key: cord-265027-kvtqpgqq
authors: Chatterjee, S.; Sarkar, A.; Karmakar, M.; Paul, R.
title: Studying the progress of COVID-19 outbreak in India using SIRD model
date: 2020-05-13
journal: nan
DOI: 10.1101/2020.05.11.20098681
sha: 
doc_id: 265027
cord_uid: kvtqpgqq

We explore a standard epidemiological model, known as the SIRD model, to study the COVID-19 infection in India, and a few other countries around the world. We use (a) the stable cumulative infection of various countries and (b) the number of infection versus the tests carried out to evaluate the model. The time-dependent infection rate is set in the model to obtain the best fit with the available data. The model is simulated aiming to project the probable features of the infection in India, various Indian states, and other countries. India imposed an early lockdown to contain the infection that can be treated by its healthcare system. We find that with the current infection rate and containment measures, the total active infection in India would be maximum at the end of June or beginning of July 2020. With proper containment measures in the infected zones and social distancing, the infection is expected to fall considerably from August. If the containment measures are relaxed before the arrival of the peak infection, more people from the susceptible population will fall sick as the infection is expected to see a three-fold rise at the peak. If the relaxation is given a month after the peak infection, a second peak with a moderate infection will follow. However, a gradual relaxation applied well ahead of peak infection, leads to a two-fold increase in the peak infection. The projection of the model is highly sensitive to the choice of the parameters and the available data. Our model provides a semi-quantitative overview of the progression of COVID-19 in India, with model projections reasonably replicating the current progress. However, since the pandemic is an ongoing dynamic phenomenon, the reported results are subjected to regular updates in consonance with the acquired real data.

In the post-WW-2 era, the world probably has not witnessed such catastrophic morbidity and the looming threat of severe economic challenges caused by the worldwide outbreak of the disease COVID-19 caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). The detection of the disease in the human host was first reported in Wuhan, China on 31 December 2019 as a cluster of cases of pneumonia. As the highly contagious disease transmitted rapidly all over the globe, the outbreak was declared as a pandemic by the WHO on 11 March 2020. Tackling the outspread of the disease is found to be very challenging across the world for the following reasons: (a) conventional flu-like symptoms in human carriers and (b) human-to-human transmission via asymptomatic human hosts and (c) the absence of a proper clinical doctrine (e.g. vaccine, drugs, concrete ideas about the immunological response, etc.). Extensive testing and imposition of containment measures to maintain social distancing turn out to be the effective remedies to prevent disease transmission at the current stage of the epidemic at several places. To evaluate the impact of these preventive measures on infection spread, recovery, death tolls, and various other associated factors, mathematical models become useful in predicting realistic, quantitative estimates. A preliminary analysis * sspsc6@iacs.res.in † sspas2@iacs.res.in ‡ sspsc5@iacs.res.in § mcsmk2057@iacs.res.in ¶ raja.paul@iacs.res.in suggests that the classic mean-field Susceptible-Infected-Recovery-Death (SIRD) model by Kermack and McKendrick [1, 2] , can be used to obtain a quantitative picture of the epidemic [3] [4] [5] [6] [7] [8] . In this article, implementing the SIRD model, we report the temporal progress of COVID-19 transmission in India, various Indian states and compare it with some other countries around the world. A similar model used by Fernandez-Villaverde et al [5] provides a detailed overview of the pandemic situation in the USA and many other countries.

India implemented a nation-wide lockdown from March 25, 2020. On the day of the announcement of nation-wide lockdown India had about 657 corona positive cases while the first COVID-19 positive was detected on January 30, 2020. The socio-economic constraints in the Indian context alludes that: (a) 'too-prolonged' lockdown is difficult to sustain; (b) the sole imposition of containment measures without a manifold increase in testing capacity is a futile endeavor; (c) if the implementation of the lockdown measures is lenient, containment of the spread is highly improbable. Henceforth, the feasible solution for limiting the spread lies in carefully balancing various key epidemiological factors. That is where the importance of the current model predictions becomes relevant.

This study further highlights the effect of lockdown on the disease spread and predictions about the variability in the infection peak upon the severity of the containment measures (and/or the lack of it). The model predicts that, in India, the height of the peak infection decreases with stricter lockdown, but at the cost of 'time' (position of the peak shifts to a later month). Thus, with a large susceptible, the infection will stay for a long time if existing infections are not quarantined immediately or no proper medicine/vaccine is employed. The key is to quarantine the infection in small pockets while in lockdown and prevent inter-pocket transmission. The model further underlines that in the highly contagious zones ('red' zones where COVID-19 positive cases continue to grow), if the present lockdown is extended and reinforced with stricter quarantine measures, the new infections will gradually plummet down flattening the COVID-19 curve at a much faster rate.

Our study also explores the plausibility of universality in the spread of the COVID-19 outbreak amongst different countries [6] and compares the situation in India with few other countries (e.g. Germany, USA, South Korea, Spain) in the relevant time window (February -April). Due to the simplicity of the SIRD model, we found it difficult to fit the observed patterns of the pandemic using the available data. The real data for analysis in India's context is collected from the repository with an interactive interface hosted at https://www.covid19india. org. The data for other countries are taken from the repository with an interactive interface hosted at https: //www.worldometers.info/coronavirus. The purpose of this article is not to make any quantitative prediction that should be used to design policies, but for the research purpose only. MODEL We employ the standard SIRD model where the population N is divided into sub-population of susceptible (S), infected (I), recovered (R) and dead (D) for all times t. Thus, N = S +I +R+D. The following set of mean-field differential equations governs the temporal dynamics of the population of susceptible (S), infected (I), recovered (R), dead (D) and describes a comprehensive picture of the SIRD epidemic evolution:

Here, β, γ, δ are the parameters determining the characteristics of infection, recovery and deaths respectively (Fig. 1A) . When a susceptible person interacts with an infectious person, the susceptible become infected at a rate βSI/N . Large variability is observed in the rate γ that an infected individual is no longer infectious or equivalently has recovered in this simplified model. Literature [9] [10] [11] suggests that, on the average, infectiousness appears to start from 2-3 days before the symptoms are visible. The infectiousness increases to its peak before the arrival of the symptoms and remains for about 7-9 days after the peak infection. Thus an infected individual remains infectious for about 12 days on the average and then recover. In our preliminary analysis, we set the recovery rate γ ∼ 1/12, which however does not give the best fit for all the cases we studied. In essence, the numerical values of the model parameters are obtained from the best fit. Initial values (time t = 0 days) of the number of infected, recovered and deaths (I 0 , R 0 , D 0 ), are chosen from real data.

The choice for the initial number of susceptible (S 0 ) is quite difficult. In the absence of antibody, the entire population can be susceptible to the COVID-19 pandemic. Nevertheless, the geographical, social, and economical characteristics of a region (and various other demographic factors) can substantially influence this number. We used two different approaches to get an estimate of S 0 . First, we study the data for the large countries where the cumulative positive cases have reached closer to a plateau. Though, the infected population at the plateau can be determined only when the epidemic is over. Dividing this number by the total population of the country gives a fraction that appears to be of the order for 10 −3 for Germany, USA, Spain, Italy, and 10 −4 and 10 −5 for South Korea and China respectively. Thus an estimate of the susceptible may be obtained by multiplying the population of a country by this fraction. The number of susceptible obtained in this way, however, indicates a lower bound as many individuals with mild or no symptoms go unreported. Another possibility to estimate the fraction would be to test the number of positive cases by the number of tests carried out. This number would be an upper bound since there are many regions within a country that remains completely isolated and the populations in such pockets would not be susceptible. The ratio between the number of positive cases and the total number of tests for different countries are given in the following; the fraction is 0.159 for the USA, 0.016 for South Korea (as per data up to May 7), 0.1072 for Spain, 0.063 for Germany (as per data up to May 10). Conventionally, in epidemiological modeling S 0 ∼ N . In our simulation, we have reasonably varied S 0 within this range to obtain the best fit with real data in a case by case manner (i.e for India and other countries).

With the formulation of the model, comes the quantitative estimate of the speed at which the disease spreads across a population. In other words, from the deterministic SIRD model, the objective is to assess how fast a human carrier would infect people belonging to the population of susceptible? The quantity that determines the transmission speed of the pandemic is the effective reproduction number or replacement number (R e ) [12] . Often the basic reproduction number R 0 , defined as the average number of secondary infections that occur when an infec- tious person (primary or source of infection) is placed into a susceptible population, is used in the epidemiological models. R 0 can be estimated from the very early stage of the infection when the infectious person mixes freely with the susceptible population. Estimating R 0 is often challenging due to lack of unbiased data as all secondary infections cannot be determined exactly; especially for COVID-19, where asymptomatic cases are hardly identified (Fig. 1B) . The effective reproduction number (R e ), which we used in this study, evaluates the mean number of new infections (infected from the susceptible pool) directly transmitted/induced by a typical infected person and can vary over the entire duration of the infection (Fig. 1C) . In the SIRD model, R e can be represented as β/(γ + δ). From the best fit of the data, we find that γ >> δ, yielding R e ∼ β/γ. If R e > 1, the disease starts spreading in a population infecting more and more people, but spreading does not occur if R e falls below 1. It is easy to notice that longer a person remains infectious (i.e. 1/γ days), can give rise to very large R e even if the number of interactions per day (i.e. β) is small.

Containment measures in terms of social distancing and lockdown have been implemented world-wide to mitigate the transmission speed of the outbreak. We implemented the effect of lockdown in the model by modifying the infection rate and obtained the best-fit. We chose the following functional form of time-dependent infection rateβ(t) where it gradually decreases after the containment measures are enforced [5, 6] . Before lockdown, the infection rate is β which is constant. When the lockdown is imposed on day τ (counted from the initial time point t = 0 or day 0 as chosen in the simulation), the time-dependent infection rateβ(t) diminishes with every progressing day which is assume to vary exponentially in the following manner [6] :

Here, ζ ∈ [0,1] is the infection parameter (or interaction parameter) and T is the delay in the number of days before the effect of lockdown is visible in the propagation of infection. Without lockdown ζ=1, referring to rapid infection while ζ = 0 means that infection is contained (e.g no interaction between infected and susceptible population, hence no transmission). ζ ∈ [0,1] reflects the asymptotic mitigation of the infection rateβ(t), when containment measures are imposed. Lower the value of ζ, stricter is the containment measures (or the manifestation of the same) [5, 6] . Here the initially chosen value of β ends with ζβ. Essentially, the initial value of β determines the characteristic properties of the disease which depends on the effective interaction of people in a region, social behavior, density of population, etc. The terminal value ζβ reflects the effect of the containment and how the social distancing is being maintained. The functional form ofβ(t) is arbitrary and chosen semi-empirically to obtain the best fit with the available data. The model simulation, data analysis, and plotting are carried out in python. The analysis of the COVID-19 data, using the deterministic compartmental SIRD model, sheds light on the primary characteristics of the temporal evolution of the pandemic. Relevant parameter values chosen for the India and few Indian states are listed in Table S1 -S2. The best-fit parameters chosen for foreign countries are listed in Table S3 .

We carried out the SIRD model analysis on COVID-19 progression in India's context (and few other countries) All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. . https://doi.org/10.1101/2020.05.11.20098681 doi: medRxiv preprint with realistic variations in following parameters: rates of infection (β), recovery (γ) and deaths (δ), the initial number of susceptible (S 0 ) and the effective reproduction number (R e ). Detailed results are described in the following and illustrated in Fig. 2 -10, Fig. S1 -S2. In a nutshell, we start with the initial susceptible population (S 0 ) varying in the range ∼ 1-3 million S 0 , keeping the effective reproduction number R e fixed at ∼ 4.0, and show how the model prediction fits with the Indian data without a lockdown, the location of the infection peak and the relative deviation from the real data ( Fig. 2A) . The best fit is obtained by tuning the rates of infection (β), recovery (γ), and deaths (δ) keeping S 0 constrained in the mentioned range. Then, we incorporate the effect of containment-measures/lockdown in the functional form of β and show how the effect of the containmentmeasures has altered the location and the height of the infection peak (Fig. 2B ). Next, we explore how the variability in the effective reproduction number R e influences the infection peak ( Fig. 3-4 ). Furthermore, we analyze the COVID-19 progression in few Indian states e.g Kerala, Maharashtra, Delhi, and West Bengal ( Fig. 5-6 ) and foreign countries e.g South Korea, Germany, USA, Spain ( Fig. 7-8) . Lastly, we explore, in brief, what happens to the outspread, if the lockdown is lifted (in other words, containment measures are relaxed) in the Indian context ( Fig. 9-10 ).

India without lockdown: What could have happened?

The first COVID-19 positive human host was reported in India on 30th January 2020. The exponential growth of the number of infections, from 30th January onward, reached a number 657 on 25th March 2020, the day on which India imposed a nation-wide lockdown (Fig. S1A) .

Using the SIRD model, we first explored what could have happened, if the containment measures had not been undertaken. As mentioned earlier, we chose the factor ∼ 10 −3 (obtained in case of Germany and few other countries by dividing the cumulative population at infection peak by the actual population of the country) and multiplied it the Indian population of ∼ 10 9 to estimate the lower bound of the susceptible population (S 0 ). It turns out that it would be a 'good' estimation to have a 'working' S 0 in the range ∼ 10 6 . With susceptible population S 0 varied in the range ∼ 1-3 million (for fixed R e ∼ 4.0), the peak of the infection occurs in the first half of May ( Fig. 2A ). As expected, the peak height (infected population at the peak) increases with increasing S 0 . For an initial susceptible pool of S 0 ∼ 10 6 , the peak reaches a height of 0.4 million, whereas the peak jumps to ∼ 1.2 million for S 0 ∼ 3 million ( Fig. 2A) . The total death toll is estimated to reach about 30,000-100,000 for S 0 in range ∼ 1-3 million, during July -August, 2020 ( Fig. 2A) .

Next, we introduced the effect of containmentmeasures in the infection rateβ(t) (Eq. 5-6). Numerical analysis is carried out to investigate whether the pro-All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. . https://doi.org/10.1101/2020.05.11.20098681 doi: medRxiv preprint gression of the outbreak is mitigated after the lockdown is imposed.

Straightening of the growth curve:

From the real data, it appears that the infection rate begins to reduce, 15-20 days after the national lockdown is implemented ( Fig. 2A, inset) . We further observed that the growth curve for the infected population displays a straightening feature during the lockdown time frame. This is expected to be observed if containment measures are initiated; the unhindered exponential growth before the lockdown slows down due to the effect of containment measures during the lockdown. While slowing down and deviating from the exponential trajectory, the infection growth curve (time progression of the infected population size) acquires a distinctive straightening feature until the very recent surge ( Fig. 2A , inset).

Next, adding the lockdown effects into the picture, we fit the theoretically obtained infection growth curve with the real data. The best fit with the current set of parameters demonstrates that, due to the effect of the present lockdown, the infection peak dwarfs down to about 0.10 million from about 0.40 million in 'without lockdown' scenario (dashed curve, Fig. 2B and inset). The infection peak is projected to reach a peak at the end of June, tentatively ( Fig. 2B, 2B , inset). The estimated death toll also reduces substantially compared to the earlier scenario without containment measures.

However, the model also shows that the situation can be improved further. The infection growth curve can be dwarfed down further if the lockdown is extended and reinforced stringently in COVID-19 prone zones. In that case, the infection growth curve noticeably flattens with the infection peak reduced further.

As mentioned earlier, S 0 is a very crucial parameter in governing the position and the height of the infection peak. In the following, we summarize how the variations in the size of the susceptible population S 0 influence the infection growth curve.

Keeping R e fixed at ∼ 4.0, we varied the size of the total susceptible population within a range of 1-3 million ( Fig. 2A-2B ). The model analysis shows that the larger the size of the susceptible population, the higher the infection peak (Fig. 2) . Moreover, for the larger size of the susceptible population, attainment of the infection peak is delayed with the infection peak shifted to a later time zone (Fig. 2B, 2B , inset). These characteristic features are consistent in both without and with lockdown scenarios.

Thus, it is evident that the key to containing the outspread lies in keeping S 0 small. This is feasible only when interactions between a demographic region with the recent occurrence of infections and a region with no 'latest' instance of infection are strictly barred. Besides the 'global' lockdown (in a nation-wide sense), locally keeping an infected region isolated from other proximal unaffected regions may help to keep S 0 in check.

The next question that crops up is what happens to the magnitude of effective reproduction number (R e ) when containment measures are put in place. We discuss in the following, how the effective R e changes with time during the lockdown (Fig. 3) . Effective reproduction number (Re) and lockdown:

We start with R e in range 3.5 ≤ R e ≤ 4.4 in the beginning. As the lockdown is implemented, less number of people interact. Therefore, the effective infection rate (β (t)) starts decreasing over time. How much the reduction would be forβ(t) in longer time regime, is determined by the factor ζ in Eq. 5. The reducedβ(t) settles at a value ζβ due to the containment effects. Thus, if the recovery rate γ is fixed, the R e will diminish and reach a valueβ (t) γ ∼ ζβ γ . The decrease in R e due to the effect of lockdown is evident in Fig. 3 where the effective R e All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. . https://doi.org/10.1101/2020.05.11.20098681 doi: medRxiv preprint reduces to ∼ 40 % of its initial value (before lockdown). The smeared color shades enveloping the dashed lines label the variations in R e within the mentioned range. As expected, the higher the value of R e , the taller the infection peak. This feature is consistent both in presence and absence of lockdown (Fig. 3, 3, inset) .

Next, we investigate, whether the value of R e , extracted from the best fit with real data, is unique (of course with marginal variation) or the variation is nonmarginal. What would be the scenario if Re is large?

Instead of fixing R e in the beginning, we varied the rates of infection (β), recovery (γ), and deaths (δ) without any restriction on the resulting value of R e . We aimed to verify whether the best fit of real data with the theoretical curves (infection, recovery, and death) can be obtained for a set of (β, γ, δ), other than the already chosen values in Fig. 2-3 , with no apparent constraint on the values of R e .

We find that, for a fixed size of susceptible population of about 0.8-1.2 million, the real data can still be fitted with the theoretical curves, even if the R e is large (R e ∼ 7.4, Fig. 4 ). Similar to Fig. 2-3 , the active infection cases deviate from the theoretical infection curve without lockdown, as the enforced lockdown effectively slows down the progression of infection (Fig. 4, 4, inset) . Note here, that in comparison with Fig. 2-3 , the location and height of the infection peaks change (both in the cases of without lockdown and with lockdown), as effective R e is increased ∼ 2-fold (Fig. 4) . In tandem with the definition of R e , we observe that greater the value of R e , larger the size of the infected population (compare Fig. 2B , R e ∼ 4.0; and Fig. 4 , R e ∼ 7.4). It is also noteworthy to mention that here the recovery rate γ = 0.0315 day −1 correspond to about ∼ 31 days compared to 12 days as discussed earlier. A prolonged infectiousness leads to the rise in the R e and consequently the total number of infected people.

From the above observations, we connote that the exactness of R e can be ascertained, when we have more data points in the time evolution of the infected, recovered, and dead population. The current model setup may not be able to precisely pinpoint the exact 'real' R e .

How well the individual Indian states are doing?

In India, the first COVID-19 positive case was reported in Kerala on January 30, 2020, and now almost half of the active COVID-19 positive cases are from Maharashtra. In this note, we explore the COVID-19 progression in these states along with Delhi and West Bengal and compare the features of pandemic progression with each other (Fig. 5-6, S1B) .

After the first case being detected in Kerala on January 30, the second and third cases were reported on Feb 2-3. After February 3, there was no new case detected in Kerala till March 7. The previous three cases were all recovered within February 20. The 'second-wave' of infections started from March 8. From March 8 onwards, there was a rapid upsurge of infections. However, about two weeks after the national lockdown is imposed, Kerala reached its infection peak. It is evident from Fig. 5A -5B that the downfall of the infection is rapid, as the infection curve moved past its peak. If the lockdown was not enforced, the infection was projected to occur around mid-May. But, Fig. 5B alludes that Kerala implemented the containment measures so well that the infection peak occurred early at a much lower height (Fig. 5A-5B) . The model analysis further projects that due to the effect of lockdown, R e reduces to ∼ 18 % of its initial value during the upsurge of infections before lockdown. The reduced value of R e is << 1, which means Kerala is on the way to become a COVID-19 free state soon if the trend continues.

In Maharashtra, the first case was detected on March 9. The total infected population is yet to attain its peak. The projected infection peak would occur around the end of May or early June if the present trend continues and containment measures remain enforced in places (Fig. 5C-5D) .

Similarly, in Delhi and West Bengal, the infection growth curves are yet to attain their respective peaks (Fig. 6A-6D ). The first cases in these states were reported on March 2 and March 17 respectively. The peaks are projected to be reached at the end of June and mid-All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. July for Delhi and West Bengal respectively, if the enforced lockdown remains deployed and the current trend continues (Fig. 6B, 6D ).

It is important to note that, in states, Maharashtra, Delhi and West Bengal, the estimated R e plummets down to value > 1.0, even after staying months under lockdown. Among the Indian states we analyzed, Kerala turns out to be the only exception where the effective R e reduces to a value << 1, meaning that further 'out-ofbound' spreading is unlikely to occur there if the current trend is followed.

It is evident from Fig. 7A-7D and S1A that both Germany and South Korea have moved past the infection peak. The infected population is decreasing day by day in those countries. The best fit with real data is obtained for the initial R e ∼ 3.0 and ∼ 4.5 for Germany and South Korea respectively. However, as the containment measures were undertaken in those countries, the effective transmission (or R e ) reduced to ∼ 20-25 % from the initial values for the respective countries (Fig. 7B,7D ). This observation, suggests that the counter-measures to fight the pandemic (e.g containment measures, social dis-All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. . https://doi.org/10.1101/2020.05.11.20098681 doi: medRxiv preprint tancing, quarantining, testing, etc.), undertaken in these countries, were reasonably successful in repressing the outspread. Moreover, the reduction of R e to 20-25 % of its initial values, rescales the R e for the respective countries to a value < 1.0 which alludes that new infections are declining and any more 'out-of-bound' infection growth is unlikely to occur if the current trend is followed.

We analyzed COVID-19 progression data for two more countries: USA and Spain (Fig. 8A-8D) . The USA is approaching the infection peak and will reach its peak shortly if the current trend continues (Fig. 8B) . However, contrary to the USA, Spain has already passed the infection peak and the infected population is decreasing gradually (Fig. 8D ). Spain imposed a nation-wide lock-down on March 14. Model analysis (fitting parameter optimization) suggests that, due to the effect of lockdown, R e for Spain reduced to 25 % of its initial value. But in the USA, the reduction in R e is only ∼ 42 % meaning that the implementation of local containments was not that stringent.

Contrary to the infection curves for Germany, South Korea, Spain (Fig. 7-8) , India is yet to reach the infection peak (the USA is about to reach the peak.). The following remarks briefly summarize where India stands compared to these countries. 1. The effective containment during the present lockdown in India indicates that the infected population might reach its peak at the end of June (Fig. 2) whereas Ger-All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. many, South Korea, and Spain have already moved past the peak and daily new infections are decreasing ( Fig.  7-8 ). 2. Since India has a large population, the infection is expected to stay for a longer duration. Germany, South Korea, and Spain might have the advantage of a smaller population of the susceptible. We allude that the higher the actual population of a country, the higher would be the effective size of the susceptible pool for that country while making the previous statement. The key is to contain the infections in small zones and prevent transmission between infectious and non-infectious zones. 3. The growth of the infected population in Germany, South Korea, and Spain were greater than that experi-enced in India which gave India an additional advantage of 'buying precious time'. Slow growth rate alludes to a smaller peak value at the zenith of the infection. However, as mentioned earlier, the height of the peak is subjected to the effective size of the initial susceptible pool (S 0 ).

For better clarity and wider accessibility to general readers, we discuss and summarize the important observations from our study in Q&A form in the following: All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. . https://doi.org/10.1101/2020.05.11.20098681 doi: medRxiv preprint How accurate are the model predictions?

The SIRD model is a drastically simplified approach to thoroughly understand the dynamics of COVID 19. From the available data, it is now clear that a susceptible person goes through a latent period of 2-3 days after coming in contact with an infected individual. Subsequently, the person remains infectious for several days (∼5 days). The infectious individual may or may not develop symptoms. The current model does not incorporate any of these de-tails and hence fitting is imperfect. Moreover, data used to fit with the model also vary between different locations leading to uncertain predictions.

How sensitive are the model predictions to parameter variations?

We investigated the sensitivity of the model to parameter variations, focusing in particular on the parameters All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. . https://doi.org/10.1101/2020.05.11.20098681 doi: medRxiv preprint To investigate the universality in the COVID-19 outspread across different countries, we looked into iterative time lag maps for the cumulative confirmed infected (C = I+R+D), recovered (R), and dead (D) population [6] . Using the iterative maps, we try to extract the correlation between a population on the day n and day n + 1. From the recurrence plots (population count on n th day vs population count on (n + 1) th day) in Fig. S2 , we ob-All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. . https://doi.org/10.1101/2020.05.11.20098681 doi: medRxiv preprint serve that the real data for all the cases follow the same power law of the following kind: f (x) = ax b . The factor and exponent a and b are similar for all the countries considered in the plots. This finding indicates that there exists an underlying universality in the outspread of the pandemic across various countries. Rapid lifting of the local containment paves the way for inter-mingling between regions with no positive cases (green) and regions with positive cases found lately (red). This, in turn, results in rapid transmission of the disease across zones (all zones becoming red), rendering the purpose of preceding lockdown futile. On the contrary, initially, the partial lifting of local containment only in the green zones bars the import of transmissions from red zones. When a red zone becomes green, the local containment can be lifted from that region. Due to the effect of this gradual lifting, the red zones diminish over time, with 'greens' taking over the 'reds'.

How effective are the quarantine measures? Can quarantining single-handedly contain the transmission?

A common perception of flu and other infectious diseases is that an infected individual spreads infection when symptoms appear. In the case of seasonal flu, infection mostly occurs when a person has symptoms [13] . However, as we understand from the literature survey, an individual with COVID-19 would be contagious before developing symptoms. The incubation period for COVID-19 is 5 days, and maximum infectiousness appears to be 2-3 days before the symptoms appear. Thus infection spread by an individual is maximum he/she becomes sick [10, 14] . Due to limitations of the testing procedure diagnosis takes about 5 days after symptoms are visible, i.e., 10 days from the day of infection. Clearly, on the average, an infected individual is beyond the peak of maximum infectiousness after this time. Thus, a reduced rate of infection demands early diagnosis and isolation of positive patients. This means that a COVID-19 patient needs to be identified in the pre-symptomatic stage as evidence suggests the infectiousness of the patient before developing symptoms which is extremely challenging (effectively, RT-PCR needs to be carried out for every individual who might have come in contact with the patient). The epidemic becomes even more complex due to a majority of the infected individual who develops mild or no symptoms [14, 15] . Therefore even with isolating/quarantining all the infected COVID-19 would not be eliminated for two reasons: a) normally an individual would be tested after symptoms appear which is when he/she has passed the peak of the contagiousness, b) asymptomatically infected person, in general, are not tested but he/she is also contagious like the symptomatic individual.

Gradual relaxation of the containment versus extended lockdown?

Relaxing the containment: We have investigated the effect of relaxing the containment measures at 3 different time points for India ( Fig. 9A-9D) . We find if the containment is relaxed before the peak infection is reached at the end of June, the infection would rise rapidly to a great extent (Fig. 9A) . The peak height reduces, if the containment measures are relaxed, when the infection is close to the peak, a time point around the 3rd week of June (Fig. 9B) . However, if the relaxation occurs a month after the peak infection, a second peak arrives which is lower than the first infection-peak (Fig. 9C) . A third possibility is to gradually relax the containment measures after May 17. The model shows that in this case the original peak does not shift its position but becomes two-fold higher than before (Fig. 9D) . A gradual relaxation could be carried out in steps: (a) First, identify all the sensitive (red) and safe (green) zones having positive and no cases respectively. Smaller the size of such zones, easier they can be managed by the administration, and necessary supplies can be arranged. It is important to seal the boundary of the red zones. (b) Test for new cases carrying symptoms and randomly test a few having no symptoms. (c) Dissolve the boundary between red and neighboring green zones once the red zone does not report a case for two weeks. This process will increase the size of the green zone where more and more people can communicate and business can restart. Successively extending the relaxation from the local neighborhood to the cities, districts, states, the containment measures can be relaxed across the country. Nevertheless, social distancing is mandatory even after the containment is officially lifted as there might be many undetected cases that can trigger the spread of the disease again. A schematic diagram in Fig. 10 summarizes the above-mentioned steps of relaxation and the consequential aftermaths, pictorially. In a nutshell, 'too-early' lifting of containment measures, All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. . https://doi.org/10.1101/2020.05.11.20098681 doi: medRxiv preprint long before the infection reached its peak, makes the purpose of lockdown 'null-and-void'. The reduced infection rate β, again starts increasing to its initial value yielding a larger R e for the outspread.

Our model projects that with containment measures in effect the number of active infection would be about 80,000 -140,000 at the peak which is expected to appear sometime after mid-June (Fig. 3) . Note that, due to containment the number of susceptible populations decreases drastically to only several million while the total population of India is about 1.35 billion. If the lockdown is suddenly relaxed locally after May 17, the peak infection would rise sharply to 0.3 -0.4 million (Fig. 9A) . Note that, this number is estimated without altering the susceptible population which is about 10 −3 times the actual population observed for many large countries. The remaining population considered to be shielded from the infection due to containment and demographic segregation. In the event of complete mixing of the population of the entire country, the peak infection might see a 10 3 fold rise which would be challenging for any health care system. Thus, social distancing measures must remain unless the infectious population is drastically removed.

An important aspect of COVID-19 is the number of patients who do not develop any symptoms (Fig. 1B) . This means they would be infecting healthy people unknowingly. According to the WHO and Indian Council of Medical Research (ICMR), as much as 80% of the infected individual can be asymptomatic. Thus all the symptomatic cases reported so far, contributes only about 20% of the total infection. Going by the reported number of cumulative infection 61,356 as on May 9, almost all of which are symptomatic, this would correspond to about 245,424 people who also had the virus but did not show any symptoms. Together, about 306,780 people have actually been infected so far in India carrying symptoms or no symptoms. Therefore, the number of people in the country who are still susceptible to the infection is sim 1.37 billion. One can realize that, with nearly 50,000 active infections, free mixing of the countrywide population after 17 May would cause a huge surge in the total number of infections which is nearly impossible to manage by any health care system.

It is noteworthy to mention that the total number of cases reported in all over India as well as in various Indian states are negligible compared to the total population of the country and states respectively. Besides, the severity of the infection with symptoms is relatively less in India than in the USA and other large European countries. Whether it is due to the effect of hot and humid weather of India or other meteorological parameters such as high UV index, future research would be able to evaluate. 

initial susceptible population 1.0-3.0 × 10 6 TABLE S1. List of parameters chosen for the best fit with real data in Indian context. All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 13, 2020. . https://doi.org/10.1101/2020.05.11.20098681 doi: medRxiv preprint

Mathematical Models in Epidemiology

Epidemic Modelling: An Introduction

No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity

ACKNOWLEDGMENTS S.C. * and A.S. were supported by a fellowship from the University Grants Commission (UGC), India. M.K. was supported by a fellowship from CSIR, India. S.C. ‡ thanks Indian Association for the Cultivation of Science, Kolkata for financial support. R.P. thanks IACS for support and Grant No. EMR/2017/001346 of SERB, DST, India for the computational facility.