key: cord-0529916-skey1zfo
authors: Cadoni, Mariano
title: How to reduce epidemic peaks keeping under control the time-span of the epidemic
date: 2020-04-05
journal: nan
DOI: nan
sha: 86f989f2bde0f90c445d81a38f0edbd2d6f15ea3
doc_id: 529916
cord_uid: skey1zfo

One of the main challenges of the measures against the COVID-19 epidemic is to reduce the amplitude of the epidemic peak without increasing without control its timescale. We investigate this problem using the SIR model for the epidemic dynamics, for which reduction of the epidemic peak $I_P$ can be achieved only at the price of increasing the time $t_P$ of its occurrence and its entire time-span $t_E$. By means of a time reparametrization we linearize the equations for the SIR dynamics. This allows us to solve exactly the dynamics in the time domain and to derive the scaling behaviour of the size, the timescale and the speed of the epidemics, by reducing the infection rate $alpha$ and by increasing the removal rate $beta$ by a factor of $lambda$. We show that for a given value of the size ($I_P$, the total, $I_E$ and average $hat I_P$ number of infected), its occurrence time $t_P$ and entire time-span $t_E$ can be reduced by a factor $1/lambda$ if the reduction of $I$ is achieved by increasing the removal rate instead of reducing the infection rate. Thus, epidemic containment measures based on tracing, early detection followed by prompt isolation of infected individuals are more efficient than those based on social distancing. We apply our results to the COVID-19 epidemic in Northern Italy. We show that the peak time $t_P$ and the entire time span $t_E$ could have been reduced by a factor $0.9 le 1/lambdale 0.34$ with containment measures focused on increasing $beta$ instead of reducing $alpha$.

The recent COVID-19 epidemics is posing formidable challenges both to the health and economic systems worldwide. In order to tackle the ongoing epidemic, the countries faced with the COVID-19 epidemic have used different strategies. In order to control the epidemic one has to lower the basic reproduction number (also called basic reproduction rate) ρ, eventually below the threshold, i.e ρ < 1. In a simple SIR Model [1, 2, 3, 4, 5, 6] , which we consider in this paper, ρ depends on the infection rate α, the removal rate β and the total population N . If ρ > 1 the epidemic starts with the number of infected individuals I growing exponentially, it reaches a maximum I P at a time t P , then decreases down to zero at time t E when the epidemic ends.

Although for severe epidemics like the COVID-19 it is very difficult to keep the reproduction number ρ from the beginning below the threshold, nevertheless containment measures soften the epidemic because they lead to a reduction of the epidemic peak I P .

Basically, there are therefore three types of containment measures one can use. One can act on α by lowering it, for instance by forcing social distancing or by prophylaxis measures. One can achieve the same result increasing β, e.g. by means of a prompt strict isolation of infected individuals. Last but not least one can reduce the total population N by separating it in strictly non-communicating compartments.

The policies of different countries for fighting the COVID-19 usually contemplate a mixture of the three types of measures mentioned above. Obviously, the choice of focusing on one kind of measure instead of another depends on a number of factors, which include not only its effectiveness but also its feasibility and its social and economic impact. For instance most of the European countries have focused their COVID-19 fighting strategies on measures aimed at reducing the infection rate α. Countries like Korea and Singapore favoured instead measures aimed at rising β.

A general, unpleasant, feature of the epidemic dynamics is that the reduction of the epidemic peak I P can be only achieved at the expenses of increasing t P and t E , i.e of increasing the time-span of the epidemic. On the one hand for epidemics like COVID-19 having a high rate of hospitalised infected with severe symptoms, reduction of I P and dilution of the epidemic is necessary in order to allow the health systems to treat them properly. On the other hand, increasing t P means increasing the time-span in which the containment measures are effective, with potentially disruptive effects both on the economies and the live of the populations in the involved countries.

In this paper we investigate, in the framework of the SIR model, the impact of containment measures, which act on α and β, on I P , t P , t E and on the epidemic speed (dρ/dt| P ). The structure of the paper is as follows. In Sect 2, by means of a time reparametrization, we are able to linearize the equations for the SIR dynamics. This allows us to solve exactly the dynamics in the time domain and to derive the scaling behaviour of I P , t P , t E and dρ/dt| P , by reducing the infection rate α → α/λ or by increasing the removal rate → λβ. We show that keeping I P fixed its occurrence time t P and t E can be be reduced by a factor 1/λ by acting on the removal rate β instead on the infection rate α. This will be discussed in Sect. 3. In Sect. 4 we discuss approximate solutions of the SIR model in which the reproduction number ρ < e. In Sect. 5 we apply our results to the COVID-19 epidemic in Northern Italy. We show that the peak time t P and the entire time-span of the epidemic could have been reduced by a factor 0. 9 1/λ 0.34 with containment measures focused on increasing β instead of the reducing α. Finally, in Sect. 6 we state our conclusions.

The SIR model describes the deterministic dynamics of an infective epidemic, characterized by the fact that individuals, which have been infected and have recovered gain permanent immunity [1, 2, 3, 4, 5, 6] . Although the model is quite simple, it can be used to give at least rough estimates of epidemic dynamics, and in particular of the COVID-19 epidemic [7, 9, 10, 11] . A generalisation of the SIR model to take into account a large number of asymptomatic infectives -hence more apt to describe the COVID-19 epidemichas been proposed in Ref. [9, 10, 11] .

The homogeneous and isolated population of N individuals exposed to the epidemic, is characterised at time t by the number of susceptible S(t), infected and infectives I(t) and removed (recovered, dead or isolated) R(t) individuals, with the conservation law N = S(t) + I(t) + R(t). The timescale of the epidemics is assumed to be relatively short so that N can be assumed constant.

The dynamic describing the evolution of the epidemic is deterministic and described by the following, non linear, dynamical system:

The infective epidemic is characterised by two parameters: (1) The infection rate (also called contact rate) α, which gives the transition rate between the class of susceptible and that of infected; (2) the removal rate β, which gives the transition rate between the class of infected and that of removed (1/β gives the characteristic time for the removal of infected from the dynamics).

From equation (1b) it is immediately evident that number of infected individuals grows, i.e the epidemic spreads, only if

where γ is the epidemic threshold. Equivalently, one can introduce the basic reproduction number ρ(t):

which represents the expected number of new infections generated by a single infection. The epidemic spreads if ρ > 1. The parameters α and β depend on several factors. Some of them are attributes of the pathogen causing the disease and cannot be changed. Other are influenced by the social behaviour of the individuals and can be therefore changed with containment and prophylaxis measures.

The system (1a),(1b),(1c) is difficult to solve, analytically in the time domain. Usually one proceeds by eliminating dt from Eqs. (1a),(1b), then after integration one easily finds the function I(S) = I 0 + (S 0 − S) + γ log(S/S 0 ), where I 0 , S 0 are the initial data (see e.g. Ref. [10] ). This form of I(S) allows one to derive some qualitative and quantitative features of the epidemic dynamics but not its explicit time evolution. This latter can be only obtained by numerical integration of Eqs. (1a),(1b),(1c). For instance, one can easily find that, if initially we are above the threshold ρ 0 > 1, I(t) grows till it reaches a maximum I P , then it goes down to zero at a time t E when the epidemic ends.

The function I(S) allows to determine analytically the value of the peak I P but not the time t P of its occurrence, nor its entire time-span t E , nor the speed of the reproduction number V P := dρ/dt| P , nor the average value of the number of infected individuals at the peakÎ P . t P , t E , V P andÎ P have to be determined after solving numerically the dynamics. This is a quite unpleasant feature because it prevents a clear understanding of the dependence of t P ,t E , V P andÎ P from the parameters α, β, which is a crucial information for fighting the epidemic.

In order to solve analytically the temporal dynamics let us reparametrize the time introducing a new time coordinate τ defined by dτ /dt = I(τ ), i.e.:

where t 0 = t(τ 0 ) is the initial time. Using time-translations we can put without loss of generality, τ 0 = t 0 = 0. The new time coordinate has a simple intuitive meaning, τ (t)/t gives the average valueÎ(t) the number of infected at time t:Î

The time reparametrization (4) allows to linearise the system (1a),(1b),(1c):

This can be easily integrate to give:

where S 0 , I 0 are initial data and R 0 = N − S 0 − I 0 . In the following we will take R 0 = 0. The function τ (t) is defined implicitly by

Exact solutions of the SIR model, which are equivalent to our Eqs. (7a),(7b),(7c), (8) have been derived in Refs. [12, 13] using a completely different approach. Although, the integral (8) cannot be evaluated analytically, it allows to solve the temporal dynamics of the SIR model and to investigate the scaling behaviour of the relevant quantities characterizing the epidemic when the parameters α, β change. The previous expressions allow us to compute easily all the relevant quantities for the epidemic peak. From Eqs. (6b) and (7a) one gets immediately τ P , the τ -time coordinate of the peak. The other quantities are readily computed using Eqs. (7a),7b),(7c), (8):

The entire time-span of the epidemic t E can be computed setting I = 0 in Eq. (7b). Because I 0 is usually small compared to S 0 it can be neglected, t E is obtained by first finding the (higher) root τ E of the transcendental equation:

and then using Eq. (8) to compute t E . An other important quantity, which describes the intensity of the epidemics is the total number I E of individuals that are infected over the whole time-span of the epidemics. Taking into account that I 0 is rather small and that initially S 0 = N , I E can expressed in terms of the lower root S E of the transcendental equation obtained by setting I = 0 in Eq. (7b) (see Ref. [9] ) for details. We have

where S E is the lower root of the transcendental equation

In this section we investigate the scaling behavior of the peak quantities (9a...(9d), the total number of infected (11) , I E and t E by changing of the parameters α and β. It is already known that Eqs.

. This scaling transformation leaves invariant the epidemic threshold γ and tells us that we can increase (reduce) the timescale of the epidemic by simultaneously reducing (increasing) both α and β. However, this is not what we are interested in. Actually, we want to know what happens to the epidemic parameters listed above when we increase the threshold γ.

Let us first observe that both the number of infected at the peak I P (see Eq. (9a) and the total number of infected I E (see Eq. (11) are decreasing functions of the parameter γ. In fact, we get from Eq. (9a) and Eq.

We see that above the epidemic threshold (S 0 /γ > 1), dI P /dγ is always negative, while being S E < N , dS E /dγ is always positive. It follows that if we want to reduce the peak and the total number of infected we have to increase γ by a factor λ > 1. Because one can increase γ either by reducing α or by increasing β, we have to compare the effects on the peak parameters of these two different ways of increasing γ.

We are therefore lead to consider two different scaling transformations: transformation T (1) , which reduces the infection rate:

and transformation T (2) , which increases the removal rate:

β → λβ, γ → λγ, λ 1.

The peak quantities in Eqs. (9a...(9d) and the τ E of Eq. (10) do not transform in a simple way under T (1) and T (2) , however the ratios of T (1) and T (2) -transformed quantities follow simple scaling laws. In particular, they remain invariant whenever the quantity depends only on their ratio γ and not on α and β separately.

Using the following notation to denote rescaled quantities: I

(1)

P = I P (λβ) and similarly for the others quantities, we get,

The transformation law for t P and t E can be derived by first acting with the transformation T (1) on the integral (9d), then acting with T (2) on the same integral and finally redefining the integration variable in the second integral τ → λ −1 τ . One obtains in this way:

Finally, using Eqs. (5), (16b), (17) and taking into account that I E depends only on γ (see Eqs. (11), (12)) we can easily show that the average number of infectedÎ P and the total number of infected I E are invariant, i.e.Î 

An important result follows from equations (16a),(16b),(17) and (18): epidemic containment measures, which have the same effect for what concerns I P , R P , S P ,Î P and I E , have different impact on the occurrence time t P of the peak, the whole time span of the epidemic t E and on the epidemic speed V P . Choosing measures increasing the removal rate β by a factor λ instead of reducing the infection rate α by a factor 1/λ allows to drop t P and t E by a factor 1/λ. For instance by implementing epidemic containment measures with λ = 2 we can reduce by a half both the time needed for the epidemic to reach the peak and the whole time-span of the epidemic. It should be noticed that this epidemic timescale reduction effect becomes more relevant for epidemics with high reproduction number ρ 0 >> 1. In fact the factor λ is limited by λ < ρ 0 , simply because for λ > ρ 0 the epidemic does not develop at all. Thus, if we have for instance ρ 0 = 5 we can reduce the peak time and the entire time-span of the epidemic until a factor of 1/5. Therefore, increasing β represents an efficient way to fight epidemics. If by increasing it we manage to bring ρ below the threshold we simply stop the epidemic, but even if we do not go so far, we can still reduce the size of an epidemic keeping under control its timescale.

The behaviour of the reproduction number speed V P in Eq. (16b) explains clearly what is going on. If one acts on β instead on α, V P increases, as expected, by a factor of λ. In short, increasing β instead of reducing α, allows one to speed up the epidemic dynamics keeping constant the number of infected at the peak, the average number of infected and the total number of infected. This is possible because the increasing of the removal rate allows prompt removal of infected individuals.

In the general case the integral (8) cannot be computed analytically. Therefore the function τ = τ (t) has to be computed numerically, by first performing numerical integration of the integral in (8) to find t = t(τ ) and then inverting it. There is, however, a situation in which the integral (8) can be computed analytically and the dynamics of the epidemic until the peak, can be expressed analytically in closed form in terms of the time t, albeit in approximate form.

For ατ << 1 we can approximate the exponential in Eq. (8) by e −ατ ≈ 1 − ατ . This approximation allows to solve the integral and to invert the function t = t(τ ). We find,

With this position we can easily write down the approximate form of the solutions (7a),(7b),(7c) in terms of the time t. We quote here only the form of I(t) and t P as a function of I P ,

Eqs. (19) and (20) are a good approximation only for ατ < 1. Because τ (t), is an increasing function of t, the approximation for the dynamics is good until the peak, if ατ P < 1, which implies from equation (9c):

The recent development of the COVID-19 epidemic in Northern Italy represents an interesting case for applying the results described in the previous sections. The epidemic developed in the three main regions of Northern Italy (Lombardia, Veneto and Emilia-Romagna) we consider in this paper, starting from end of February 2020 (although it may be possible that the epidemic was circulating in the regions before that date). Altogether these three regions have around 20 millions of inhabitants, we will therefore take N = 2·10 7 in our computations. We take as initial value I 0 = 100, which approximately corresponds to the known cases of COVID-19 infected Northern Italy on February 23, 2020. The determination of the initial values of the other two parameters of the SIR model α 0 and β 0 is more involved. These initial values are completely determined by the pathogen because they are not affected by the epidemic containment measures put into play. The value of α 0 can be determined from the exponential behaviour of the early dynamics, or equivalently from the initial doubling time [7, 8] . Using the raw data for the early dynamics of the epidemic in Northern Italy, Gaeta [7, 8] has given the estimate:

The determination of β 0 is even more problematic. This is because we expect it to be sensitive to the presence of large cohort of asymptomatic infectives. We can estimate β 0 from the basic reproduction number ρ 0 , using Eq. (3). Rough evaluations of ρ 0 give a number between 2 and 2.5, however the indeterminacy related to the presence of large number of asymptomatic infectives may result in a much higher value for ρ 0 . To be rather conservative we assume here ρ 0 = 3, so that Eqs. (3) and (22) give:

The COVID-19 containment strategies put in place in Northern Italy are a mixture of social distancing, social confinement, early detection and infection tracing. Although mainly focused of social distancing, these strategies contain all the previous ingredients, which modify in different ways the parameters α and β. Social distancing acts by reducing α, whereas systematic, prompt, and strict isolation of infected individuals as a result of early detection and tracing enhances β. Both effects rise γ and reduce in the same way the amplitude of the peak I P , the average number of infectedÎ P and the total number of infected I E .

Because it is almost impossible to disentangle the effects of the various containment measures on α and β in the real situation, we will discuss and compare two hypothetical situations in which the rising of γ, γ → λγ 0 , with γ 0 = (2/3) · 10 7 , is obtained in two fully distinct and complementary ways:

• (1) We have exclusively social confinement containment measures: β is held fixed to its initial value β 0 , whereas α 0 is reduced by a factor 1/λ.

We have exclusively containment measures consisting in prompt and strict isolation of infected individuals triggered by early detection and tracing of infected: α is held fixed to its initial value α 0 , whereas β 0 is increased by a factor λ.

Being 1 λ < ρ 0 we consider the following values: λ = 1, 1.5, 2, 2.5, 2.9. Using Eqs. (9a)...(9d) we compute for these values of the parameters α and β the peak quantities: peak amplitude I P , the time t P (in days) of occurrence of the peak (computed by numerical evaluation of the integral (9d)), average number of infected I P and absolute value of the epidemic speed at the peak |V P | (in days −1 ). Moreover, using Eqs. (11), (12), (10), together with Eq. (9d), we compute numerically the total number I E of infected individuals during the epidemic and its whole time span t E (in days). The results are shown in Table I . Our results are in accordance with the scaling behaviour given by Eqs. (16a),(16b), (17)and (18). We see from Table I that the raising of the epidemic threshold for γ from the initial value γ 0 first to 2γ 0 then to 2.9 γ 0 let both the number of infected at the peak and their average number drastically sink from the order of magnitude 10 6 first to 10 5 and then to 10 3 − 10 4 . This reduction is the same independently of the fact that if it is achieved by reduction of α (way (1)) or by increase of β (way (2)). Similarly, the total number I E of infected individuals drops from the huge value 1 1.88 · 10 7 till 1.16 · 10 7 (for λ = 2) and then to 1.3 · 10 6 (for λ = 2.9)

On the other hand, the two ways of reducing I P andÎ P and I E affect differently the occurrence time of the peak t P and the whole time span of the epidemic t E . By acting on β (way (2)) instead of on α (way (1)) we can shorten these times by 33% (for λ = 1.5), by 50% (for λ = 2) and even reduce it by almost 1/3 (for λ = 2.9). Correspondingly, the speed of variation of the basic reproduction number |V P | will be enhanced by the same factors.

If the containment measures can manage to increase λ above 3, we go below the threshold for ρ and the epidemic does not start at all. Obviously, in a real situation reducing α or increasing β is not performed once for all at the beginning, but occurs in steps. Our main result is that in order to try to stop the epidemic it is much more convenient to rise β instead of lowering α because even if we do not manage to stop it, we are able to reduce its size and at the same time to shorten its timescale. Table I . Comparison of the effect of reduction of the infection rate α → (1/λ)α versus increase of the removal rate β → λβ on epidemic parameter: peak amplitude I P , average value of infectedÎ P , peak time t P , speed of basic reproduction number |V P | at the peak, total number of infected individuals I E and whole time-span of the epidemics t E . The total population is N = 2 · 10 7 , β 0 = 1/9 and α 0 = (1/6)10 −7 . The values of I P ,Î P , t P , |V P |, I E , t E are tabulated for values of λ = 1, 1.5, 2, 2.5, 2.9. For sake of clarity we also show in the table the values of α and β corresponding to a given value of λ.

In this paper we have analysed, in the context of the standard SIR model for epidemic dynamics, the impact of different containment measures on size (the epidemic peak I P , the average number of infectedÎ P and the total number of infected I E ), the timescale (the occurrence time of the peak t P and the whole time-span t E ) and the speed (time variation of the reproduction number |V P |) of epidemics. Using an exact solution for the epidemic dynamics we have been able to derive the scaling behaviour of these quantities under change of the two parameters (the infection rate α and the removal rate β) of the SIR model, which can be controlled by the containment measures. This allowed us to compare the impact on size, timescale and speed of the epidemic of containment measures acting either on α or on β

We have shown that for a given reduction of I P ,Î P , I E , the timescale and the speed of the epidemic are to a great extend sensitive to the kind of measures we put into play. By increasing the removal rate β instead of reducing the infection rate by a factor λ one can reduce the timescale of the epidemic by a factor 1/λ and increase the speed of the epidemics by a factor λ. In the case we have analysed in detail, namely the COVID-19 epidemic in Northern Italy, the reduction factor λ, in principle could also take values around 3.

An important point we have not addressed in this paper is the determination of the exact way in which the usual containment measures used to fight epidemic, impact on the values of the parameters α and β. Whereas it is quite clear that social distancing reduces the parameter α and does not change β, the effect of other measures like, early detection and contacts tracing is not a priori evident. Early detection and contact tracing increase β only if implemented on a large scale and followed by prompt and strict isolation of the detected infectives. If this is not the case, it is likely that these measures just bring a small reduction of α

The recent analysis of Gaeta [10] of the different strategies used in Northern Italy to tackle the COVID-19 epidemic seems to confirm this result. He found that simple early detection and contact tracing, while having an impact on the epidemic peak, do not substantially affect the timescale of the epidemic. On the other hand he also showed that contact tracing if followed by prompt isolation is the only efficient way to reduce the size of the epidemic, without having to live with it a long time. The Veneto experience shows that this was one of the factors underlying the success of the containment strategy in that region. Thus the main lesson one can draw from our results is that, epidemic containment measures focused on tracing, early detection followed by prompt removal of infected individuals are more efficient to fight epidemics than those based on social distancing.

Let us conclude this paper with some comments about the range of validity of our results. The SIR model is an oversimplified model for epidemic dynamics. Generalisations of it are necessary in order to give a good descriptions of real epidemics. For instance, in the case of the COVID-19 epidemic a generalization of the SIR model seems to be necessary in order to take into account the presence of a large set of a asymptomatic infective [9, 10, 11] . On the other hand, the SIR model gives the bare bones of deterministic epidemic dynamics. For this reason we believe that, at least at qualitative level, the main result of this paper -the possibility to reduce the epidemic peak keeping under control its timescale by acting on removal rates-could remain true for generalized and improved SIR-like models.

Contributions to the Mathematical Theory of Epidemics

Proc. R. Soc. Lond. A

Mathematical Biology. I: An Introduction

The Mathematics of Infectious Diseases

Epidemic Modeling: An Introduction

Data analysis for the COVID-19 early dynamics in Northern Italy

Data Analysis for the COVID-19 early dynamics in Northern Italy

A simple SIR model with a large set of asymptomatic infectives

Social distancing versus early detection and contacts tracing in epidemic management

Asymptomatic infectives and R0 for COVID

Exact analytical solutions of the Susceptible-Infected-Recovered (SIR) epidemic model and of the SIR model with equal death and birth rates

A note on the derivation of epidemic final sizes

Acknowledgements I thank Giuseppe Gaeta for several useful discussions and comments.