplicity,
but it can be easily done).
dr,u(t): The reported cumulative number of estimated undetected COVID-19 deaths, at time t.
Hosr(t): The reported number of people in hospital (i.e. Hosr), at time t.
qr(t): The number of people in quarantine at home, due to their infection with COVID-19.
We note that, sometimes, there is only partial information and the gaps can be filled with suitable techniques (see,
e.g., [17]).
Now, with that information we may estimate some of the model parameters and functions, as follows:
-Ratio θ(t)of new infected people that are detected: To estimate θ(t), first, we compute an estimation of iCFR ( ω(t)
θ(t))
as follows:
ω(t)
θ(t)=
dr(tiCFR +1
γHD
)
cr(tiCFR), if ttiCFR ,
dr(t+1
γHD
)dr(tr+1
γHD
)
cr(t)cr(tr), if tiCFR < t < Tmax 1
γHD
, r is the minimum natural number such that
cr(t)cr(tr)6= 0 and this quotient is 1,
1, if tiCFR < t < Tmax 1
γHD
, r is the minimum natural number such that
cr(t)cr(tr)6= 0 and the previous quotient is >1,
linear interpolation , otherwise,
where tiCFR 7days is the first date for which cr(t)cr(t1) 6= 0 and dr(t+1
γHD
)6= 0. We note that, in this
formula, we estimate the iCFR by taking into account the delay between the detection of a case and its death (instead
of the ratio between cases and deaths reported at the same date). This delay could also be considered in the system of
differential equations, instead of the exponential-decay-formulation, leading to a system of delay differential equations
(see, e.g., [23]).
15
After that, to regularize the obtained iCFR curve, we apply a filter that consists in computing the average value
of the last week. We denote the resulting filtered curve by ωCFR(t):
ωCFR(t) =
ω(tiCFR)
θ(tiCFR), if ttiCFR ,
1
7P6
i=0
ω(ti)
θ(ti), if t>tiCFR.
(3)
To avoid unrealistic situations (due to poor data), if ωCFR (t)0.01 (which is assumed too low, see, e.g., [53]) for
some t > tiCFR, then we extend the average of the last week used to compute this ωCFR(t)by including the previous
days until getting ωCFR(t)>0.01.
Finally, we define θ(t)by:
θ(t) =
θ0, if ttθ0,
ω(t)
ωCFR(t), if t>tθ0,
where θ0=ω(tθ0)
ωCFR(tθ0)and tθ0=tiCFR + 6 days to guarantee that ωCFR(tθ0)is computed by using the average value
defined by the second line of (3), for the sake of smoothness. Notice that, if we simulate until the current date, we do
not have enough data to compute ω(t)
θ(t). In this case, we set θ(t)to a constant value for the last days of the simulation
without data. When simulating scenarios for future dates, we can also keep the last value of θor use alternative
possible values.
For particular territories (as we do here for the case of Italy) we can use different values of IFR found in the
literature and use them as ωin the previous formula. That will provide different estimations for θ(t).
-Contact rates βHR(t)and βHD(t): We use the following estimation of the fraction of infected healthcare workers
(assumed to be due to contact with hospitalized people, see, e.g., [58]):
p(θω)βHR
γHR
+ωβHD
γHD
p(θω)βHR
γHR
+ωβHD
γHD+βE
γE+βI
γI+ (1 θωu)βIu
γIu+ωu
βIDu
γIDu
=η.
To compute η, we first define eηas
eη(t) =
hrtη+1
γE+1
γI
crtη+1
γE+1
γI, if ttη,
hrt+1
γE+1
γIhrtr+1
γE+1
γI
crt+1
γE+1
γIcrtr+1
γE+1
γI, if tη< t < Tmax 1
γE1
γI;ris the minimum natural number
such that hrt+1
γE+1
γIand hrtr+1
γE+1
γIare known
and crt+1
γE+1
γIcrtr+1
γE+1
γI6= 0;
and this quotient is 1,
1, under the same conditions as in the previous line but when
the quotient is >1,
linear interpolation , otherwise,
16
and tηis the first date for which hrtη+1
γE+1
γI6= 0 is reported. We point out that, in the previous formula, we
have made a distinction about the knowledge of the values of hr, because they are rarely given every day. We remark
that we are taking into account the delay of 1
γE+1
γIdays between the infection of a person and its detection by
authorities. Thus, we set eη(t)to a constant value for the last 1
γE+1
γIdays. Finally, to avoid unrealistic situations
(due to poor data), we take the following seven-day average value:
η(t) =
1
7P6
i=0 eη(t0+i), if t<t0+ 3,
1
7P3
i=3eη(t+i), if t0+ 3 tt0+Tmax 3,
1
7P6
i=0 eη(t0+Tmax i), if t>t0+Tmax 3.
(4)
Assuming, for the sake of simplicity, that βHR=βHD, we have that
βHR=βHD=
ηβE
γE+βI
γI+ (1 θωu)βIu
γIu+ωu
βIDu
γIDu
(1 η)p(θω)1
γHR
+ω1
γHD.
In the previous expression, the coefficients depend on time t.
-Other unknown parameters: We can identify them with a multiobjective technique, as described below.
3.3 Identification of some parameters with a multiobjective optimization technique
First, we determine the set of unknown model parameters to be identified by the approach presented here (which will
depend on the available information in the considered territory). For each of those parameters, we establish a suitable
interval of possible values (given by upper and lower bounds of each parameter) and denote by SRnthe feasible
region given by the cartesian product of all those intervals, where nis the number of unknown parameters.
Once Sis defined, we follow the methodology proposed in [15]. This approach aims at minimizing the differences
between the temporal series reported by the authorities and the corresponding ones obtained as model outputs when
considering particular values S. Then, given two weight coefficients ρ1>0and ρ2>0and a temporal series X,
we define the function
fX(Ω) =
XrX
m
2
L2(t0,t0+Tmax)+ρ1
max(XrX
m,0)
2
L2(t0,t0+Tmax)
kXrk2
L2(t0,t0+Tmax)
+ρ2max(Xr(t0+Tmax)X
m(t0+Tmax),0)2
|Xr(t0+Tmax)|2,
(5)
that measures the error between the reported data, which is denoted by Xr, and the model output denoted by X
m.
Notice that, as done in previous works [15, 33], it is based on the relative error. However, as a novelty, we add
two new terms to penalize those obtaining a temporal series that underestimate the reported data. We recall
that, to reduce the risk of over-fitting to the reported data [50], we use the L2norm in [t0, t0+Tmax], which is
kgkL2([t0,t0+Tmax]) =Rt0+Tmax
t0|g(t)|2dt1/2.
The number of functions fX(Ω) depends on the number of available temporal series of the territory under study.
The authorities usually report at least the cumulative daily number of cases and deaths (i.e. crand dr), see, e.g., [62].
In some cases, they also include other data as, for instance, the number of people in hospital (i.e. Hosr), in quarantine
at home (i.e. qr), or the cumulative number of undetected deaths (i.e. dr,u) (see, e.g., [44]).
17
Then, we formulate a multiobjective problem that may have as many objective functions as available temporal
series. For instance, we could solve
(min fc(Ω), fd(Ω), fHos(Ω), fdu(Ω)
subject to SRn,(6)
where Sis the feasible region (i.e. upper and lower bounds of each parameter) and nis the number of parameters in
. We can also add together (with suitable weights) several functions, in order to reduce the number of functions to
be minimized (and the complexity of the corresponding multiobjective problem).
Next, to solve this multiobjective optimization problem, the algorithm called Weighting Achievement Scalarizing
Function Genetic Algorithm (WASF-GA) is applied [16, 17, 52]. We choose this particular multiobjective algorithm as
it has been applied with success in a previous work to identify some parameters of an epidemiological model simulating
the evolution of Ebola Virus Disease outbreaks in Democratic Republic of the Congo [17].
From a general point of view, WASF-GA is an evolutionary algorithm that manages a population of possible
solutions (called individuals), trying to improve them iteratively towards the optimal set (called Pareto optimal set).
To do that, it uses two procedures at each iteration: reproduction and replacement. Firstly, the reproduction method
applies some classical genetic operators (here: selection, mutation, and crossover) to the population of the previous
iteration to obtain a new offspring population. Secondly, the replacement procedure decides which solutions from
the previous iteration and the offspring populations should compose the population for the next iteration. The main
feature of WASF-GA is that this decision is based on preferences, which are defined as reference values for the
objectives. Indeed, it classifies the individuals using an achievement scalarizing function (see [52]) characterized by
the Ldistance between the objective function values and the reference values with different weights. The considered
sample of weight vectors is built following [52] to ensure they are well distributed.
As a result, WASF-GA returns a set of solutions covering a region of the Pareto optimal front.
In this work, we set the number of iterations to 15000, the population size to 50 individuals, the number of weight
vectors to 50, and the reference values for the objectives to zero, corresponding to the ideal situation of a total matching
between the model outputs and the reported data. The remaining input parameters of WASF-GA are those reported
in [52]. Those values were set experimentally in order to obtain a good ratio between computational time and precision
of the results.
Since WASF-GA is a stochastic algorithm, it may return different Pareto front approximations for different exe-
cutions of the same instance. Then, to guarantee that results do not depend on the stochasticity of the algorithm,
we perform 60 runs of each experiment. Among all the 60 sets of solutions, we extract the most balanced solution,
meaning the one whose objective values are the closest to the reference values (in our case, zero) when considering the
Euclidean distance.
In the 8th, 9th, 10th, 11th and 12th equations of system 1, which, as said previously, are not coupled with the
other equations, there is a parameter, namely γQ, that only appears in those equations. Therefore, its estimation can
be done once the multiobjective process is over. For instance, if the authorities report the time series corresponding to
the number of people in quarantine at home (i.e. qr), the parameter γQcan be estimated separately as a final step of
the identification methodology. To do that, we assume that γQcan be parameterized by n0Nparameters denoted
by Φ = {Φ1,· · · ,Φn0} ∈ S0Rn0(see next section for a particular example). Then, we formulate the following
mono-objective problem: (min fq(Φ)
subject to ΦS0,(7)
where fq(Φ) measures the error (as defined by (5)) between the reported data qrand qΦ
m, the number of people in
quarantine returned by the model when considering γQdescribed by Φ.
Here, we solve this mono-objective problem by using a genetic algorithm configured with the same selection,
mutation, and crossover operators as WASF-GA and the same input parameters (i.e. we set the number of iterations
to 15000 and the population size to 50 individuals). Furthermore, we also perform 30 runs to guarantee that the
results do not depend on the stochasticity of the algorithm.
18
4 Numerical simulations and calibration of the model for the case of Italy
We present here the results obtained with our model for the case of Italy, as an example of application. The situation
in that country, like in other European countries such as Spain, the United Kingdom and France, has been especially
complicated since it has been strongly affected by the disease. The first two known cases of Coronavirus in Italy, a
couple of Chinese tourists, were confirmed on January 30. The first known case of secondary transmission occurred
in Codogno, Municipality of Lombardy, on February 18, 2020. After that, the pandemic grew and extended to other
countries, which needed the establishment of strict control measures (see [42]).
In Section 4.1, we introduce the data and some model parameters specific to the case of Italy. Then, in Section
4.2, we show and analyze some results returned after the numerical experiments performed to illustrate the efficiency
of our approach.
4.1 Considered data and model parameters
The information regarding the first cases in Italy is available in [42]. More precisely, the first two known cases
of Coronavirus in Italy, a couple of Chinese tourists, were confirmed on January 30. On February 26, they were
declared recovered. The first known case of secondary transmission occurred in Codogno, Municipality of Lombardy,
on February 18, 2020. Thus, in the numerical simulations, we assume that the index case started his incubation period
in Italy on t0=19 January 2020, as the earliest symptom onset of confirmed patients can be traced back to 30 January
2020 (reported on 31 January 2020) and we assume a delay of dI+dE=10.5 days (rounded to 11 days) between the
beginning of the incubation period and the symptoms onset. Thus, we set, E(t0) = 1,S(t0) = N1and all other
compartments are set to 0.
The different control measures implemented in Italy are detailed in [42]. They started on λ1=23 February 2020,
where a ’red zone’ was instituted on 11 municipalities with local transmission of SARS-CoV-2 infection, with lockdown
to contain the emerging threat. Before this date we set m0=m1= 1 and after this date we consider m2[0,1] (to
be estimated). On λ2=11 March 2020, lockdown was declared on the entire country. Thus, we assume m3=c3×m2
with c3[0,1]. On λ3=22 March 2020 new additional measures regarding the containment and management of the
epidemiological emergency from COVID-19, applicable on the whole national territory were adopted including the
closure of non-essential or strategic production activities until May 3 (after several extensions). Consequently, we
set m4= 0, corresponding to the most restrictive social distancing control measures. Starting on λ4= 4 May 2020,
some production activities are authorized and we assume m5=c5×m2with c5[0,1]. The Decree of 16 May 2020
establishes the end of regional travel restrictions and the reopening of production activities from λ5=18 May 2020.
We set m6=m5+c6×(1 m5)with c6[0,1]. Finally, this Decree also governs the end of total travel restrictions
(traveling to and from abroad can only be limited with measures derived from the legal system of the European Union
and international obligations) and the reopening of production, commercial and social activities from λ6=3 June to
31 July ([43]). Hence, we consider m7=m6+c7×(1 m6)with c7[0,1].
For Italy we do not include the compartment IDu(i.e. people who die undetected pass from compartment Ito
compartment Du). This simplification is done since, according to [31], almost all undetected deaths affect to vulnerable
persons that die quickly after developing clinical sign (e.g. aged persons in nursing homes). All the other compartments
were used in this case study.
Furthermore, according to [26], noticeable changes in the delay of detection of infected persons (i.e. dI) have not
been observed during the course of the epidemic in Italy. Thus, we set dg= 0.
The information by regions concerning the number of cases (i.e. people who enter each day in compartments HR
or HD) and COVID-19 deaths per day is daily reported in the website of the Italian Ministry of Health, (see [44]).
Moreover, the Italian National Institute of Health supervises daily the data related to cases, deaths, tests carried out,
etc. and the Italian civil protection publishes them completely here:
https://github.com/pcm-dpc/COVID-19
In our simulations, we take cr(t),dr(t),Hosr(t),qr(t)from the previous github repository, hr(t)from [28] and τ1,r(t),
τ2,r(t)from [62].
19
We use the multiobjective technique described above to identify the parameters
Ω=(βI,0, κ2, κ3, κ4, κ5, κ6, κ7, c2, c3, c5, c6, c7, p0, Cu, ωu,0,1HR)R16 ,
taking ρ1= 10 and ρ2= 100. We reduce the number of objective functions to two functions, f1(Ω) and f2(Ω) that
combine the four time-series errors fc(Ω), fd(Ω), fHos(Ω) and fdu(Ω), using some weights as follows:
f1(Ω) = 0.5fc(Ω) + 0.5fHos(Ω) and f2(Ω) = 0.9fd(Ω) + 0.1fdu(Ω).(8)
Thus, we solve the following two-objective problem:
(min f1(Ω), f2(Ω)
subject to SR16 ,(9)
instead of solving a four objective one, that is more complex and time-consuming. We use ρ1= 10 (penalization to avoid
the underestimation of the reported data in [t0, t0+Tmax)) and ρ2= 100 (penalization to avoid the underestimation
of the reported data at t0+Tmax).
Regarding the data denoted by qr(t), reported by the Italian authorities, after several numerical experiments, we
decided to define γQ(t)as follows:
γQ(t) =
γQ,1, if ttγQ,
γQ,2, if t>tγQ.
Next, we solve the mono-objective problem described by (7) to identify the parameters Φ=(γQ,1, γQ,2, tγQ)
[0,1] ×[0,1] ×[0,+).
The effective reproduction number given by (2) has been computed using the Composite Simpson’s rule.
4.2 Results
First, we use the methodology and techniques detailed in Sections 2 and 3 to identify suitable model parameters for
the case of Italy, validating the model proposed here. Then, we show how the model may be used to estimate the
spread of the disease for different futures scenarios.
4.2.1 Validation of the model.
There is much uncertainty about θbut, as explained in the Section 2.1 , we can first estimate ω(t)
θ(t)and then θ(t)for
different values of ω(which is the instantaneous infection detected fatality ratio, iIdFR). That shows that, actually,
the uncertainty of θis related to the uncertainty of ω. Hence, we have considered different scenarios for the value of
ω, using several estimations found in the literature for the Infection Fatality Ratio (IFR). We remind that the IFR is
the expected value of the probability of dying for a person who is infected, i.e. the number of deaths due to the virus
(detected or not) over the number of infections (detected or not). The IFR is one of the most critical features of the
coronavirus disease, being quite difficult to estimate. More precisely:
According to [55], published on 4 May 2020, IFR=1.2%. Then, since on 5 May 2020 there were 29315 detected
deaths (see [13]) and 9154 estimated undetected deaths (see [31, 32]), an associated estimation of ωwould be
ω= 1.2% 29315
29315+9154 = 0.9145%.
According to [21], published on 18 May 2020, IFR=1.91%. Then, an associated estimation of ωwould be
ω= 1.91% 29315
29315+9154 = 1.4555%.
According to [11], published on 21 May 2020 but received on 4 April 2020, IFR=5.7%. Then, since on 4 April
2020 there were 15362 detected deaths and on 14 April 2020 there were 6773 estimated undetected deaths (see
[30, 32]), an associated estimation of ωwould be ω= 5.7% 15362
15362+6773 = 3.9559%.
20
We have used these three values of ωup to 21 July 2020.
The whole set of model parameters obtained when considering ω= 1.4555% is reported in Tables 1 and 2. For the
sake of simplicity, we only present here the parameters obtained with this particular value. Results for other values ω
can be found at https://github.com/momat-ucm/T-SIR-T.
In Figures 2 and 3, we show the comparison of some of the outputs of the simulation run with the chosen parameters
and the official real data, when ω= 1.4555%. Similar figures have been obtained when considering ω= 0.9145% and
3.9559% (obviously, with different parameters) and, thus, are not reported here.
More precisely, in Figure 2, we show a very good fitting of the model outcomes with respect to four important time
series reported by the Italian authorities regarding the pandemic: the cumulative number of cases and deaths, people
in hospital, people in quarantine and recovered people. Furthermore, the model estimates that on 14 April 2020 and 5
May 2020 there were around 7509 and 9192 undetected deaths (persons in Du), respectively. According to the Italian
authorities (see [32, 31]), on 14 April 2020 and 5 May 2020 (the only two reported values) this number was estimated
to be 6773 and 9154, respectively. This shows that our model is able to estimate well the biological behavior of the
pandemic.
Next, in Figure 3 (Top-Left) and (Top-Right), we show the evolution of the new detected cases and deaths, per
day. When available, we compare those outputs with the real data. Again we observe a good fitting between simulated
results an observed data. This feature of the model is interesting as it allows to estimate the date and magnitude of
the peaks of all those outcomes (also reported in this figure), which may help to plan some health system measures
(such as increasing the number of beds in hospital).
In Figure 3 (Bottom-Left) and (Bottom-Right), we show the evolution of the functions corresponding to the
social distancing measures and the effective reproduction number, when ω= 1.4555%. We observe that the effective
reproduction number went below 1 after 14 March 2020, which should indicate the start of a decrease of the spread
in Italy. This value is reached soon after the implementation of some intensive control measures (lockdown in the
entire country) in Italy on 11 March 2020, showing that such restrictive control measures were necessary to control
the pandemic (before this date the model estimates that Reremains above 1). We also note that the initial value of
the effective reproduction number, which is the basic reproduction number R0, is equal to 8.0, which is a value higher
than those generally reported in the literature (between 1.4 and 3.9, see e.g. [8, 39]). However, we also show in this
figure the contribution of each infectious state (i.e. E,I,Iu,HRand HD) to the value of Re(see Section 2.2 for more
details). In particular, we observe that the compartment Iuof undetected infectious people contributes to R0with a
value of 3.9, which means that detected cases contribute to R0with a value of less than 4.1, which is consistent with
the values usually found in the literature.
Next, in Figure 4, we show the evolution of some of the outcomes of the model for different values of ω. First, in
this figure (Top-Left) and (Top-Right), we present the evolution of Reand the effect of the social distancing measures
when considering ω= 0.9145%,ω= 1.4555% and 3.9559%. We observe that, before 11 March 2020, the magnitude and
shape of those functions are affected by the value of ω. After this date, they are quite similar for the different values
of ω. Focusing on the total number of cases (including undetected), presented in Figure 4 (Bottom-Left), the model
estimates that, taking into account the different values of ωconsidered previously, between 1 and 4 million people
could have been infected by COVID-19 in Italy. Regarding the value of θ, depicted in Figure 4 (Bottom-Right), we
see that for ω= 3.9559% the proportion of detection of new cases oscillates between 15% and 67%. For ω= 0.9145%
or ω= 1.4555%, it is between 5.7% and 25%. From those results we see that, in all cases, the model estimates that
a large amount of new cases have not been detected during this pandemic. However, the proportion of detection is
highly dependent on the value of ωand large variations of the results are observed when this value changes. The
results presented in Figure 4, indicate the necessity to approximate well this critical parameter.
4.2.2 Estimation of possible future scenarios
We considered just a few cases, to illustrate the possibilities that can be explored with the model to estimate the
possible spread of the disease. We simulate the future scenarios described below until 1 September 2020. We recall
that the meaning of parameters λi,κiand mithat appear below may be seen in Section 2.1. In all cases, we consider
λ7= 22 July 2020 (date of a potential change of control measures; taking into account that the date of our last data
is 21 July 2020), κ8=κ2= 62.0015. Since we cannot estimate θfor the future (it depends on the authority actions)
21
19 Jan 2020
08 Feb 2020
28 Feb 2020
20 Mar 2020
09 Apr 2020
30 Apr 2020
20 May 2020
10 Jun 2020
30 Jun 2020
21 Jul 2020
0
0.5
1
1.5
2
Cumulative number of people
105
Rep. Cases
Rep. Deaths
Est. Cases
Est. Deaths
19 Jan 2020
08 Feb 2020
28 Feb 2020
20 Mar 2020
09 Apr 2020
30 Apr 2020
20 May 2020
10 Jun 2020
30 Jun 2020
21 Jul 2020
0
0.5
1
1.5
2
2.5
3
Number of people
104
05-Apr-2020
Rep. Hos
Est. Hos
19 Jan 2020
08 Feb 2020
28 Feb 2020
20 Mar 2020
09 Apr 2020
30 Apr 2020
20 May 2020
10 Jun 2020
30 Jun 2020
21 Jul 2020
0
1
2
3
4
5
6
7
8
Number of people
104
Rep. Q
Est. Q
19 Jan 2020
08 Feb 2020
28 Feb 2020
20 Mar 2020
09 Apr 2020
30 Apr 2020
20 May 2020
10 Jun 2020
30 Jun 2020
21 Jul 2020
0
5
10
15
Number of people
104
Rep. Recovered
Est. Recovered
Figure 2: Comparison of some of the outputs of the simulation run with ω= 1.4555% (Est.) and the Italian official
reported data (Rep.). Top-left: Cumulative number of cases and deaths. Top-right: People in hospital. Bottom-left:
People in quarantine. Bottom-right: Cumulative number of recovered people.
22
Table 1: Summary of all parameters used for COVID-19 in Italy when considering ω= 1.4555%. A brief description
(Description) of each parameter is given. Furthermore, if available, the range of the considered values (Value)
and the reference in the literature (Ref.) are reported. The notation * means that this value is identified during
the multiobjective or monoobjective optimization processes. The notation – means this parameter is omitted during
this work. The notation & means that this parameter (depending on time) is estimated directly from the reported
data. The notation means that the value of the parameter is set by the authors of this article, after some numerical
experiments.
Notation Value Description Ref.
N60317000 Number of people in the country before starting the pandemic [59]
γE1
5.5Transition rate of a person in compartment E(day1) [64]
γI(t)1
5Transition rate of a person in compartment I(day1)
at time t[25]
γIu(t)1
9Transition rate of a person in compartment Iu
to compartment Ru(day1) at time t[25]
γHR(t)1
14.2729 Transition rate of a person in compartment HR
to compartment Q(day1) at time t*
γHD(t)1
5Transition rate of a person in compartment HD
to compartment D(day1) at time t[25]
γIDu(t)Transition rate of a person in compartment IDu
to compartment Du(day1) at time t[31]
γQ,11
36.0450 Transition rate of a person in compartment Q
to compartment Rd(day1) at time ttγQ*
γQ,21
24.8646 Transition rate of a person in compartment Q
to compartment Rd(day1) at time t>tγQ*
tγQ2 May 2020 Date when the duration of a person in quarantine changes *
dg0 Maximum number of days that dIcan be decreased [26]
CE0.3806 Ratio between the disease contact rates βEand βI[33]
Cu0.3293 Ratio between βIand the disease contact rate βI*
CIDu Ratio between the disease contact rates βIDuand βI
βI,00.4992 Disease contact rate of a person in compartment I
without control measures (day1) *
βIDu,0 Disease contact rate of a person in compartment IDu
without control measures (day1) –
βHR& Disease contact rate of a person in compartment HR&
βHD& Disease contact rate of a person in compartment HD&
t019 Jan 2020 Initial time [42]
ω1.4555 iIdFR when the implemented control measures are fully applied (in %) [21]
δω0 to compute iIdFR when no control measures are applied ω=ω+δω
ωu,00.42 iIuFR when t=t0(in %) *
ωu,10.42 iIuFR when t=tu,1(in %)
ωu,20 iIuFR when t=tu,2(in %)
tu,13 May 2020 One day before the last report of dr,u[44]
tu,24 May 2020 Date of the last report of dr,u[44]
θ(t)& Proportion of infected people that are detected by the authorities &
p00.7382 Ratio of the number of new detected infected people that will survive
the disease and are hospitalized at time tθ0[44]
23
Table 2: Summary of all parameters related to control measures used for COVID-19 in Italy when considering ω=
1.4555%. A brief description (Description) of each parameter is given. Furthermore, if available, the range of the
considered values (Value) and the reference in the literature (Ref.) are reported. The notation * means that this
value is optimized during the multiobjective optimization process. The notation – means this parameter is omitted
during this work.
Notation Value Description Ref.
λ019 Jan 2020 First day of application of the control strategy
that was being used before t0, if any [42]
λ123 Feb 2020 First day of application of the 1st control strategy
(lockdown on 11 municipalities) [42]
λ211 Mar 2020 First day of application of the 2nd social distancing strategy
(lockdown on the entire country) [42]
λ322 Mar 2020 First day of application of the 3rd social distancing strategy
(closure of non-essential or strategic production activities) [42]
λ44 May 2020 First day of application of the 4th social distancing strategy
(end of closure of non-essential or strategic production activities) [42]
λ518 May 2020 First day of application of the 5th social distancing strategy
(end of regional travel restrictions and reopening of production activities) [42]
λ63 Jun 2020 First day of application of the 6th social distancing strategy
(progressive end of total travel restrictions, reopening of production,
commercial and social activities) [43]
m01 Intensity of the social distancing strategy used before t0[42]
m11 Intensity of the control strategy that was being used in (t0, λ1][42]
m20.5332 Intensity of the social distancing strategy used in (λ1, λ2]*
m30.1369 Intensity of the social distancing strategy used in (λ2, λ3]*
m40 Intensity of the social distancing strategy used in (λ3, λ4][33]
m50.0549 Intensity of the social distancing strategy used in (λ4, λ5]*
m60.0577 Intensity of the social distancing strategy used in (λ5, λ6]*
m70.0578 Intensity of the social distancing strategy used in (λ6, λ7]*
m(t) Intensity of the contact tracing control strategy
m(s) Intensity of the health system control strategy
κ1 Efficiency of the social distancing strategy used in (λ0, λ1]
κ262.0015 Efficiency of the social distancing strategy used in (λ1, λ2]*
κ30.5139 Efficiency of the social distancing strategy used in (λ2, λ3]*
κ40.0362 Efficiency of the social distancing strategy used in (λ3, λ4]*
κ590.6408 Efficiency of the social distancing strategy used in (λ4, λ5]*
κ614.6514 Efficiency of the social distancing strategy used in (λ5, λ6]*
κ752.0747 Efficiency of the social distancing strategy used in (λ6, λ7]*
κ(t) Efficiency of the contact tracing control strategy
κ(s) Efficiency of the health system control strategy
by using the methodology described in Section 3.2, from 22 July 2020 on, we assume that θremains equal to its value
on 21 July 2020.
First, for ω= 0.9145%,ω= 1.4555% and ω= 3.9559%, we consider m8= 0.25. Some results of this simulation
are reported in Figure 5 (Top-Left). In the three cases we observe a slight increase in the number of cases.
Next, for ω= 1.4555%, we simulate four future scenarios with: (i) m8=m7= 0.0578; (ii) m8= 0.15; (iii)
m8= 0.25, and (iv) m8= 0.4. In Figure 5 (Top-Right), (Bottom-Left) and (Bottom-Right), we show some outputs
24
19 Jan 2020
08 Feb 2020
28 Feb 2020
20 Mar 2020
09 Apr 2020
30 Apr 2020
20 May 2020
10 Jun 2020
30 Jun 2020
21 Jul 2020
0
1000
2000
3000
4000
5000
6000
Number of people/day
Rep. New Cases
Est. New Cases
19 Jan 2020
08 Feb 2020
28 Feb 2020
20 Mar 2020
09 Apr 2020
30 Apr 2020
20 May 2020
10 Jun 2020
30 Jun 2020
21 Jul 2020
0
200
400
600
800
Number of people/day
Rep. New Deaths
Est. New Deaths
19 Jan 2020
08 Feb 2020
28 Feb 2020
20 Mar 2020
09 Apr 2020
30 Apr 2020
20 May 2020
10 Jun 2020
30 Jun 2020
21 Jul 2020
0
1
2
3
4
5
6
7
8
Re
14-Mar-2020
Est. Re =1.4555%
Est. ReE =1.4555%
Est. ReI =1.4555%
Est. ReIu =1.4555%
Est. ReHR =1.4555%
Est. ReHD =1.4555%
19 Jan 2020
08 Feb 2020
28 Feb 2020
20 Mar 2020
09 Apr 2020
30 Apr 2020
20 May 2020
10 Jun 2020
30 Jun 2020
21 Jul 2020
0
0.2
0.4
0.6
0.8
1
m
23-Feb-2020
11-Mar-2020
22-Mar-2020
04-May-2020
18-May-2020
03-Jun-2020
Est. social distancing
measures =1.4555%
Figure 3: Comparison of some of the outputs of the simulation run with ω= 1.4555% (Est.) and, when available, the
Italian official reported data (Rep.). Top-left: New detected cases per day. Top-right: New detected deaths per day.
Bottom-Left: Function modeling the social distancing measures. Bottom-Right: Effective reproduction number and
the contribution of each infectious state (i.e. E,I,Iu,HRand HD) to this number. The vertical line corresponds to
the first date when Re<1.
of the model for these four scenarios. As we can see, there is a remarkable possibility of new outbreaks in Italy if the
control measures are relaxed too much. For instance, focusing on the effective reproduction number, it goes back over
1 when m8= 0.25 and m8= 0.4, which would mean a second wave of uncontrolled outbreaks. In particular, if control
measures are relaxed too much (e.g. m8= 0.4) the new cases and deaths may increase dramatically. This clearly
indicates that, the pandemic is not over yet and efforts must continue, keeping enough control measures.
25
19 Jan 2020
08 Feb 2020
28 Feb 2020
20 Mar 2020
09 Apr 2020
30 Apr 2020
20 May 2020
10 Jun 2020
30 Jun 2020
21 Jul 2020
1
2
3
4
5
6
7
8
Re
Estimated Re
13-Mar-2020
14-Mar-2020
=0.9145%
=1.4555%
=3.9559%
19 Jan 2020
08 Feb 2020
28 Feb 2020
20 Mar 2020
09 Apr 2020
30 Apr 2020
20 May 2020
10 Jun 2020
30 Jun 2020
21 Jul 2020
0.2
0.4
0.6
0.8
1
m
Estimated social distancing measures
=0.9145%
=1.4555%
=3.9559%
19 Jan 2020
08 Feb 2020
28 Feb 2020
20 Mar 2020
09 Apr 2020
30 Apr 2020
20 May 2020
10 Jun 2020
30 Jun 2020
21 Jul 2020
0
1
2
3
4
Cumulative number of people
106Estimated number of total cases
(including undetected)
=0.9145%
=1.4555%
=3.9559%
19 Jan 2020
08 Feb 2020
28 Feb 2020
20 Mar 2020
09 Apr 2020
30 Apr 2020
20 May 2020
10 Jun 2020
30 Jun 2020
21 Jul 2020
0.1
0.2
0.3
0.4
0.5
0.6
Estimated
=0.9145%
=1.4555%
=3.9559%
Figure 4: Some of the outputs of the simulation run with ω= 1.4555%, ω = 0.9145% and 3.9559% for the Italian case.
Top-left: Effective reproduction number. Top-right: Function modeling the social distancing measures. Bottom-left:
Cumulative number of total cases (including undetected cases). Bottom-right: θ(t).
5 Conclusions and Discussion
We have developed an epidemiological model for simulating the dynamics of COVID-19 in a considered territory.
This model may be used by epidemiologists and policy makers for understanding the spread of the pandemic, study
the efficiency of control measures and propose suitable ways to tackle the problem. It is intended to be as simple
as possible, but with sufficient complexity to be able to simulate the biological and sociological mechanisms which
influence the disease spread (including the three main types of control measures: social distancing, contact tracing
and health system measures), to fit real data and to simulate possible future scenarios. Of course, for particular cases
the model may be simplified (for instance removing the IDucompartment, as we showed for the case of Italy, or the
Qcompartment, if the territory does not keep asymptomatic people in quarantine at home).
The model can serve as a basis for studies including higher complexity. Indeed, it can be modified to introduce
26
New Cases
19 Jan 2020
13 Feb 2020
09 Mar 2020
03 Apr 2020
28 Apr 2020
23 May 2020
17 Jun 2020
12 Jul 2020
06 Aug 2020
01 Sep 2020
0
1000
2000
3000
4000
5000
6000
Number of people/day
Reported
Est. =0.9145%
Est. =1.4555%
Est. =3.9559%
19 Jan 2020
13 Feb 2020
09 Mar 2020
03 Apr 2020
28 Apr 2020
23 May 2020
17 Jun 2020
12 Jul 2020
06 Aug 2020
01 Sep 2020
1
2
3
4
5
6
7
8
Re
Re =1.4555%
Est. m8=m7
Est. m8=0.15
Est. m8=0.25
Est. m8=0.4
New Cases =1.4555%
19 Jan 2020
13 Feb 2020
09 Mar 2020
03 Apr 2020
28 Apr 2020
23 May 2020
17 Jun 2020
12 Jul 2020
06 Aug 2020
01 Sep 2020
0
1000
2000
3000
4000
5000
6000
7000
Number of people/day
Reported
Est. m8=m7
Est. m8=0.15
Est. m8=0.25
Est. m8=0.4
New Deaths =1.4555%
19 Jan 2020
13 Feb 2020
09 Mar 2020
03 Apr 2020
28 Apr 2020
23 May 2020
17 Jun 2020
12 Jul 2020
06 Aug 2020
01 Sep 2020
0
200
400
600
800
Number of people/day
Reported
Est. m8=m7
Est. m8=0.15
Est. m8=0.25
Est. m8=0.4
Figure 5: Outputs obtained when simulating possible future scenarios (using κ8=κ2= 62.0015). Top-left: New
detected people when m8= 0.25 and ω=0.9145%, 1.4555% or 3.9559%, Top-right: Effective reproduction number
when ω= 1.4555% and m8= 0.15, 0.25 or 0.4, Bottom-left: New detected cases per day, when ω= 1.4555% and
m8= 0.15, 0.25 or 0.4, Bottom-right: New detected deaths per day, when ω= 1.4555% and m8= 0.15, 0.25 or 0.4.
structures of age, sex, etc., stochasticity, levels of uncertainty, mobility between territories, spatial structure, etc.
In the model that we propose, there are several important points to highlight that, to the best of our knowledge,
have not been previously considered:
1. We have included a compartment for undetected deaths mainly happening in nursing homes or residences for
disabled people. The number of these deaths and the associated infections does not seem to be negligible, which
makes necessary to take them into account in the model.
2. Some dynamic coefficients are based on reported time series, which allows the model to reproduce faithfully the
spread of the disease, fitting very well the reported data.
3. The model includes the effects of the different control measures described above.
27
4. The model takes into account quarantine and hospitalized people in order to estimate the number of hospital
beds that are needed.
We point out that the model, in this current simple form, has the following limitations:
1. Epidemics/pandemics are phenomena that respond to complex processes and all epidemiological models are
simplifications of this complex reality. Furthermore, the model is based on the interactions of human beings,
which do not behave in reasonably predictable ways like molecules, cells or particles (see [22]).
2. The prediction of its evolution is very complicated because of the characteristics of the system of differential
equations governing the model. In particular, in the exponential growth phases, small variations in some pa-
rameters can lead to large variations in model outcomes (see, [33]). Moreover, there are data quality problems:
a) uncertainty about the characteristics of SARS-CoV-2 (new virus); b) changes in criteria to count cases; c)
uncertainty about undetected cases (and, therefore, the mortality rate), etc.
3. Since there is no clear scientific evidence of the effect of humidity and temperature (see, e.g. [40, 56]), we have
not included these two factors in our model (if such evidence is demonstrated, these variables should be included
in the model).
4. It is assumed that the spatial distribution of the population is homogeneous in the territory under study (if this
is not a suitable assumption, the territory can be divided into smaller ones).
5. In this paper we have only considered the spread of disease within a territory, where only local transmission and
known imported / exported cases are modeled. The between-country spread has not been modeled in this work,
but can be considered in the future employing ideas from the BeCoDiS model (see [35]).
Acknowledgments
This research was partially supported by the Spanish Government under projects PID2019-106337GB-I00, MTM2015-
64865-P and MTM2017-89423-P; the research group MOMAT (Ref. 910480) supported by the Complutense University
of Madrid; and the Regional Government of Andalusia under project P12-TIC301, partially financed by the European
Regional Development Fund (ERDF). Also thanks to M. Ramos-Rubio for his help with the manuscript.
Annex
In the version of this article initially submitted to Physica D: Nonlinear Phenomena, the case of Italy studied in
Section 4 was written using official data up to 21 July 2020 to validate the model and analyze several scenarios up
to 1 September 2020. In the current version, written later as a revised version of the submitted paper, we have kept
everything that was computed using those data (including the values of the parameters). Furthermore, following the
reviewers’ suggestions, in this Annex we show an update of the situation in Italy, using official data up 21 December
2020, to compare with our simulations. It is important to remark that in all the simulations shown in this Annex,
we keep the same parameter values as those given in the version of the paper initially submitted to the journal (with
data up to 21 July 2020, as mentioned above).
In all cases, we consider ω= 1.4555% and we estimate θby using the methodology described in Section 3.2 until 21
December 2020 assuming that, since then, θremains equal to its value on this day. Between 14 and 15 August 2020,
the cumulative number of deaths reported by the Italian authorities shows a sudden increase, because they updated
unreported deaths that occurred in previous weeks. According to their weekly reports (see [27]), from 10 to 23 August
there were 14 new deaths. Thus, 218 deaths must be smoothly distributed before the 10 August as done in [33] for the
case of China. First, we distribute the 14 deaths from 10 to 23 August and, then, we assign the 218 deaths between
22 July and 23 August, avoiding the modification of the data that we had already used in the paper until 21 July.
Besides the values of the model parameters shown in Section 4, in Table 3 we show the values of the parameters
that we have used to model the control measures (as implemented for the Italian Government) corresponding to the
days after the 21 July 2020.
28
Table 3: Summary of the parameters used to model the control measures corresponding to the days after the 21 July
2020. Time intervals with different measures are indexed with i∈ {8,9,10,11,12,13}.
Date of application Efficiency Intensity Type Detail
(λi1) (κi) (mi)
1 Aug 2020 62.0015 0.3 Relaxation “New normality”
20 Sep 2020 62.0015 0.53 Relaxation Back from vacation,
school, work. Regional
elections
3 Nov 2020 62.0015 0.3 Restriction Curfew
13 Nov 2020 62.0015 0.22 Restriction More regions with
restrictive measures
23 Dec 2020 62.0015 Different Different Christmas
scenarios scenarios
7 Jan 2020 62.0015 Different Different End of Christmas
scenarios scenarios
In Figure 6 we show some results corresponding to five different scenarios up to 31 January 2021:
Red scenario: More restrictive control measures are implemented from 23 December to 31 January.
Yellow scenario: More restrictive control measures are implemented from 23 December to 6 January. From 7
January, the measures come back to the ones before 23 December.
Purple scenario: The control measures are maintained without changes.
Green scenario: The control measures are relaxed for Christmas from 23 December to 6 January. From 7 January,
more restrictive control measures are implemented.
Cyan scenario: The control measures are relaxed for Christmas from 23 December. These measures continue
during January, i.e., no restrictions are implemented after Christmas.
As we can see, Italy could be on 22 December at a crucial moment where any changes in control measures could
have notable consequences. The model estimates that on 22 December the effective reproduction number Recould be
slightly higher than 1 and the intensity of the control measures could be of 0.22 (considering 1 the absence of measures
before the pandemic). If they tighten the measures from 23 December, as the Italian government has announced,
the effective reproduction number could drop below 1 and the pandemic would be controlled (see the red scenario
in Figure 6). However, if the measures are relaxed too soon in January, the cases and deaths would increase again
before the end of January (see the yellow scenario in Figure 6). As we can see in the purple scenario, if no changes
are implemented in the measures from 22 December to 31 January, the new cases and deaths may increase. Finally,
if the measures are relaxed on 23 December, there would be a third wave of uncontrolled outbreaks whose magnitude
depends on the restrictive measures they apply after Christmas (see the green and cyan scenarios in Figure 6).
References
[1] D. Adam. A guide to R — the pandemic’s misunderstood metric. Nature, News Feature, 3 July 2020. doi:
https://doi.org/10.1038/d41586-020-02009-w.
[2] M. M. Arons, K. M. Hatfield, S. C. Reddy, A. Kimball et al. Presymptomatic SARS-CoV-2 infections and
transmission in a skilled nursing facility. New England Journal of Medicine, 382(22):2081–2090, 2020. doi:
10.1056/NEJMoa2008457.
29
19 Jan 2020
01 Mar 2020
12 Apr 2020
24 May 2020
05 Jul 2020
16 Aug 2020
27 Sep 2020
08 Nov 2020
20 Dec 2020
31 Jan 2021
0
1
2
3
4
5
6
7
8
Re
Re =1.4555%
Est. m12=0.15, m13=0.15
Est. m12=0.15, m13=0.22
Est. m12=0.22, m13=0.22
Est. m12=0.35, m13=0.15
Est. m12=0.35, m13=0.35
New Cases =1.4555%
19 Jan 2020
01 Mar 2020
12 Apr 2020
24 May 2020
05 Jul 2020
16 Aug 2020
27 Sep 2020
08 Nov 2020
20 Dec 2020
31 Jan 2021
0
2
4
6
8
10
12
Number of people/day
104
Reported
Est. m12=0.15, m13=0.15
Est. m12=0.15, m13=0.22
Est. m12=0.22, m13=0.22
Est. m12=0.35, m13=0.15
Est. m12=0.35, m13=0.35
New Deaths =1.4555%
19 Jan 2020
01 Mar 2020
12 Apr 2020
24 May 2020
05 Jul 2020
16 Aug 2020
27 Sep 2020
08 Nov 2020
20 Dec 2020
31 Jan 2021
0
500
1000
1500
2000
2500
3000
3500
Number of people/day
Reported
Est. m12=0.15, m13=0.15
Est. m12=0.15, m13=0.22
Est. m12=0.22, m13=0.22
Est. m12=0.35, m13=0.15
Est. m12=0.35, m13=0.35
19 Jan 2020
01 Mar 2020
12 Apr 2020
24 May 2020
05 Jul 2020
16 Aug 2020
27 Sep 2020
08 Nov 2020
20 Dec 2020
31 Jan 2021
0
0.2
0.4
0.6
0.8
1
m
Social distancing measures =1.4555%
23-Feb-2020
11-Mar-2020
22-Mar-2020
04-May-2020
18-May-2020
03-Jun-2020
01-Aug-2020
20-Sep-2020
03-Nov-2020
13-Nov-2020
23-Dec-2020
07-Jan-2021
Est. m12=0.15, m13=0.15
Est. m12=0.15, m13=0.22
Est. m12=0.22, m13=0.22
Est. m12=0.35, m13=0.15
Est. m12=0.35, m13=0.35
Figure 6: Outputs obtained when simulating possible future scenarios when ω= 1.4555%. Top-left: Function modeling
the social distancing measures. Top-right: Effective reproduction number. Bottom-left: New detected cases per day.
Bottom-right: New detected deaths per day.
[3] A. Basu. Estimating the infection fatality rate among symptomatic COVID-19 cases in the United States. Health
Affairs, 39(7):1229–1236, 2020. https://doi:10.1377/hlthaff.2020.00455.
[4] E. Bertuzzo, L. Mari, D. Pasetto, S. Miccoli, R. Casagrandi, M. Gatto and A. Rinaldo. The geography of COVID-
19 spread in Italy and implications for the relaxation of confinement measures. Nature Communications, 11:4264,
2020. doi:https://doi.org/10.1038/s41467-020-18050-2.
[5] F. Brauer and C. Castillo-Chávez. Mathematical Models in Population Biology and Epidemiology. Texts in Applied
Mathematics, Springer, 2001. URL: https://www.springer.com/gp/book/9781441931825.
[6] S. Chen, J. Yang, W. Yang, and T. Barnighausen. COVID-19 control in China during mass population movements
at New Year. The Lancet, 395(10226):764–766, 2020. doi:https://doi.org/10.1016/S0140-6736(20)30421-9.
[7] H. Chu, J. Li, D. Blakeley, and R.J. Smith. The failure of R0.Computational and Mathematical Methods in
Medicine, 527610, 2011. doi:10.1155/2011/527610.
[8] M. D’Arienzo and A. Coniglio. Assessment of the sars-cov-2 basic reproduction number, r0, based on the early
phase of covid-19 outbreak in italy. Biosafety and Health, 2(2):57 – 59, 2020. doi:https://doi.org/10.1016/
j.bsheal.2020.03.004.
30
[9] Centers for Disease Control and Prevention. Transmission of SARS-COV-2 Infections in Households — Tennessee
and Wisconsin. Mobidity and Mortality Weekly Report, 69(44):1631-1634, 2020. https://www.cdc.gov/mmwr/
volumes/69/wr/pdfs/mm6944e1-H.pdf.
[10] F. Della Rossa, D. Salzano, A. Di Meglio et al. A network model of Italy shows that intermittent regional strategies
can alleviate the COVID-19 epidemic. Nature Communications, 11:5106, 2020. doi:https://doi.org/10.1038/
s41467-020-18827-5.
[11] G. De Natale, V. Ricciardi, G. De Luca, D. De Natale et al. The COVID-19 infection in italy: A statistical study
of an abnormally severe disease. J. Clin. Med., 9(5):1564, 2020. doi:https://doi.org/10.3390/jcm9051564.
[12] O. Diekmann, H. Heesterbeek, and T. Britton. Understanding Infectious Disease Dynamics. Princeton Series
in Theoretical and Computational Biology, Princeton University Press, 2013. URL: http://www.jstor.org/
stable/j.cttq9530.
[13] Dipartimento della Protezione Civile. Aggiornamento casi COVID-19. URL: http://opendatadpc.maps.arcgis.
com/apps/opsdashboard/index.html#/b0c68bce2cce478eaac82fe38d4138b1, 2020.
[14] European Centre for Disease Prevention and Control. Surveillance of COVID-19 at long-term
care facilities in the EU/EEA. URL: https://www.ecdc.europa.eu/sites/default/files/documents/
covid-19-long-term-care-facilities- surveillance-guidance.pdf, May 2020.
[15] M. R. Ferrández, B. Ivorra, J. L. Redondo, A. M. Ramos, and P. M. Ortigosa. A multi-objective approach to
estimate parameters of compartmental epidemiological models. Application to Ebola Virus Disease epidemics.
Preprint in ResearchGate, 2020. doi:https://doi.org/10.13140/RG.2.2.25778.56006.
[16] M. R. Ferrández, S. Puertas-Martín, J. L. Redondo, B. Ivorra, A. M. Ramos, and P. M. Ortigosa. High-
performance computing for the optimization of high-pressure thermal treatments in food industry. The Journal
of Supercomputing, 75:1187–1202, 2018. doi:10.1007/s11227- 018-2351-4.
[17] M. R. Ferrández, J. L. Redondo, B. Ivorra, A. M. Ramos, and P. M. Ortigosa. Preference-based multi-
objectivization applied to decision support for High-Pressure Thermal processes in food treatment. Applied
Soft Computing, 79:326–340, 2019. doi:https://doi.org/10.1016/j.asoc.2019.03.050.
[18] M. Gatto, E. Bertuzzo, L. Mari, S. Miccoli, L. Carraro, R. Casagrandi and A. Rinaldo. Spread and dynamics of
the COVID-19 epidemic in Italy: Effects of emergency containment measures. PNAS, 117(19):10484-10491, 2020.
doi:https://doi.org/10.1073/pnas.2004978117.
[19] G. Giordano, F. Blanchini, and R. Bruno et al. Modelling the COVID-19 epidemic and implementa-
tion of population-wide interventions in Italy. Nature Medicine, 2020. doi:https://doi.org/10.1038/
s41591-020-0883-7.
[20] D.H. Glass. European and US lockdowns and second waves during the COVID-19 pandemic. Mathematical
Biosciences, 330:108472, 2020. doi:https://doi.org/10.1016/j.mbs.2020.108472.
[21] R. Grewelle and G. A. De Leo. Estimating the global infection fatality rate of COVID-19. medRxiv, 2020.
doi:https://doi.org/10.1101/2020.05.11.20098780.
[22] H.W. Hethcote. The basic epidemiology models: models, expressions for R0, parameter estimation, and applica-
tions. In Mathematical Understanding of Infectious Disease Dynamics, Chapter 1, pages 1–61. (eds.) S. Ma and
Y. Xia. World Scientific, 2009. doi:10.1142/9789812834836_0001.
[23] G. Huang, Y. Takeuchi, W. Ma and D. Wei. Global Stability for Delay SIR and SEIR Epidemic Models with
Nonlinear Incidence Rate. Bulletin of Mathematical Biology, Volume 72, 1192–1207, 2010. doi:https://doi.
org/10.1007/s11538-009-9487-6.
31
[24] Instituto de Salud Carlos III. Vigilancia de los excesos de mortalidad por todas las causas. MoMo. URL: https:
//www.isciii.es/QueHacemos/Servicios/VigilanciaSaludPublicaRENAVE/EnfermedadesTransmisibles/
MoMo/Documents/informesMoMo2020/MoMo_Situacion%20a%2021%20de%20diciembre_CNE.pdf, December
2020.
[25] Istituto Superiore di Sanità. Characteristics of SARS-CoV-2 patients dying in Italy Report based on
available data on May 21st, 2020. URL: https://www.epicentro.iss.it/en/coronavirus/bollettino/
Report-COVID-2019_21_may_2020.pdf, May 2020.
[26] Istituto Superiore di Sanità. Epidemia COVID-19, aggiornamento nazionale. URL: https://www.epicentro.
iss.it/coronavirus/bollettino/Bollettino-sorveglianza-integrata-COVID-19_23-giugno- 2020.pdf,
June 2020.
[27] Istituto Superiore di Sanità. Epidemia COVID-19, aggiornamento nazionale. URL: https://www.epicentro.
iss.it/coronavirus/bollettino/Bollettino-sorveglianza-integrata-COVID-19_25-agosto- 2020.pdf,
August 2020.
[28] Istituto Superiore di Sanità. Infografica. URL: https://www.epicentro.iss.it, 2020.
[29] Istituto Superiore di Sanità. Recommendation for people in self-isolation and for the family members assist-
ing them. 2020. https://www.iss.it/documents/20126/0/Poster+Raccomandazioni+isolamento+rev+12+
marzo+\%281\%29.pdf/d409de34-8bba-a220-10e8-e7e8a33963c3.
[30] Istituto Superiore di Sanità. Survey nazionale sul contagio COVID-19 nelle strutture residenziali e sociosan-
itarie. URL: https://www.epicentro.iss.it/coronavirus/pdf/sars-cov-2-survey-rsa-rapporto-3.pdf,
April 2020.
[31] Istituto Superiore di Sanità. Survey nazionale sul contagio COVID-19 nelle strutture residenziali e sociosani-
tarie. URL: https://www.epicentro.iss.it/coronavirus/pdf/sars- cov-2-survey-rsa-rapporto-finale.
pdf, May 2020.
[32] Istituto Nazionale di Statistica ed Istituto Superiore di Sanità. Impatto dell’epidemia COVID-19 sulla mortalità to-
tale della popolazione residente primo quatrimestre 2020. URL: https://www.epicentro.iss.it/coronavirus/
pdf/Rapp_Istat_Iss_3Giugno.pdf, June 2020.
[33] B. Ivorra, M. R. Ferrández, M. Vela-Pérez, and A.M. Ramos. Mathematical modeling of the spread of the
coronavirus disease 2019 (COVID-19) taking into account the undetected infections. the case of China. Commu-
nications in Nonlinear Science and Numerical Simulation, 88:105303, 2020. doi:https://doi.org/10.1016/j.
cnsns.2020.105303.
[34] B. Ivorra, D. Ngom, and A.M. Ramos. Stability and sensitivity analysis of Be-CoDiS, an epidemiological model
to predict the spread of human diseases between countries. Validation with data from the 2014-16 West African
Ebola Virus Disease epidemic Electronic Journal of Differential Equations, 2020, No. 62, pp. 1–29. https:
//ejde.math.txstate.edu/Volumes/2020/62/ivorra.pdf.
[35] B. Ivorra, A. M. Ramos, and D. Ngom. Be-CoDiS: A mathematical model to predict the risk of human diseases
spread between countries. Validation and application to the 2014 Ebola Virus Disease epidemic. Bulletin of
Mathematical Biology, 77(9):1668–1704, 2015. doi:https://doi.org/10.1007/s11538-015-0100-x.
[36] A.B. Kubik. Parameter estimation for a mathematical model predicting the COVID-19 spread in the Àrea
Metropolitana de Barcelona. Ms Thesis, 2020. URL: https://upcommons.upc.edu/handle/2117/328166.
[37] A. J. Kucharski, T. W. Russell, C. Diamond, Y. Liu, J. Edmunds, S. Funk, R. M.Eggo et al. Early dynamics of
transmission and control of COVID-19: a mathematical modelling study. The Lancet Infectious Diseases, 2020.
doi:https://doi.org/10.1016/S1473-3099(20)30144-4.
32