2023 International Conference on Next-Generation Computing, IoT and Machine Learning (NCIM)

16-17 June 2023, Gazipur-1707, Bangladesh

A Study on Future Lockdown Predictions using

ANN
Shuvra Smaran Das Anirban Saha Anik Md. Muzakker Hossain
Department of Computer Science Department of Computer Science Department of Computer Science
American International University- American International University- American International University-
Bangladesh (AIUB) Bangladesh (AIUB) Bangladesh (AIUB)
2023 International Conference on Next-Generation Computing, IoT and Machine Learning (NCIM) | 979-8-3503-1600-1/23/$31.00 ©2023 IEEE | DOI: 10.1109/NCIM59001.2023.10212686

Dhaka, Bangladesh Dhaka, Bangladesh Dhaka, Bangladesh

shuvradas59@[Link] anirbansaha002@[Link] 18-37801-2@[Link]

Md. Kishor Morol Fariha Jahan Md. Abdullah - Al - Jubair

Department of Computer Science Department of Computer Science and Department of Computer Science
American International University- Engineering American International University-
Bangladesh (AIUB) Daffodil International University Bangladesh (AIUB)
Dhaka, Bangladesh Dhaka, Bangladesh Dhaka, Bangladesh
kishor@[Link] [Link]@[Link] abdullah@[Link]

Abstract—More than three years have passed since the first future. During a pandemic, it is critical to determine when a
detection of COVID-19, and our generation had no clue as to lockdown is necessary for the nation and its citizens.
what was coming or how dangerous the disease was.
Unfortunately, this is not the end of such diseases, and in order A lockdown is a strategy that mandates individuals or
to minimize losses, we may have to deal with many more communities to remain in a designated location due to
lockdowns over new diseases. With the data from the recent particular risks to their health or others' well-being if they are
pandemic, we investigated how to anticipate the potential permitted to travel and interact freely. Lockdowns that affect
lockdown date in the future. We took into account the first a geographic area rather than specific sites are often known as
detection, the first death by COVID-19, the WHO emergency "stay-at-home" or "shelter-in-place" orders. The
declaration, the total GDP of the various nations, the GDP implementation of a lockdown and adherence to social
growth, the population density, etc. This paperwork predicts the distancing measures have proven to be effective in limiting the
day passed after the first COVID-19 positive detection inside the virus's transmission. Depending on the region, some
country when the country's government declared a lockdown lockdowns may have complete movement restrictions, while
(FDL) and the day when the WHO announced COVID-19 as an others may be subject to time-based limitations. Only essential
emergency of international concern (WEL). This research uses businesses are usually allowed to operate during these
a variety of machine learning models, including a modified restrictions.
Artificial Neural Network (ANN), for the prediction. This study
presents the successful tuning and training of an ANN model A pandemic, such as COVID-19, initially found in China
using the lockdown dataset, ultimately achieving an impressive in late December 2019, is causing one of the world's most
accuracy rate of 96% for FDL (First Detection Lockdown) and dreadful public health disasters. It has expanded to more than
98% for WEL (World Emergency Lockdown), which is more 200 nations in less than eight months. It is transmitted between
accurate than any other applied traditional and/or tree-based people through in-person encounters or interactions with an
machine-learning models. infected person via respiratory organs, saliva, and droplets
when coughing. According to current evidence, these are
Keywords—Machine learning, Lockdown, Neural Network,
ejected from infected ones by sternutation, singing, or talking.
Pandemic.
Dry cough, exhaustion, and fever are all common symptoms.
I. INTRODUCTION People over the age of sixty and those with medical issues
such as high blood pressure, heart People with heart disease,
The pandemic is having a terrible effect on the economy diabetes, obesity, or cancer are more likely to develop a
and society. Tens of millions of people are at risk of falling serious illness. When a person is infected with the
into extreme poverty. The global community must prepare for coronavirus, symptoms usually appear after five to six days
this type of pandemic [18]. More than that, the polar regions but can take up to fourteen days. Covid-19 was designated a
are warming two to four times faster than the rest of the planet pandemic by the World Health Organization (WHO) on
as global warming intensifies. This is forcing an important soil March 11, and governments were asked to take steps to stop it
layer known as permafrost to melt for the first time, which is from spreading[11,12].
melting everything that has been frozen in this icy prison for
hundreds of thousands of years. Many of the microorganisms The COVID-19 pandemic resulted in the closure of
that are currently awakening are descendants of bacteria that schools, universities, and colleges, affecting a significant
caused prior pandemics that have since died out. Their portion of the world's student population. Many countries
extended slumber in subzero temperatures has preserved their enforced lockdowns in 2020 and 2021 to control the spread of
response systems and kept them from decomposing. Extreme the virus, starting with China's Hubei region and later in Italy.
cold also prevents deterioration. When these microbes wake India's lockdown was the largest in terms of population size,
up and find themselves floating in liquid instead of solid ice, with over 1.3 billion people staying at home. The longest
they are called Zombie Bacteria. As a result, the world may continuous lockdown was in Buenos Aires, Argentina, lasting
once again confront a pandemic as severe as Corona in the 234 days. Melbourne, Australia, holds the record for spending

979-8-3503-1600-1/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Dhaka Univ of Engineering & Technology Gazipur. Downloaded on June 24,2025 at 17:04:37 UTC from IEEE Xplore. Restrictions apply.
 the most cumulative days in lockdown as of October 2021, linear regression. This paper does an excellent job of
totaling 267 days multiple times. While experts generally describing graphs and datasets. In [13], the impact of the
supported lockdown restrictions to curb the spread of the lockdown on air quality in Uttarakhand, India, during the
virus, these measures have had significant health, social, and covid-19 outbreak has been studied. The presence of SO2,
economic consequences. They have been met with protests in NO2, and particulate matter (P.M.10) shows the ambient air
some areas [12,17]. It is essential to note that the impacts of quality. With two distinct techniques of measuring distance in
the lockdown before and after its implementation will differ. MATLAB, the clustering potential of the K-means clustering
The lockdown's primary aim was to prevent viral methodology is studied as this algorithm is modest and fast
transmission, resulting in a lower rate of dissemination than and gives a good performance. The Euclidean distance is the
before its implementation. first method, whereas the cosine distance is the second. In this
study, mixed air quality data from two years were analyzed,
The primary objective of this research is to estimate the namely 2019 (normal period) and 2020 (lockdown period of
lockdown duration necessary to prevent this type of pandemic. Covid-19). Using the K-means clustering algorithm in
A dataset was required in order to remedy the situation. A MATLAB, they were categorized into two clusters, namely
database has been constructed, which contains lockdown polluted air and clean air. The findings of two distinct ways of
starting times, country GDPs, population densities, and other
determining distance have been compared. The Euclidean
pertinent information. The lockdown start date information distance strategy of K-means clustering produces a
was acquired from the COVID pandemic because there were substantially better clustering result for the given data set than
insufficient datasets established for earlier pandemics. Using the cosine distance approach. Narayana Darapaneni et al. [12]
WHO and World Bank statistics, this dataset was generated. conducted another study utilizing supervised learning
Different Machine learning algorithms were used in this approaches on the impact of lockdown and post-lockdown on
research. The algorithms improve their performance Covid19 cases across India. The authors provided a detailed
adaptively as the number of learning examples grows. Various description of the data analysis. Here, data is gathered through
machine learning algorithms were used to predict when a crowdsourcing, and the findings are represented through
government would need to initiate a lockdown and when it various graphs using formulas. According to the
would be beneficial for the country's citizens. However, we investigation's findings, initiatives such as public awareness
have modified the Artificial Neural Network and trained it to of social distancing and using an obligatory face mask can be
see how it performs with other traditional machine learning helpful. Governments may gradually open up economic
models. operations while enforcing basic social distance rules,
allowing economies to recover while restricting the spread of
The objectives of this paper are- the Covid-19 virus.
• The study of how and when different countries and R. M. NASSR et al. [14] conducted second research on
organizations decided to announce the lockdown and educational challenges in Malaysia during the COVID-19
create a dataset for each nation with different features shutdown. This research examined the challenges faced by
such as lockdown after WHO emergencies, age students at a government-affiliated institution. This institution
percentage, total GDP, lockdown after the first death, is one of the biggest in Malaysia, with over ten campuses
density, lockdown after first detection, total lengths, located around the nation. This investigation produced 284
and first lockdown length. valid replies. The results indicate that respondents were
• Apply machine learning models to predict the impact physically, environmentally, and psychologically unprepared
after the WHO emergency lockdown announcement for this condition, with some gender, age, and state of
and after the first COVID-19 detection in a nation. residency differences in responses. Concern about the
consequences of the lockdown on the responders'
• The study of the impactful features helps predict our performance. According to the results of this research,
outcome and the possible future of lockdown for a students who have no intention to miss class experience
nation and the socio-economic situation before and considerable challenges if they abruptly convert to an online-
after the lockdown. only alternative. The results also indicate that changing the
current blended learning model, which incorporates online
II. LITERATURE REVIEW learning as a support mechanism for face-to-face learning, has
Some studies have been conducted to forecast the proved challenging, particularly for students who are
lockdown. Prathyusha K et al. [11] presented coronavirus case unprepared. Those semesters that occurred during the
analysis according to lockdown. Before, after, and during COVID-19 epidemic are likely to fall short of worldwide
lockdowns, researchers looked at the COVID-19 patients' data standards and benchmarks for higher education, thus, they
for confirmed cases, recovered cases, and deaths. Most of the should be examined. This is the first time that students and
datasets are from Kaggle, and the researchers created a instructors have experienced this unique condition and
lockdown dataset from March 1, 2020, to June 30, 2020, as challenges (academic, communication, technological, etc.)
detailed in this publication. The data is analyzed based on the that have negatively affected the quality of the educational
daily growth rate of confirmed cases, the rate of recovery, and process. The curriculum consists of lectures, laboratory
the rate of demise. Machine learning is employed in this case exercises, and exams. Universities should prepare for life
because it extracts more useful statistics from vast amounts of beyond COVID-19 since education has undergone substantial
data using an algorithm model for explicit problems. changes that may impact its future.
Regression techniques such as Linear Regression and Many countries' lockdown states have been described in
Polynomial Regression from Machine Learning (ML) have research. Researchers used data from Worldometer to create a
been used to anticipate the future spread of this virus in India. graph. This study aimed to assess the effectiveness of
As a result, polynomial regression shows better results than lockdown measures globally, taking into account the mix of

Authorized licensed use limited to: Dhaka Univ of Engineering & Technology Gazipur. Downloaded on June 24,2025 at 17:04:37 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1 Dataset Features Correlation

policies, healthcare-directed spending, and outcomes in terms developed a new dataset, namely Lockdown Prediction
of COVID-19-related mortality and incidence. The first Dataset. These data were gathered from the official websites
judgments on the success of the lockdown may be formed now of the World Bank and the WHO at
that most countries have completed the first month after the [Link] and [Link]
outbreak began. Since the pandemic began, it is apparent that respectively. The majority of the data in our dataset—which
nations in East Asia have fared far better than those in Europe has 16 features—was gathered from these two websites. To
and North America. This good conclusion was aided by the collect data particular to each country, we made use of a
fast implementation of quarantine measures and population- number of other popular websites. The complete dataset can
wide testing [15]. This study aims to demonstrate, using real- be accessed at [Link]
time incidence data, that the Lebanese government's country Lockdown-Prediction. Not all of the features, nevertheless,
lockdown will have a key and major influence on decreasing were employed in this research. We’ve used 11 features from
the spread of the virus and, eventually, containing it. The the dataset. The statistics of these features can be found in
findings of an investigation of COVID-19 instances in Table I.
Lebanon before and after the lockdown measures are
presented in this study. The results reveal that daily instances B. Dataset Preprocessing
reduced after the lockdown, indicating that the lockdown Each country in Table I has a different number of sample
tactics have successfully limited the sickness thus far [16]. data points. Therefore, we must transform it such that each
Researchers use a mathematical framework to get insight into sample has the same amount of data. We must calculate this
the benefits of staying at home and the impact of the lockdown by averaging the initial lockdown duration (days), total
on the virus's propagation. To put it another way, the purpose lockdown length (days), total GDP, percentage of people over
is to track the slowdown of the pandemic, which is most likely 65, and GDP growth characteristics. Additionally, there are
due to lockdown effectiveness. Authors [17] utilized R to several features with string values where label encoding must
build graphs for different nations using data from be used. There are four designations for a country's economic
[Link] category: "High Income," "Upper Income," "Lower-Middle
Income," and "Lower Income." To help our machine learning
III. METHODOLOGY algorithms better determine how those labels should be
applied, we have used the label encoder to transform these text
A. Dataset Description
values into a numerical value in this case. Last but not least,
We required a customized dataset that matched our we utilized Standard Scaler to scale our dataset using the
requirements in order to carry out the study. Therefore, Standard Normal Distribution (SND). As a result, it adjusts
integrating data from 144 different countries and key features the data to unit variance and puts the mean at 0. Fig.1 displays
like lockdown duration, population density, country economic a correlation among the features of the dataset. Our dataset
type, GDP growth, and the percentage of persons over 60, we consists of a variety of numerical data types, including days,

TABLE I. STATISTICS OF FEATURES.

Count Mean Standard Deviation Minimum Value Maximum Value
First lockdown length (days) 95 49.16842 35.57601 6 234
Total lengths (days) 95 79.47368 72.09536 7 405
Lockdown After First Detection 144 27.88194 76.19128 -251 443
Density (/km², 2022) 144 407.0208 1728.449 2 18235
Lockdown after First Death 144 -17.4375 107.4128 -739 407
GDP Growth 142 2.633099 4.230566 -35 9.5
Total GDP (millions $) 143 595642 2228472 852 21372572
% of age > 65 132 10.35606 6.660964 1 28
Lockdown after WHO Emergency 144 61.79167 63.90037 -7 455

Authorized licensed use limited to: Dhaka Univ of Engineering & Technology Gazipur. Downloaded on June 24,2025 at 17:04:37 UTC from IEEE Xplore. Restrictions apply.
 money, category data, etc. We have utilized the SND to min
1
‖𝑋𝑤 − 𝑦‖22 + 𝛼𝜌‖𝑤‖1 +
𝛼(1 − 𝜌)
‖𝑤‖22
redistribute our dataset from a single location in order to 𝑤 2𝑛𝑠𝑎𝑚𝑝𝑙𝑒𝑠 2
improve the performance of the machine learning algorithms.
5) Stochastic Gradient Descent
C. Description of Applied Models Gradient Descent is a fairly general optimization approach
1) Linear Regression that may locate the best answers to a variety of issues.
Gradient Descent's main principle is the iterative adjustment
Linear regression analysis is a statistical method used to
of parameters to minimize a cost function. Gradient Descent
examine the relationship between two variables by predicting
measures the gradient of the error function at the local level
the future value of one based on the current value of the other.
with respect to the parameters vector, and it moves downward.
The "dependent variable" represents the variable of interest,
The lowest is attained when the gradient is zero [8].
while the "independent variable" is the basis for predicting its
future value. Linear regression is a widely used technique in Instead, Stochastic Gradient Descent (SGD) [8] updates
various fields, including economics, social sciences, and parameters for each training example x(i) and label y(i) :
engineering [1].
𝜃 = 𝜃 − 𝜂. Δ𝜃 𝐽(𝜃; 𝑥 (𝑖) ; 𝑦 (𝑖) )
The equation for a linear regression line can be expressed
as Y = a + bX, where Y is the dependent variable, and X is the 6) Polynomial Regression
independent variable. The slope of the line represents the The employment of linear models that have been trained
change in Y per unit change in X, while the intercept indicates on nonlinear data functions is a typical trend in machine
the value of Y when X is equal to zero. learning. This strategy enables linear algorithms to fit a
considerably larger variety of data while maintaining their
2) Ridge Regression typically quick performance.
Ridge regression is a statistical technique used to address
For instance, by creating polynomial features from the
the issue of multi-collinearity in linear regression. It achieves
coefficients, a straightforward linear regression may be
this by adding a penalty term to the least-squares cost function,
improved. For two-dimensional data, you may have a model
which shrinks the regression coefficients toward zero while
that looks like this in the case of ordinary linear regression:
maintaining their relationship with the dependent variable [5].
The regularization parameter α controls the amount of 𝑦̂(𝑤, 𝑥) = 𝑤0 + 𝑤1 𝑥1 + 𝑤2 𝑥2
shrinkage applied to the coefficients, with higher values Instead of fitting the data with a plane, we may mix the
resulting in greater regularization. The ridge regression characteristics in second-order polynomials to create the
equation is given by: following model:
𝛽̂ridge = 𝑎𝑟𝑔𝑚𝑖𝑛 ‖𝑌 − 𝑋𝛽‖22 + 𝛼‖𝛽‖22 𝑦̂(𝑤, 𝑥) = 𝑤0 + 𝑤1 𝑥1 + 𝑤2 𝑥2 + 𝑤3 𝑥1 𝑥2 + 𝑤4 𝑥12 + 𝑤5 𝑥22
Where 𝛽̂ridge is the vector of ridge regression coefficients, 𝑧 = 𝑥1 + 𝑥2 + 𝑥1 𝑥2 + 𝑥12 + 𝑥22
y is the dependent variable, X is the matrix of independent 𝑦̂(𝑤, 𝑥) = 𝑤0 + 𝑤1 𝑧1 + 𝑤2 𝑧2 + 𝑤3 𝑧3 + 𝑤4 𝑧4 + 𝑤5 𝑧5
variables, β is the vector of ordinary least squares coefficients,
and ‖. ‖22 represents the L2 norm of the vector. We can see that the generated polynomial regression
belongs to the same class of linear models as the ones we
Overall, ridge regression is a powerful technique that can previously discussed (the model is linear in w) and can be
significantly improve the predictive performance of linear solved using the same methods. The model is flexible enough
regression models, particularly when dealing with high- to fit a much wider variety of data by taking into account linear
dimensional data with collinear predictors [19]. fits in a higher-dimensional space constructed with these basis
3) Lasso Regression functions [10].
A linear model for sparse coefficient estimation. It is 7) Regression Evaluation Metrics
mathematically made up of a linear model that was trained The average absolute value of the mistakes is known as the
using the l1 prior as a regularizer. The goal is to reduce the Mean Absolute Error (MAE) [3]. Simply put, MAE is the
following function: average error, making it the most intuitive measure to employ.
1 𝑛
min ‖𝑋𝑤 − 𝑦‖22 + 𝛼‖𝑤‖1 1
𝑤 2𝑛𝑠𝑎𝑚𝑝𝑙𝑒𝑠 ∑|𝑦𝑖 − 𝑦̂𝑖 |
𝑛
𝑖=1
The lasso estimate thus minimizes the least-squares
penalty with 𝛼‖𝑤‖1 added, where α is a constant and ‖𝑤‖1 is The average of all the squared mistakes is known as the
the ℓ1−norm of the parameter vector. Mean Squared Error (MSE) [4]. Unlike MAE, MSE
"punishes" greater errors, which is generally helpful in
4) Elastic Net Regression practice.
A linear regression model that was regularized using the 𝑛
L1 and L2 priors. With this combination, it is possible to train 1
∑(𝑦𝑖 − 𝑦̂𝑖 )2
a sparse model, similar to lasso, in which only a small number 𝑛
𝑖=1
of weights are non-zero while preserving Ridge's
regularization qualities. When there are several characteristics More popular than MSE due to its interpretability in "y"
that are associated with one another, an elastic net is helpful. units, the Root Mean Squared Error (RMSE) [2] is calculated
While elastic-net [6] is more likely to choose both, the lasso is by taking the square root of the mean of the squared errors.
more likely to choose one at random. Elastic-Net may inherit 𝑛
part of Ridge's stability under rotation as a result of the trade- 1
√ ∑(𝑦𝑖 − 𝑦̂𝑖 )2
off between Lasso and Ridge, which is a useful benefit. In this 𝑛
𝑖=1
situation, the desired outcome is to reduce.

Authorized licensed use limited to: Dhaka Univ of Engineering & Technology Gazipur. Downloaded on June 24,2025 at 17:04:37 UTC from IEEE Xplore. Restrictions apply.
 Fig. 2 Overview of the modified Artificial Neural Network.

8) Modified Artificial Neural Network

An artificial neural network [7] is a type of artificial
intelligence that takes its cues from the brain and other
biological systems. Artificial neural networks are often
implemented in computer code, and their design is inspired by Fig. 4 Impact score of the features for WEL.
the biological neural networks that form the basis of the
human brain. Artificial neural networks mimic the functioning artificial neural network. For every matrix, the modified
of the human brain, consisting of layers of interconnected artificial neural network has the best result. For example, in
neural nodes. Nodes refer to these neurons. the R2 squared, we have achieved 0.958064, which is better
than any other model and closer to 1.
The inputs of a biological neural network are represented
by the dendrites; the nodes are the cell nuclei; the weights are In this experiment, we have analyzed eight characteristics
the synapses, and the output is the axons. to predict the day of lockdown after the first identification of
COVID-19 (Fig. 3). Here, the lockdown After the WHO
The modification of ANN (Fig. 2) has been done using the Emergency feature has the highest impact score of 62 percent.
Keras tuner library. It provides the base of the network, and Then, the lockdown date after the first death and a nation's
then some hyper tuning has been done, such as setting the overall GDP has the most significance. These characteristics
learning rate to 1e-5, the dropout to 0.2, the batch size to 32, will play crucial roles if a nation initiates a lockdown after the
tweaking the number of nodes to achieve the best possible first detection in the country. In the recent pandemic, we can
result after tuning with different settings. The final see that many countries also considered the emergency
architecture of the modified ANN has an input layer with eight announcement from WHO and announced lockdowns inside
nodes, two hidden layers of twenty-four nodes, and one output their borders. Then, some countries also considered the first
layer. We have trained our model for 50 epochs by using the death inside the country and their current economic situation
Adam optimizer. when declaring a lockdown.
IV. RESULTS AND DISCUSSION B. WHO Emergency Lockdown
A. First Detection Lockdown To predict the WHO Emergency Lockdown (Table III),
we have trained different traditional machine learning models
To predict the first lockdown (Table II) after COVID-19 and compared them with a modified artificial neural network.
patient(s) detection, we have trained different traditional For every matrix, the modified artificial neural network has
machine learning models and compared them with a modified the best result. For example, in the R2 squared, we have
achieved 0.979, which is better than any other model and
closer to 1.

TABLE II. DIFFERENT MATRICS RESULTS FOR FIRST DETECTION

LOCKDOWN
Model MAE MSE RMSE R2 Squared
Ridge Regression 22.515 1450.321 38.083 0.6887
Linear Regression 19.676 1129.185 33.603 0.7576
Stochastic
19.947 1127.919 33.584 0.7579
Gradient Descent
Lasso Regression 19.558 1123.466 33.518 0.7588
Elastic Net
19.501 1111.180 33.334 0.7615
Regression
Polynomial
21.223 1094.942 33.089 0.7650
Regression
Random Forest
19.424 681.981 26.114 0.8909
Regressor
Modified
Artificial Neural 8.111 262.142 16.190 0.9581
Network

Fig. 3 Impact score of the features for FDL.

Authorized licensed use limited to: Dhaka Univ of Engineering & Technology Gazipur. Downloaded on June 24,2025 at 17:04:37 UTC from IEEE Xplore. Restrictions apply.
 [4] Wang, Zhou, and Alan C. Bovik. "Mean squared error: Love it or leave
it? A new look at signal fidelity measures." IEEE signal processing
TABLE III. DIFFERENT MATRICS RESULTS FOR WHO magazine 26, no. 1 (2009): 98-117.
EMERGENCY LOCKDOWN
[5] Hoerl, Arthur E., and Robert W. Kennard. "Ridge regression: Biased
Model MAE MSE RMSE R2 Square estimation for nonorthogonal problems." Technometrics 12, no. 1
Ridge Regression 18.980 1608.568 40.107 0.6113 (1970): 55-67.
Lasso Regression 17.159 880.241 29.668 0.7873 [6] Zou, H., & Hastie, T. (2005). Regularization and variable selection via
Stochastic the elastic net. Journal of the royal statistical society: series B
16.066 703.796 26.529 0.8299
Gradient Descent (statistical methodology), 67(2), 301-320.
Elastic Net [7] Jain, Anil K., Jianchang Mao, and K. Moidin Mohiuddin. "Artificial
15.923 693.750 26.339 0.8323
Regression neural networks: A tutorial." Computer 29, no. 3 (1996): 31-44.
Linear Regression 15.985 692.435 26.314 0.8327
[8] Bottou, Léon. "Stochastic gradient descent tricks." In Neural networks:
Random Forest Tricks of the trade, pp. 421-436. Springer, Berlin, Heidelberg, 2012.
9.437 214.765 14.654 0.9443
Regressor
Modified [9] Grömping, Ulrike. "Variable importance assessment in regression:
Artificial Neural 3.606 84.387 9.186 0.9796 linear regression versus random forest." The American Statistician 63,
Network no. 4 (2009): 308-319.
[10] Edwards, Jeffrey R., and Mark E. Parry. "On the use of polynomial
In this study (Fig. 4), we also assessed eight factors to regression equations as an alternative to difference scores in
estimate the day on which the WHO will declare an organizational research." Academy of Management journal 36, no. 6
(1993): 1577-1613
emergency lockdown for COVID-19. Here, the Lockdown
after First Detection feature has the highest impact score of 80 [11] Prathyusha, K., K. Helini, Ch V. Raghavendran, and NSL Kumar
Kurumeti. "COVID-19 in India: lockdown analysis and future
percent. The impact of the lockdown after the first death predictions using regression models." In 2021 11th International
feature is the greatest. If WHO intends to evaluate an Conference on Cloud Computing, Data Science & Engineering
emergency statement for a future pandemic, these attributes (Confluence), pp. 899-904. IEEE, 2021.
will play a vital role. In the current pandemic, WHO analysed [12] Darapaneni, N., Reddy, D., Paduri, A. R., Acharya, P., & Nithin, H. S.
early COVID-19 detection and mortality rates in several (2020, October). Forecasting of COVID-19 in India using ARIMA
countries to assess the situation and proclaim a global model. In 2020 11th IEEE Annual Ubiquitous Computing, Electronics
& Mobile Communication Conference (UEMCON) (pp. 0894-0899).
emergency lockdown for COVID-19. IEEE.
[13] Darapaneni, Narayana, Praphul Jain, Rohit Khattar, Manish Chawla,
V. CONCLUSION Rijy Vaish, and Anwesh Reddy Paduri. "Analysis and Prediction of
This study was started to determine the primary cause of COVID-19 Pandemic in India." In 2020 2nd International Conference
why different governments declared lockdowns at various on Advances in Computing, Communication Control and Networking
(ICACCCN), pp. 291-296. IEEE, 2020.
periods. The days between the earliest COVID-positive
[14] Nassr, Rasheed Mohammad, Abdulaziz Aborujilah, Danah Ahmed
detection and the government-issued lockdown were two of Aldossary, and Alia Ahmed Abdullah Aldossary. "Understanding
the many aspects we thought about including, as were the days education difficulty during COVID-19 lockdown: Reports on
between the date of the WHO emergency lockdown Malaysian university students’ experience." IEEE access 8 (2020):
declaration and the government-issued lockdown. Most of the 186939-186950.
time, we see tree-based and/or traditional machine learning [15] Tang, Julian W., Miguela A. Caniza, Mike Dinn, Dominic E. Dwyer,
models surpass the neural network for tabular data. However, Jean-Michel Heraud, Lance C. Jennings, Jen Kok et al. "An exploration
of the political, social, economic and cultural factors affecting how
in this experiment, we have indicated that, with proper different global regions initially reacted to the COVID-19 pandemic."
architecture and tuning, neural network models can provide Interface Focus 12, no. 2 (2022): 20210079.
better results than traditional machine learning models. [16] Van Hoof, Elke. "Lockdown is the world's biggest psychological
Moreover, we found out that our modified artificial neural experiment-and we will pay the price." In World Economic Forum, vol.
network produces some outstanding results in predicting the 9. 2020.
future lockdown date after the first detection of the disease in [17] Ribeiro, Sérvio Pontes, Alcides Castro e Silva, Wesley Dáttilo,
a person and the WHO emergency announcement regarding Alexandre Barbosa Reis, Aristóteles Góes-Neto, Luiz Carlos Junior
Alcantara, Marta Giovanetti, Wendel Coura-Vital, Geraldo Wilson
lockdown. The lack of significant data and other factors, such Fernandes, and Vasco Ariston C. Azevedo. "Severe airport sanitarian
as the nation's economic dependence, the way in which the control could slow down the spreading of COVID-19 pandemics in
public's and government's attitudes toward the new disease are Brazil." PeerJ 8 (2020): e9446.
changing, and the extent to which the government is [18] Kramer, Amit, and Karen Z. Kramer. "The potential impact of the
dependent on imports as opposed to exports, etc., poses some Covid-19 pandemic on occupational status, work from home, and
limitations, though. Because of this, we were unable to occupational mobility." Journal of Vocational Behavior 119 (2020):
103442.
determine how long the first lockdown lasted. Therefore, this
paper's future research may include determining the length of [19] James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). An
introduction to statistical learning (Vol. 112). New York: springer.
the lockdown, making projections about the nation's economic
[20] G. Modak, S. S. Das, M. A. Islam Miraj and M. K. Morol, "A Deep
future, and determining what more can be done to minimize Learning Framework to Reconstruct Face under Mask," 2022 7th
the impact of a pandemic on a new disease. International Conference on Data Science and Machine Learning
Applications (CDMA), Riyadh, Saudi Arabia, 2022, pp. 200-205, doi:
REFERENCES 10.1109/CDMA54072.2022.00038.
[1] Montgomery, Douglas C., Elizabeth A. Peck, and G. Geoffrey Vining. [21] M. S. Nafiz, S. S. Das, M. K. Morol, A. Al Juabir and D. Nandi,
Introduction to linear regression analysis. John Wiley & Sons, 2021. "ConvoWaste: An Automatic Waste Segregation Machine Using Deep
[2] Willmott, Cort J., and Kenji Matsuura. "Advantages of the mean Learning," 2023 3rd International Conference on Robotics, Electrical
absolute error (MAE) over the root mean square error (RMSE) in and Signal Processing Techniques (ICREST), Dhaka, Bangladesh,
assessing average model performance." Climate research 30, no. 1 2023, pp. 181-186, doi: 10.1109/ICREST57604.2023.10070078.
(2005): 79-82. [22] Morol, M. K., Rokon, M. S. J., Hasan, I. B., Saif, A. M., Khan, R. H.,
[3] Chai, T., & Draxler, R. R. (2014). Root mean square error (RMSE) or & Das, S. S. (2022, March). Food recipe recommendation based on
mean absolute error (MAE)?–Arguments against avoiding RMSE in ingredients detection using deep learning. In Proceedings of the 2nd
the literature. Geoscientific model development, 7(3), 1247-1250. International Conference on Computing Advancements (pp. 191-198).

Authorized licensed use limited to: Dhaka Univ of Engineering & Technology Gazipur. Downloaded on June 24,2025 at 17:04:37 UTC from IEEE Xplore. Restrictions apply.