Chaos, Solitons and Fractals 135 (2020) 109864

Contents lists available at ScienceDirect

Chaos, Solitons and Fractals

Nonlinear Science, and Nonequilibrium and Complex Phenomena
journal homepage: www.elsevier.com/locate/chaos

Time series forecasting of COVID-19 transmission in Canada using

LSTM networksR
Vinay Kumar Reddy Chimmula∗, Lei Zhang
Faculty of Engineering and Applied Science, University of Regina, Regina, Saskatchewan, S4S0A2 Canada

a r t i c l e i n f o a b s t r a c t

Article history: On March 11th 2020, World Health Organization (WHO) declared the 2019 novel corona virus as global
Received 6 April 2020
pandemic. Corona virus, also known as COVID-19 was first originated in Wuhan, Hubei province in China
Revised 4 May 2020
around December 2019 and spread out all over the world within few weeks. Based on the public datasets
Accepted 4 May 2020
Available online 8 May 2020
provided by John Hopkins university and Canadian health authority, we have developed a forecasting
model of COVID-19 outbreak in Canada using state-of-the-art Deep Learning (DL) models. In this novel
Keywords: research, we evaluated the key features to predict the trends and possible stopping time of the current
Epidemic transmission COVID-19 outbreak in Canada and around the world. In this paper we presented the Long short-term
Time series forecasting memory (LSTM) networks, a deep learning approach to forecast the future COVID-19 cases. Based on the
Machine learning results of our Long short-term memory (LSTM) network, we predicted the possible ending point of this
Corona virus
outbreak will be around June 2020. In addition to that, we compared transmission rates of Canada with
COVID-19
Italy and USA. Here we also presented the 2, 4, 6, 8, 10, 12 and 14th day predictions for 2 successive
Long short term memory (LSTM) networks
days. Our forecasts in this paper is based on the available data until March 31, 2020. To the best of our
knowledge, this of the few studies to use LSTM networks to forecast the infectious diseases.
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction matical models presented in [6,7] helped clinicians to better un-

derstand the characteristics of human liver and transmission of
Every infectious disease outbreak exhibits certain patterns and dengue outbreak.
such patterns needed to be identified based on transmission dy- Most of the Data driven approaches used in previous stud-
namics of such outbreaks. Intervening measures to eradicate such ies [8] are linear methods and often neglects the temporal com-
infectious diseases rely on the methods used to evaluate the out- ponents in the data. They depend upon regression without non-
break when it occurs. Any outbreak in a country or province usu- linear functions and failed to capture the dynamics of transmis-
ally occurs at different levels of magnitude with respect to time i.e. sion of infectious diseases like novel corona virus. Statistical mod-
seasonal changes, adaptation of virus over time. Usually patterns els such as Auto Regressive Moving Average (ARIMA), Moving Aver-
exhibited in such scenarios are non-linear in nature and this mo- age (MA), Auto Regressive (AR) methods overwhelmingly depends
tivates us to design the system that can capture such non-linear on assumptions and such models are difficult for forecasting real-
dynamic changes. With the help of these non-linear systems, we time transmission rates. Wide range of statistical and mathematical
can describe the transmission of such infectious diseases. In [1,2] a models [9,10] have been proposed to model the transmission dy-
transmission model for malaria and in [3] a mathematical model namics of current COVID-19 epidemic. In many cases, these models
for analysing dynamics of tuberculosis has been developed to study are not able to fit the given data perfectly and accuracy is also low
the transmission using mathematical models. In [4] a laplacian while predicting the growth of COVID-19 transmission.
based decomposition is used to solve the non-linear parameters R0 is a popular statistical method specifically used to model
in a Pine Witt disease. A modified SIRS model in [5] successfully an infectious disease. Often referred as âreproduction numberâ be-
helped to control the syncytial virus in infants. Similarly mathe- cause, the infections reproduce itself with respect to time. R0 fore-
casts the number of people can get the infection from the infected
person. In this model, an extra weight is applied to the person
R
This document is the results of the research project funded by the Saskatchewan who never infected the current disease nor vaccinated. If the value
Centre for Patient Oriented Research (SCPOR), Saskatchewan, Canada
∗
of R0 of a disease is 10, then the infected person will spread the
Corresponding author.
disease to 10 other people surrounding him. In [11] authors used
E-mail addresses: vcw015@uregina.ca (V.K.R. Chimmula), lei.zhang@uregina.ca (L.
Zhang). R0 method to find the infection rate of novel virus on diamond

https://doi.org/10.1016/j.chaos.2020.109864
0960-0779/© 2020 Elsevier Ltd. All rights reserved.
2 V.K.R. Chimmula and L. Zhang / Chaos, Solitons and Fractals 135 (2020) 109864

princes cruise ship [11]. However, in such method it is difficult to The rest of this paper is structured as follows: section II de-
find the starting point of the infectious disease by identifying pa- scribes methods, datasets and LSTM models used in this paper. In
tient zero and the people he interacted with during his incubation Section III, we have discussed our findings and in Section IV, con-
period. It is worth noting that mathematical models presented in cussion and future work was discussed
[12–14] can be used to solve the complex non-linear patterns of
infectious diseases. 2. Methods and models
Even though these epidemiological models are good at captur-
ing vital components of an infectious disease, parameters of these 2.1. Dataset
models required several assumptions. Such hypothesized parame-
ters would not fit the data perfectly and precision of such mod- The COVID-19 data used in this research is collected from Johns
els will be low. Meanwhile, in engineering applications [15], model Hopkins University and Canadian Health authority, provided with
parameters are calculated with the help of real-time data. Similar number of confirmed cases until March 31, 2020. The data set also
approach was used in this research to find the model parameters includes number of fatalities and recovered patients by the end of
instead of assumptions. each day. The dataset is available in the time series format with
In order to overcome the barriers of statistical approaches, date, month and year so that the temporal components are not
we developed the Deep Learning based network to predict the neglected. A wavelet transformation [25] is applied to preserve the
real-time transmission. Our model could help public health care time-frequency components and it also mitigates the random noise
providers, policy makers to make necessary arrangements to tackle in the dataset. The fundamental point to represent and forecast the
the rush of potential COVID-19 patients. This experiment is based trends of current is to select conventional functions to fit the data.
on the data sets of confirmed COVID-19 cases available until March The COVID-19 dataset is divided into training set (80%) on which
31, 2020. our models are trained and testing set (20%) to test the perfor-
Artificial Intelligence and mobile computing are one of the key mance of the model.
factors for the success of technology in health care systems [16].
In the world of smart devices, data is being generated in the un- 2.2. LSTM Network for modelling time series
precedented way than ever before and promoted the role of ma-
chine learning in healthcare [16]. The world today is more con- A large part of real-world datasets are temporal in nature. Due
nected than ever before this helped to share the real time infec- to its distinctive properties, there are numerous unsolved problems
tious data between the countries. The distinctive feature of artifi- with wide range of applications. Data collected over regular inter-
cial intelligence is its flexibility, domain adaptation and econom- vals of time is called time-series (TS) data and each data point is
ical to integrate with existing systems. Over the last few weeks, equally spaced over time. TS prediction is the method of forecast-
many researchers came up with several mathematical models to ing upcoming trends/patterns of the given historical dataset with
predict the transmission of novel corona virus [17,18]. The ma- temporal features. In order to forecast COVID-19 transmission, it
jor drawbacks of the existing models are linear, non-temporal and would be effective if input data has temporal components and it
several assumptions while modelling the network. First of all, the is different from traditional regression approaches. A time series
covid-19 is a time series data set and it is highly recommended (TS) data can be break downed into trend, seasonality and error.
to use the sequential networks to extract the patterns from it. A trend in TS can be observed when a certain pattern repeats
Second of all, the data we are dealing with is dynamic in nature on regular intervals of time due to external factors like lockdown
so by using statistical and epidemiological models, results are of- of country, mandatory social distancing, quarantines etc. In many
ten vague [19,20]. In [21–24] researchers used deep learning based real-world scenarios, either of trend or seasonality are absent. Af-
LSTM networks to forecast COVID-19 infections. The LSTM models ter finding the nature of TS, various forecasting methods have to
used in the above networks could not able to represent the spatio- be applied on given TS
temporal components simultaneously. In this paper we addressed Given the TS, it is broadly classified into 2 categories i.e. sta-
the above problem by modifying the internal connections. In our tionary and non-stationary. A series is said to be stationary, if it
modified LSTM cells, We have established the alternative connec- does not depend on the time components like trend, seasonality
tions between the input and output cells. This type of connections effects. Mean and variances of such series are constant with re-
not only helps the networks to preserve spatio-temporal compo- spect to time. Stationary TS is easier to analyze and results skil-
nents, but also to transfer the historical information to the next ful forecasting. A TS data is said to non-stationary if it has trend,
units. seasonality effects in it and changes with respect to time. Statis-
In this paper, we made an effort to predict the outbreak of tical properties like mean, variance, sand standard deviation also
COVID-19 based on past transmission data. First of all, coherence changes with respect to time.
of input data needs to be analyzed in order to find the key feature In order to check the nature (stationarity and non-stationarity)
i.e. number of new cases reported with respect to the previous day of the given COVID-19 dataset, we have performed Augmented
infections. After selecting the key parameters of the network, sev- Dickey Fuller (ADF) test [26] on the input data. ADF is the stan-
eral experiments was conducted to find the optimal model that can dard unit root test to find the impact of trends on the data and
predict future infections with minimum error. Previous studies on its results are interpreted by observing p-values of the test. If P is
COVID-19 predictions, did not considered the recovery rate while between 5-1%, it rejects the null hypothesis i.e. it does not have a
developing the model. In this research, we considered the recovery unit root and it is called stationary series. If P is greater than 5% or
rate as one of the features while building our model. From the de- 0.05 the input data has unit root so it is regarded as non-stationary
sign point of view, when a crisis occurs, algorithms tend to assign series.
high probability and completely neglects the previous information Before diving into the model architecture, it is crucial to explain
which leads to biased predictions. We addressed this issue in our the internal mechanisms of LSTM networks and reasons behind us-
literature and solved this by using LSTM networks. ing it instead of traditional Recurrent Neural Networks. Recurrent
Our results are expected to alert the public health care LSTM networks has capability to address the limitations of tradi-
providers of Canada to prepare themselves for the crisis against tional time series forecasting techniques by adapting nonlinearities
COVID-19. With the help of this real-time forecasting tool, front- of given COVID-19 dataset and can result state of the art results
line clinical staff will be alerted before the crisis. on temporal data. Each block of LSTM operates at different time
 V.K.R. Chimmula and L. Zhang / Chaos, Solitons and Fractals 135 (2020) 109864 3

the output of forget gate at time step âtâ for generating the out-
put. The internal block diagram of LSTM block used in this study
is shown in 1
The motivation behind initiating self-loops is to create a path so
that gradients or weights can be shared for long durations. Espe-
cially, this is useful while modelling deep networks where vanish-
ing gradient is a frequent issue to deal with. By adjusting weights
as self-looped gates, we can adjust the time scale to detect the dy-
namically changing parameters. Using the above techniques, LSTMs
Fig. 1. LSTM internal architecture. are able to produce the state-of-the-art results in [27]. The network
architecture used in this study is shown in 2

step and passes its output to next block until the final LSTM block 3. Results and discussion
generates the sequential output.
As of this writing, RNNs with blocks (LSTM) are the efficient The methods used in this study are based on data guided ap-
algorithms to build a time series sequential model. The fundamen- proaches and are completely different from previous studies. Our
tal component of LSTM networks is memory blocks, which was in- approaches and predictive outcomes will provide assistance for re-
vented to tackle vanishing gradients by memorizing network pa- stricting the infections and possible elimination of current COVID-
rameters for long durations. Memory block in LSTM architecture 19 pandemic. We trained our network with data until March 31,
are similar to the differential storage systems of a digital systems. 2020 reported by Canadian health authority. In this study we found
Gates in LSTM helps in processing the information with the help of that policies or decisions taken by government will greatly af-
activation function (sigmoid) and output is in between 0 or 1. Rea- fect the current outbreak.Several studies on forecasting of COID-
son behind using sigmoid activation function is because, we need 19 transmission are based on the R0 method however, they didn’t
to pass only positive values to the next gates for getting a clear include the sensitivity analysis to find the important features. We
output. The 3 gates of LSTM network are represented with the fol- examined our model predictions using mean square error (MSE).
lowing equations below: In Fig. 4 we plotted the total number of confirmed cases and fore-
casted COVID-19 cases in Canada as a function of time. From the
Jt = sigmoid (wJ [ht−1 , kt ] + bJ ) (1)
figure we can observe that, Canada didnât witness its peak yet and
it is expected number of cases will soon increase exponentially de-
Gt = sigmoid (wG [ht−1 , kt ] + bG ) (2) spite the social distancing.
Although our model achieved better performance when com-
pared with other forecasting models, it is unfortunate that trans-
Pt = sigmoid (wP [ht−1 , kt ] + bP ) (3)
missions are following increasing trend. The rate of infections in
Where: Jt = function of input gate USA, Italy and Spain are growing exponentially meanwhile, the
Gt = function of forget gate number of infections in Canada are increasing linearly in Fig. 3. If
Pt = function of output gate Canadians follow the regulations strictly, the number of confirmed
Wx = coefficients of neurons at gate (x) cases will soon decline.
Ht−1 = result from previous time step In our LSTM model-1 we trained and tested our network on
kt = input to the current function at time-step t Canadian dataset; the RMSE error is 34.83 with an accuracy of
bx = bias of neurons at gate (x) 93.4% for short term predictions in Canada. Meanwhile, based on
Input gate in the first equation gives the information that needs our testing/validation dataset the RMSE error is about 45.70 with
to be stored in the cell state. Second equation throws the informa- an accuracy of 92.67% for long term predictions. The predictions of
tion based on the forget gate activation output. The third equation LSTM model are shown in 4 with solid red line. It shows that our
for output gate combines the information from the cell state and model was able to capture the dynamics of the transmission with

Fig. 2. LSTM Architecture.

4 V.K.R. Chimmula and L. Zhang / Chaos, Solitons and Fractals 135 (2020) 109864

Fig. 3. a) Number of infections in Canada vs USA vs Italy as of March 31, 2020. b) Distribution of confirmed cases in Canada as of March 31, 2020.

Fig. 4. Predictions of the LSTM model on current exposed and infectious cases (Red solid line). The red dotted lines represents the sudden changes from where number
of infections started following exponential trend. The black dotted lines in the figure represents the training data or available confirmed cases. (For interpretation of the
references to colour in this figure legend, the reader is referred to the web version of this article.)

minimum loss. From the Fig. 4 we can say that Canada witnessed parameters and several unknown variables led to uncertainty in
linear growth in cases until March 16, 2020 after its first confirmed their predictions. LSTM model is different from statistical meth-
case. The current epidemic in Canada is predicted to continue until ods in many ways for instance, the proposed LSTM network fits
June 2020. Our second LSTM model-2 is trained on Italian dataset the real-time data and without any assumptions while selecting
to predict short-term and long-term infections in Canada. For short hyperparameters. It was able to overcome the parameter assump-
term predictions, the RMSE error is about 51.46 which is higher tions using cross validation and achieved better performance by
than previous model. According to this second model within 10 reducing the uncertainty. After reaching the inflection point, the
days, Canada is expected to see exponential growth of confirmed recovery rate will start decrease rapidly and death rate may in-
cases. crease at the same time as shown in Fig. 5. In order to find the
It was a challenging task to forecast the dynamics of trans- trend of the infections we decomposed the given series and the
mission based on small dataset. Even though COVID-19 outbreak trend of infections is increasing with respect to time. Further, num-
started in Canada around early January, the consistent epidemio- ber of infections followed increasing trend from Sunday to Tues-
logical data wasn’t released until early February. Because of small day and followed decreasing trend until Saturday as shown in
dataset several statistical models struggled to select the optimal Fig. 6.
 V.K.R. Chimmula and L. Zhang / Chaos, Solitons and Fractals 135 (2020) 109864 5

Fig. 5. a) Mortality rate of COVID-19 in Canada and the average mortality stands around 3.2% b) Recovery rate of COVID-19 patients shows that it is decreasing with respect
to time because of rise in number of infections.

Fig. 6. Trend of infections in Canada.

As we are still under the stage of dilemma about the current current pandemic is that, people travelling between the provinces.
situation of COVID-19 because, the accuracy of our estimates is Based on our sensitivity analysis our projections may go down if
bounded with a lot of external factors. So, it is recommended to current trials on potential vaccines achieves fruitful results. Finally,
conduct the follow-up study after this experiment to be more pre- in order to minimize the bias on our training algorithm we intro-
cise about the dynamics of this novel infectious disease. The actual duced regularization.
number of cases might be higher than the cases reported by the Further, by training our network inversely, we found that out-
government because, of the backlog of test results and some peo- break in Canada started around early January but, it was not re-
ple will be immune before even testing. All the above factors may ported until January last week. Even without the knowledge of 1st
lead to discrepancy of our model estimations. Even though we ad- case, our inverse training will help governments to better under-
dressed data imbalance by using statistical methods like interpola- stand the outbreak of COVID-19 and helps then to prevent such
tion and re-sampling yet we couldnât represent patients who are outbreaks in future.
on incubation period or not tested. Other problem while modelling
6 V.K.R. Chimmula and L. Zhang / Chaos, Solitons and Fractals 135 (2020) 109864

4. Conclusion and future work CRediT authorship contribution statement

The patterns from the data reveals that prompt and effective Vinay Kumar Reddy Chimmula: Conceptualization, Methodol-
approaches taken by Canadian public health authorities to mini- ogy, Software, Writing - original draft, Writing - review & editing.
mize the human exposure is showing a positive impact when com- Lei Zhang: Supervision, Writing - review & editing, Data curation.
pared with other countries like USA and Italy 3. Rate of transmis-
sion in Canada is following linear trend while in USA is witnessing References
an exponential growth of transmissions. However, it is too early to
[1] Basing A, Tay S. Malaria transmission dynamics of the anopheles mosquito in
draw the conclusions about the current epidemic. kumasi, ghana. Int J Infect Dis 2014;21:22.
After simulations and data fitting, our model predicted Canada [2] Chitnis N, Cushing JM, Hyman J. Bifurcation analysis of a mathematical model
would reach peak within 2 weeks from now. However, the cur- for malaria transmission. SIAM J Appl Math 2006;67(1):24–45.
[3] Sharomi O, Podder CN, Gumel AB, Song B. Mathematical analysis of the trans-
rent outbreak resembles early 20th century Spanish flu [28], which mission dynamics of hiv/tb coinfection in the presence of treatment. Math
killed millions of people and lasted for 2 years. Based on our model Biosci Eng 2008;5(1):145.
simulations, the current COVID-19 pandemic is expected to end [4] Shah K, Alqudah MA, Jarad F, Abdeljawad T. Semi-analytical study of pine wilt
disease model with convex rate under caputo–febrizio fractional order deriva-
within 3 months from now. Due to some unreported cases, a small
tive. Chaos Solitons Fractals 2020;135:109754.
number infection clusters may appear until December 2020. How- [5] Jajarmi A, Yusuf A, Baleanu D, Inc M. A new fractional hrsv model and its op-
ever, recent technological improvements and international coop- timal control: a non-singular operator approach. Physica A 2019:123860.
[6] Baleanu D, Jajarmi A, Mohammadi H, Rezapour S. A new study on the math-
eration between countries may even reduce the duration current
ematical modelling of human liver with caputo–fabrizio fractional derivative.
pandemic. Chaos Solitons Fractals 2020;134:109705.
To sum up, this is the first study to model the infections disease [7] Jajarmi A, Arshad S, Baleanu D. A new fractional modelling and control strategy
transmission model to predict the gravity of COVID-19 in Canada for the outbreak of dengue fever. Physica A 2019;535:122524.
[8] Knight GM, Dharan NJ, Fox GJ, Stennis N, Zwerling A, Khurana R, et al. Bridging
using deep learning approaches. Based on our current findings, the gap between evidence and policy for infectious diseases: how models can
provinces that have implemented social distancing guidelines be- aid public health decision-making. Int J Infect Dis 2016;42:17–23.
fore the pandemic has less confirmed cases than other provinces [9] Benvenuto D, Giovanetti M, Vassallo L, Angeletti S, Ciccozzi M. Application of
the arima model on the Covid-2019 epidemic dataset. Data Brief 2020:105340.
3. For instance, Saskatchewan issued social distancing guidelines 2 [10] Dehesh T, Mardani-Fard H, Dehesh P. Forecasting of covid-19 confirmed cases
weeks ahead than Quebec which has half of the confirmed cases in different countries with arima models. medRxiv 2020.
in Canada. Our results could help Canadian government to moni- [11] Zhang S, Diao M, Yu W, Pei L, Lin Z, Chen D. Estimation of the reproduc-
tive number of novel coronavirus (Covid-19) and the probable outbreak size
tor the current situation and use our forecasts to prevent further on the diamond princess cruise ship: a data-driven analysis. Int J Infect Dis
transmissions. 2020;93:201–4.
[12] Baleanu D, Jajarmi A, Sajjadi SS, Asad JH. The fractional features of a
harmonic oscillator with position-dependent mass. Commun Theor Phys
Declaration of Competing Interest 2020;72(5):55002.
[13] Yıldız TA, Jajarmi A, Yıldız B, Baleanu D. New aspects of time fractional optimal
control problems within operators with nonsingular kernel. Discret Cont Dyn
1. Conflict of Interest
Syst-S 2020;13(3):407.
Potential conflict of interest exists: [14] Jajarmi A, Baleanu D, Sajjadi SS, Asad JH. A new feature of the fractional eu-
We wish to draw the attention of the Editor to the following ler-lagrange equations for a coupled oscillator using a nonsingular operator ap-
facts, which may be considered as potential conflicts of interest, proach. Front Phys 2019;7:196.
[15] George D, Huerta E. Deep learning for real-time gravitational wave detec-
and to significant financial contributions to this work: The nature tion and parameter estimation: results with advanced ligo data. Phys Lett B
of potential conflict of interest is described below: 2018;778:64–70.
No conflict of interest exists. [16] Panch T, Mattie H, Celi LA. The ǣinconvenient truthǥ about ai in healthcare.
Npj Digital Med 2019;2(1):1–3.
We wish to confirm that there are no known conflicts of inter- [17] Zhao S, Lin Q, Ran J, Musa SS, Yang G, Wang W, et al. Preliminary estimation
est associated with this publication and there has been no signifi- of the basic reproduction number of novel coronavirus (2019-ncov) in china,
cant financial support for this work that could have influenced its from 2019 to 2020: adata-driven analysis in the early phase of the outbreak.
Int J Infect Diseas 2020;92:214–17.
outcome. [18] Shim E, Tariq A, Choi W, Lee Y, Chowell G. Transmission potential and severity
2. Funding Funding was received for this work. of Covid-19 in south korea. Int J Infect Diseas 2020.
All of the sources of funding for the work described in this [19] Liestøl K, Andersen PK. Updating of covariates and choice of time origin in
survival analysis: problems with vaguely defined disease states. Stat Med
publication are acknowledged below: This research is funded by
2002;21(23):3701–14.
Saskatchewan Center for Patient Oriented research (SCPOR) [20] Krätschmer V. Strong consistency of least-squares estimation in linear regres-
We confirm that we have given due consideration to the pro- sion models with vague concepts. J Multivar Anal 2006;97(3):633–54.
[21] Bandyopadhyay SK, Dutta S. Machine learning approach for confirmation of
tection of intellectual property associated with this work and that
Covid-19 cases: positive, negative, death and release. medRxiv 2020.
there are no impediments to publication, including the timing of [22] Huang C-J, Chen Y-H, Ma Y, Kuo P-H. Multiple-input deep convolutional neural
publication, with respect to intellectual property. In so doing we network model for covid-19 forecasting in china. medRxiv 2020.
confirm that we have followed the regulations of our institutions [23] Tomar A, Gupta N. Prediction for the spread of Covid-19 in india and effective-
ness of preventive measures. Sci Total Environ 2020:138762.
concerning intellectual property. [24] Pal R, Sekh AA, Kar S, Prasad DK. Neural network based country wise risk pre-
Research Ethics diction of covid-19. arXiv preprint arXiv:20 040 0959 2020.
We further confirm that any aspect of the work covered in this [25] Xu Y, Weaver JB, Healy DM, Lu J. Wavelet transform domain filters:
a spatially selective noise filtration technique. IEEE Trans Image Process
manuscript that has involved human patients has been conducted 1994;3(6):747–58.
with the ethical approval of all relevant bodies and that such ap- [26] Cheung Y-W, Lai KS. Lag order and critical values of the augmented dickey—
provals are acknowledged within the manuscript. fuller test. J Bus Econ Stat 1995;13(3):277–80.
[27] Karim F, Majumdar S, Darabi H. Insights into lstm fully convolutional networks
We confirm that the manuscript has been read and approved for time series classification. IEEE Access 2019;7:67718–25.
by all named authors. [28] de Jong Jd, Claas E, Osterhaus AD, Webster RG, Lim W. A pandemic warning?
We confirm that the order of authors listed in the manuscript Nature 1997;389(6651):554.
has been approved by all named authors.