Downloaded from https://iranpaper.

ir
https://www.tarjomano.com https://www.tarjomano.com

Rabies Outbreak Prediction Using Deep

Learning with Long Short-Term Memory

Abdulrazak Yahya Saleh1(&), Shahrulnizam Anak Medang1,

and Ashraf Osman Ibrahim2,3
1
FSKPM Faculty, Universiti Malaysia Sarawak (UNIMAS),
94300 Kota Samarahan, Sarawak, Malaysia
[email protected]
2
Faculty of Computer Science and Information Technology,
Alzaiem Alazhari University, Khartoum North 13311, Sudan
3
Faculty of Computer Science, Future University, Khartoum, Sudan

Abstract. The purpose of this article is to evaluate the Long Short-Term

Memory (LSTM) model performance for rabies outbreak prediction (ROP).
Successful forecasting of the initial epidemic outbreaks can decrease the inci-
dence of the ailment and save lives, but this type of research is costly, and an
erroneous result can trigger false alarms, and the trustworthiness of the warning
system will be at stake. As such, biosurveillance system developers are looking
for highly sensitive outbreak prediction algorithms that will minimise the
number of false alarms. Using the epidemiological data such as those of rabies to
forecast novel and vital directions is a significant issue of public health, and it
involves the collective attention of the machine learning (ML) communities. In
this study, the data are obtained from HealthData.com and utilised for the
performance evaluation of the LSTM algorithm. The algorithm performance is
evaluated based on Root Mean Square Error (RMSE) and Accuracy, and
compared with that of the traditional algorithm– the Autoregressive integrated
moving average (ARIMA) model. The results from this research prove that a
deep learning LSTM network can predict the disease prevalence, using the
rabies datasets, with a good accuracy. The performance of the proposed model is
evaluated by comparing with the ARIMA model. The LSTM model attains the
best result with 97.10% accuracy, while the traditional ARIMA obtains 72.10%.
Moreover, the LSTM model scores the lowest value of RMSE (2.04) compared
with the ARIMA model which scores the highest (3.12). Through this study, it is
obvious that the LSTM prediction model is an effective method for determining
this viral disease, evidenced by a very low RMSE value and a high accuracy
score.

Keywords: Long Short-Term Memory (LSTM)
Autoregressive Integrated

Moving Average (ARIMA)
Machine Learning (ML)
Rabies Outbreak
Prediction (ROP)

© Springer Nature Switzerland AG 2020

F. Saeed et al. (Eds.): IRICT 2019, AISC 1073, pp. 330–340, 2020.
https://doi.org/10.1007/978-3-030-33582-3_32
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Rabies Outbreak Prediction Using Deep Learning with LSTM 331

1 Introduction

Rabies is a well-known zoonotic disease. This disease has been classified under the
Lyssavirus type of the Rhabdoviridae family, which infects warm-blooded creatures
and causes lethal encephalitis [1]. A person bitten by a crazy infected dog has more
than 99% chance of contracting the disease. Although rabies is preventable, the death
toll is still high. It is estimated that 59,000 humans die from rabies every year on the
planet Earth, with sub-Saharan Africa and Asia contributing to over 95% of these
deaths [2]. According to a report of WHO [3], about 45% of the total deaths caused by
rabies happen in all parts of the world. More than 3.3 billion persons are in danger of
the rabies infection and it disproportionally affects the under-resourced and poorest
populaces living in the lower-middle income nations; the infection rate is on the rise in
a few of the countries. Ramos [4] explains that in addition to the undesirable effects on
mortality, rabies has an adverse impact on the financial health of the affected populaces.
Marí [5] says that about 75% of the uncontrollable rising numbers of maladies affecting
humans nowadays originate from the animals, and 60% of these ailments are zoonotic
in nature. Rabies is an exceptionally deadly disease that has been around for about
4000 years; it is caused by the bullet-shaped Lyssavirus that can infect all warm-
blooded creatures with a mortality rate of nearly 100%, after the beginning of clinical
signs or manifestations [6]. This illness presently kills in excess of 69,000 individuals,
the vast majority of which are from Asia and Africa [7].
Epidemic prediction around the globe is an essential public health issue. Infor-
mation about tracking, timely detection and forecasting of epidemics like outbreak time
and peak intensity are vital for efficient health interventions. Wu mentions in [8] that
epidemic control draws much attention and consideration of the government in every
nation and the World Health Organization (WHO). Among the numerous issues are
scourge control, and the epidemic prediction, which gauges the prevalence of the
pandemic in specific locales. As indicated by Fricker [9], Bio-observation is the pro-
cedure of effectively assembling and breaking down information associated with
human wellbeing and diseases; it provides early warning of an outbreak of a disease in
a community and raises the general awareness of the human ailment in the populace.
Outbreak forecasting algorithms carry out the job of computerised biosurveillance,
analysing vast volumes of clinical information progressively and distinguishing the
different phases of an ailment development. In the medicinal services industry, utili-
sation of time-series modelling and forecasting a future flare-up of certain uncontrol-
lable illnesses and infection, which happen cyclically, is exceptionally critical. The
determining of an ailment helps predict the course of an illness, cautions social
insurance specialists about the potential health cost and alerts the relevant authorities to
introduce control measures to avoid disease flare-ups. Zoonotic sicknesses are a group
of uncontrollable diseases normally transmitted from animals to people. There are
numerous methodologies that can be used to predict the spread of Zoonotic Diseases
using Machine Learning techniques [10, 11]. However, there are limitations in con-
ducting this kind of experiment when the actual datasets display irregular trends [12].
Therefore, forecasting different and crucial drifts of epidemiological data such as that of
rabies is a significant public health issue.
 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

332 A. Y. Saleh et al.

One of the powerful methods which can be utilised for epidemic prediction is deep
learning [13]. Deep learning is well known in the field of predicting the probability of
event occurrences; it entails collecting huge amounts of data and uses them to forecast
the outcome of the future [14]. It has been used to solve numerous problems in the
world such as flood prediction and the malware detection [15, 16].
In this study, Deep Learning with Long Short-Term Memory is utilised to predict
the rabies outbreaks by applying the time-series analysis. In this regard historical data
are used to identify and validate the existing data patterns, which are then used to
predict what will happen in the future. By utilising Deep Learning to predict rabies
outbreaks, the local governments can put in place the necessary preventive measures
and raise the awareness of the people about the impending danger so that they can take
precautions and appropriate actions to protect themselves. As the saying goes, fore-
warned is forearmed, and the harm would be greatly reduced.
The rest of this article is organised in the following manner: Sect. 2 presents an
explanation of the materials and methods, Long Short-Term Memory (LSTM), and the
general framework of the methodology; while Sect. 3 analyses the findings; and finally,
Sect. 4 delivers the conclusion and future work.

2 Materials and Methods

Based on the latest development in Deep Learning, a powerful new pattern of machine
learning has emerged [17]. Deep Learning refers to a subset of machine learning that
imitates the workings of the human brain in processing data and creating patterns for
use in decision-making, by using ‘deep’ artificial neural systems, for example, recurrent
neural networks (RNN), convolution neural networks (CNN) and deep neural networks
(DNN). The area of Data Science has extended its boundaries into the sphere of
healthcare and medicine. Although epidemiology and biostatistics have been in exis-
tence for several years, the current industrial flourishing in Deep Learning (DL) and
Artificial Intelligence is a driving factor for innovative research in these fields. This
study focuses on predicting the rabies outbreak and it is measured by the model
performance (Accuracy). The proposed DL model performance is evaluated by com-
paring with that of ARIMA. The datasets of rabies are obtained from HealthData.com;
it is an open-source data management platform from the United States, with 1000
samples, dataset characteristics (Univariate) and Features (2). The software used for
ROP in the current research is Python because it is very flexible and popular among
other programming languages.

2.1 Long Short-Term Memory (LSTM)

LSTM is a type of RNN; it is created to prevent the output of neural network for a
certain input from exploding or decaying as it rounds over the feedback loops [18]. The
pattern recognition of the Recurrent Networks is considered superior to those of other
neural networks because the former allows the feedback of memory cells of the
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Rabies Outbreak Prediction Using Deep Learning with LSTM 333

previous input, which are essential for settling frequent learning errands. LSTM net-
works deliver improved findings associated with other RNN models by mitigating the
problem of vanishing gradient [19].

Fig. 1. Detailed diagram of a Long Short-Term Memory block

Figure 1 demonstrates the detailed diagram of the LSTM block utilised in the RNN
hidden layers. The LSTM unit consists of four input weights and four recurrent
weights. The LSTM is appropriate for forecasting the data of time series if there is a
period venture with an arbitrary size [20]. It seems that the execution procedure of the
epidemic forecasting models may be improved by employing LSTM and the data
gathered for the research. An essential preferred standpoint of RNN is that the relevant
data are accessible if they have been mapped into the IO successions. In any case, the
gradient issue might arise as a result of the impact of a certain input on the hidden layer,
which can be enlarged or reduced meaningfully throughout the circular connection. As
certain input sources are overwritten, the affectability of the principal input diminishes
after some time. Consequently, the system is “forgotten”. The three gates are consid-
ered as non-linear additional units which control the cell activation. The forget gate
duplicates the cell’s past condition, whereas the input and output gates increase the
cell’s IO. The door enactment work f is a strategic sigmoid. The IO activation functions
g and h of the cell, for the most part, utilise hyperbolic tangents or strategic sigmoid. Be
that as it may, at times, h utilises the identity function. For whatever length of time that
the forget gate is open and the input door is shut, the memory cell keeps on recalling
the first input. Along these lines, LSTM is an algorithm that settles an issue in the
conventional RNNs [21]. The conditions for forgetting, putting away, restoring, and
outputting data in the cell appear as follows, separately:
 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

334 A. Y. Saleh et al.

f t ¼ rðwf :½ht1 ; xt  þ bf Þ ð1Þ

it ¼ rðwi :½ht1 ; xt  þ bi Þ ð2Þ

ct ¼ tanhðwc
½ht1 ; xt  þ bc Þ ð3Þ

ct ¼ f t xct1 þ it xct ð4Þ

ot ¼ rðwo :½ht1 ; xt  þ bo Þ ð5Þ

ht ¼ ot x tanhðct Þ ð6Þ

After the data (xt ) are given to the LSTM cell in Eq. 1, function f t controls the data
to be forgotten in the cell layer. In Eqs. 2 and 3, data to be again kept in the cell layer
are formed in it and ct . In Eq. 4; the cell layer ct is changed utilising ft, it , and ct . In
Eq. 5, the cell layer’s information is utilised and ht is the output. In Eq. 6, the cell state
gets a value between −1 and 1 over the tanh function. The values of ct and ht are saved
for the following repetition of LSTM.

Stage 1 Stage 2 Stage 3

Developing the
LSTM model

Evaluation
and comparing
the LSTM
Data model with
preparation ARIMA model.

Testing and Training

the dataset

Fig. 2. Detailed diagram of an LSTM block

Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Rabies Outbreak Prediction Using Deep Learning with LSTM 335

Figure 2 shows the general framework of the methodology. The first stage is
obtaining the data samples for the experiment. The second stage is the description of
the LSTM model, followed by testing and training the datasets. Finally, the last stage is
evaluation and comparing the model with ARIMA. Data pre-processing is a data
mining procedure which contains shifting CRUDE information to a reasonable for-
mation. Genuine data are normally deficient, contradictory, and at the same time poor
in some practices or points and might have many errors. Information preparation is an
established approach for solving such problems. With information preparation,
CRUDE information is ready for extra treatment. Information preparation is used in
database-driven applications like client relationship and the executives, as well as
standard-based applications (such as neural systems). Rouse in [22] states that data
preparation is the way of collecting, joining, forming and categorising data so that they
may be analysed as a key feature of data visualisation, examination and machine
learning applications. One of the basic factors of the data preparation is that data
processed for examination must be exact and expectable, so the aftereffects of BI and
analytics applications will be considerable. Poor quality of data is a common problem,
which may include missing values, mistakenness or other errors. Data collections
should be put away in discrete samples or databases that have distinctive configurations
to accommodate them. The ways of revising mistakes, providing confirmation and
joining informational indexes are a main part of the data preparation process. Data
preparation is, to a great extent, a robotised assignment, since it can take long stretches
of work by the IT staff members or information experts to physically address each field
in each document that will be utilised in an investigation, especially in a big data
application. Machine learning algorithms can speed things up by looking at information
fields, and naturally filling in clear qualities or renaming certain fields to guarantee
consistency when information documents are being joined. The data for this research is
collected from HealthData.com. It is an open-source data management platform from
the United States with 1000 samples. The dataset characteristics are univariate and have
two features. For an accurate performance comparison of diverse classifiers, it is
essential to validate the data, and select the suitable performance metrics to be utilised.
The classifiers from the data approximate the true features to build the models. The
obtained data need to be split into training and testing datasets, in order to evaluate the
constructed models. The training data are utilised to build the model, and the test data
to assess it. It is assumed that the test data labels are recognised so that they can be put
on the metrics of performance, which are accuracy and RMSE in this study.

3 Results

This section contains the results of the statistical analysis of LSTM for the rabies
datasets. In this study, the LSTM model has experimented with a different number of
epochs and different batch sizes. The findings of all datasets included are categorised
according to Accuracy and RMSE.
 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

336 A. Y. Saleh et al.

3.1 Accuracy
This section explains the results and the LSTM model performance based on accuracy.
In this research, the explained variance score is utilised to test the accuracy and the
model performance. The described variance regression score function with the
superlative probable score is 1.0; lower values indicate worse scenarios.

Table 1. Different epoch numbers with different batch size numbers (accuracy)
Number of epochs Batch size Accuracy (%)
1 32 64 128
50 ✓ 94.47
100 ✓ 97.3
120 ✓ 88
150 ✓ 84.3
200 ✓ 85.1

Each experiment shows a different outcome and we select the best result based on
the accuracy as shown in Table 1. The best accuracy from these experiments is 97.3%,
utilising 100 epochs and 1 batch size. Meanwhile, the lowest accuracy is 84.3%,
utilising 150 epochs and 32 batch size.

3.2 Root Mean Square Error (RMSE)

The RMSE metric is utilised to show the variance between the actual and forecast
values. A lower RMSE value shows a lesser variance between the actual and forecast
results, which indicates an excellent forecasting performance.

Table 2. Different epochs number with different batch size number (RMSE (testing))
Number of epochs Batch size RMSE (Testing)
1 32 64 128
50 ✓ 2.76
100 ✓ 2.04
120 ✓ 4.82
150 ✓ 4.61
200 ✓ 4.58

Table 2 above presents the results of RMSE of the LSTM model. Lower values of
RMSE indicate a better fit. Moreover, Table 2 shows that 100 epochs with batch size 1
produce the lowest RMSE value of 2.04; and 120 epochs with batch size 32 produce
the highest RMSE value of 4.82.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Rabies Outbreak Prediction Using Deep Learning with LSTM 337

Table 3 below shows a summary of the best results of the LSTM model. LSTM has
a small RMSE value of (Testing) 2.04 and the highest accuracy of 97.30%.

Table 3. Summary of LSTM best results

Model RMSE (test) Accuracy (%)
LSTM 2.04 97.30

3.3 Comparative Study and Discussion

This study focuses on predicting the rabies outbreak and it is measured by the model
performance (Accuracy). The performance of the proposed model is evaluated by
comparing with that of ARIMA. The deep learning model (LSTM) displays extraor-
dinary performance in comparison with the traditional method (ARIMA). The ARIMA
model utilised for the study was noticed to be perfect when the incidences of the
infectious diseases were steady and have zero growing or reducing drifts. But actual
data may have drifts and abnormality. Consequently, deep learning might be a better
analytical technique when analysing certain data and forecasting the upcoming con-
ditions. Based on the findings of the earlier analyses, the model of deep learning adapts
to growing and reducing drifts adequately well. Additionally, the LSTM model was
noticed to be susceptible to reducing drifts and growing drifts, respectively. Figure 3
shows a comparative analysis of the LSTM model and ARIMA model respectively.
Findings from this study evidently prove that the LSTM model is consistent in pre-
dicting the rabies outbreak. The LSTM model attains the highest accuracy of 97.30%
compared with 72.10% by the ARIMA model.

Fig. 3. Comparison of the accuracies of LSTM and ARIMA

 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

338 A. Y. Saleh et al.

Fig. 4. Comparison of RMSE values of LSTM and ARIMA

In addition, Fig. 4 shows the comparative results of RMSE values between the
LSTM and ARIMA models. The results show that LSTM has a low RMSE value of
2.04. Meanwhile, ARIMA obtains a high RMSE value of 3.118—higher than that of
LSTM.

4 Conclusion

This study focuses on assessing the LSTM performance for forecasting the rabies
outbreak and the results are compared with those of the ARIMA model. The results
achieved by this study show that a deep learning LSTM network can predict an
outbreak using a rabies dataset with a better accuracy. The LSTM model acquires the
best accuracy results of 97.30%; meanwhile, the ARIMA model with a traditional
network obtains 72.10%. As for the comparison of RMSE values, the LSTM model
scores the lowest value of 2.04, and the ARIMA model scores the highest value of
3.12. From these results, the LSTM model is superior for forecasting the outbreak of
this viral disease. The success of the LSTM model warrants it to be used for forecasting
the outbreaks of other diseases like flu, H1N1, and Ebola or related problems. In future
work, deeper architectures will be applied in order to achieve a better performance.
More data are also required to improve the prediction performance. Finally, different
machine learning algorithms should be utilised for the purpose of making more robust
comparisons.
Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

Rabies Outbreak Prediction Using Deep Learning with LSTM 339

References
1. Hemachudha, T., Ugolini, G., Wacharapluesadee, S., Sungkarat, W., Shuangshoti, S.,
Laothamatas, J.: Human rabies: neuropathogenesis, diagnosis, and management. Lancet
Neurol. 12(5), 498–513 (2013)
2. Hampson, K., Coudeville, L., Lembo, T., Sambo, M., Kieffer, A., Attlan, M., et al.:
Estimating the global burden of endemic canine rabies. PLoS Negl. Trop. Dis. 9(4),
e0003709 (2015). https://doi.org/10.1371/journal.pntd.0003709
3. Kole, A.K., Roy, R., Kole, D.C.: Human rabies in India: a problem needing more attention
(2014)
4. Ramos, J.M., Melendez, N., Reyes, F., Gudiso, G., Biru, D., Fano, G., et al.: Epidemiology
of animal bites and other potential rabies exposures and anti-rabies vaccine utilization in a
rural area in Southern Ethiopia (2015)
5. Bueno-Marí, R., Almeida, A.P.G., Navarro, J.C.: Emerging zoonoses: eco-epidemiology,
involved mechanisms and public health implications. Front. Publ. Health. 3, 157 (2015)
6. Sparkes, J., Fleming, P.J.S., Ballard, G., Scott-Orr, H., Durr, S., Ward, M.P.: Canine rabies
in Australia: a review of preparedness and research needs. Zoonoses Publ. Health 62, 237
(2014)
7. Mähl, P., Cliquet, F., Guiot, A.L., Niin, E., Fournials, E., Saint-Jean, N., Aubert, M.,
Rupprecht, C.E., Gueguen, S.: Twenty-year experience of the oral rabies vaccine SAG2 in
wildlife: a global review. Vet. Res. 45(1), 77 (2014)
8. Wu, Y., Yang, Y., Nishiura, H., Saitoh, M.: Deep learning for epidemiological predictions.
In: SIGIR, (2018). Ann. Agricul. Environ. Med. AAEM 22(1), 76–79. https://doi.org/10.
5604/12321966.1141372
9. Fricker, R.: Some methodological issues in biosurveillance. Stat. Med. 30, 403–415 (2011)
10. H, Bamaiyi: 2015 outbreak of canine rabies in malaysia: review, analysis and perspectives.
J. Vet. Adv. 5(12), 1181 (2015). https://doi.org/10.5455/jva.19691231040000
11. He, J., Luo, L., Jin, R.G., Li, J.M.: The application of ARIMA in forecasting the cases of
rabies in China different human groups. Zhonghua lao dong wei sheng zhi ye bing za
zhi = Zhonghua laodong weisheng zhiyebing zazhi = Chin. J. Ind. Hygiene Occup. Dis. 36
(7), 512–515 (2018)
12. Chae, S., Kwon, S., Lee, D.: Predicting infectious disease using deep learning and big data.
Int. J. Environ. Res. Public Health 15(8), 1596 (2018). https://www.ncbi.nlm.nih.gov/
pubmed/30060525
13. Wu, Y., Yang, Y., Nishiura, H., Saitoh, M.: Deep learning for epidemiological predictions.
In: SIGIR (2018)
14. Zhang, Q., Yang, L.T., Chen, Z., Li, P.: A survey on deep learning for big data. Inform.
Fusion 42, 146–157 (2018)
15. Saleh, A.Y., Tei, R.: Flood prediction using seasonal autoregressive integrated moving
average (SARIMA) model. Int. J. Innov. Technol. Explor. Eng. 8(8), 1037–1042 (2019)
16. Saleh, A.Y., Francis, C.: A deep learning approach to Malware detection in android
platform. Int. J. Innov. Technol. Explor. Eng. 8(8), 1043–1048 (2019)
17. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning; Nature Publishing Group, a division of
Macmillan Publishers Limited, 28 May 2015. https://doi.org/10.1038/nature14539
18. Staudemeyer, R.C.: Evaluating performance of long short-term memory recurrent neural
networks on intrusion detection data, October 2013
19. Brownlee, J.: Time series prediction with lstm recurrent neural networks in python with
keras, p. 18 (2016). machinelearningmastery.com
 Downloaded from https://iranpaper.ir
https://www.tarjomano.com https://www.tarjomano.com

340 A. Y. Saleh et al.

20. Bao, W., Yue, J., Rao, Y.: A deep learning framework for financial time series using stacked
autoencoders and long-short term memory. PLoS One 12, e0180944 (2017). https://doi.org/
10.1371/journal.pone.0180944
21. Alex, G.: Supervised Sequence Labelling with Recurrent Neural Networks; Studies in
Computational Intelligence. Springer, Berlin (2012)
22. Rouse, M.: What is data preparation? - Definition from WhatIs.com, January. 2018. https://
searchbusinessanalytics.techtarget.com/definition/data-preparation