key: cord-0057247-5ooi3qhm
authors: Chaudhary, Anshu; Sharma, Manisha
title: Multilayer Neural Network Design for the Calculation of Risk Factor Associated with COVID-19
date: 2021-03-09
journal: Augment Hum Res
DOI: 10.1007/s41133-021-00044-4
sha: 64eb1cd6727f21b3d8112022ba29ee79080641d0
doc_id: 57247
cord_uid: 5ooi3qhm

Coronavirus disease 2019 (COVID-19) is a public health emergency and is of international concern. Till now, there is no effective pharmaceutical treatment available for this disease. This paper presents a multilayer neural network algorithm to calculate the risk factor of getting COVID-19 to the individual based on the symptoms described by World Health Organization. The aim of this study is to provide an approximate risk factor of getting COVID-19 to an individual that helps for further treatment.

At present, the whole world is facing an outbreak of a new virus disease that is killing people more rapidly than any other chronic virus disease. World Health Organization (WHO) named it coronavirus disease 2019 (COVID-19) on January 10, 2020. The common symptoms of COVID-19 are fever, tiredness and dry cough. COVID-19 disease, like any other flu, is highly communicable and spreads from one person to another when a person comes in contact with an infected person. It also spreads when a person touches a surface or object that has the virus on it and then accidently touches their eyes, nose or mouth and rarely from facial contamination. This virus has affected not only society but also economics around the world, and it will permanently rephrase the living of world as it continues to unfold. A numerous number of people have been infected by this disease, and millions have been died in the world till now (https://www.int/health-topics/coronavirus2020). Scientists around the world are engaged in searching the medicines for the cure of this disease and vaccines to prevent individuals from getting COVID-19. We hope that scientist will be successful in making its effective vaccine in future [11, 12] .

''Artificial neural network'' is very connotative term. It is proposed that machines are similar to brain and likely laden with the science fiction connotative of the Frankenstein myths. One of the main features of artificial neural networks is that it has been vaguely inspired by the biological neural network that constitutes brain and nervous system. It is very useful in many branches like science, engineering, mathematics, etc. In neural network, the main objective is to do this in a nontechnical way as possible, while some mathematical notations are necessary for specifying certain rules, procedures and structures significant. A unique feature of artificial neural network is that its ability to establish empirical relationships between independent and dependent variables, and further extract important information and complex knowledge from representative datasets. Ability of this network is stored in the inter-unit weights, connection strengths acquired by the adaptive process, or learning from a set of training patterns [13] .

Artificial neural network is generally used in statistical analysis and data modelling, where their role is to recognize as a substitute to standard nonlinear regression or cluster analysis techniques (Cheng and Herington 1994). Neural networks have been applied in diverse fields such as speech recognition, temporal character recognition, medical diagnosis, geological survey for oil, and financial market indicator predictive. Perhaps artificial neural network is a thought of simplified model of neurons that occur naturally in the animal brain. From the biological angle, the necessary requirement for the neural network is that it allows learning by example from representative data that describes a physical phenomenon or a decision process. For an engineering purpose, this correspondence is not necessary and it offers an alternative form of parallel computing. The straightforward artificial neuron is the threshold logic unit. Its basic operation is to perform a weighted sum of its inputs which comes under threshold logic unit and then gives output as ''1,'' and if this sum exceeds a threshold logic unit, then it gives output as ''0.'' Threshold logic unit is the basic model of Integrate-and-fire mechanism of actual neuron [13, 14] .

Artificial Neural Network Architecture ANN neural network uses the processing of the brain as a basis to develop algorithm that can be used to model complex patterns and prediction of problems. A neural network is an accumulation of highly interconnected processing elements called neurons that have the capability to understand the problem and thereby knowledge being used by the network to solve that problem. A difficult task with ANN involves choosing the architecture parameter of the network. In a single layer network, there is only one layer of connection weight, where the units can be distinguished as input units, which receive signals from the outside world and there are output units from which the output of the network can be deliver.

A feedforward ANN with single layer is assumed in this research. In feedforward network, information flows from input units to the output units. A neural network may be viewed as a collection of communicating simple processing elements or units. These units are a serviceable abstraction of the neurons in the central nervous system. A unit is a simple processing element, connected to other units by its ''weighted'' (dendritic) connections, as shown in Fig. 1 .

A unit collects weighted (w 1 to w n ) numerical information from other units (x 1 to x n ). This knowledge sometimes increased with time and is summed to the net input (y_in). The y_in is passed through an activation function F, resulting in the activation Ai of the unit. After the summed input has been passed through the activation function, the activation (A i ) is connected by other units that are connected to other units. Activation function can be (non) deterministic binary or (non) deterministic continuous. The activation function has been chosen according to the functionality required in the neural network. Neural network has two basic functions: First they can be trained to remember some information [7] , and they can be used to perform constraint satisfaction and optimization task [4, 6] . Neural network being used as learning tools has demonstrated their ability to capture the relationship between variables that are usually difficult to relate each other analytically by learning, recalling and generalizing from training patterns as data. In other words, they are universal function approximations and are attractive for automatically learning of (nonlinear) functional relationship between the input and output variables.

A multilayer neural net has more than one layers of nodes between the input units and the output units. Thus multilayer neural network possessing an input and output layer also has one or more intermediary layer called hidden layer. Between the two adjacent levels of neurons layer (input, hidden and output), there is a layer of connection weight which carries information in terms of synaptic weight. The input layers neurons are linked to the hidden layer neurons through synaptic weight, and the weights on these links are stated as input-hidden layer weights [9] . Also, the hidden layer neurons are linked to the output layer neurons, and the equivalent weights are stated as hidden-output layer weights. MLNN can solve complicated and difficult problems, which are very time taken or sometimes impossible to solve by using single layer networks.

As the number of input and output units are verbalized by the dimensionality of the input vectors and the number of classifications, respectively. The number of hidden layer is not simply related to such obvious properties of the classification problem. The number of hidden units depends of the complexity of the decision boundary [10] . If the patterns are well separated or linearly separable, then few hidden units are required. But if the patterns are drawn from complicated compactness that are highly interspersed, then more hidden units are required. A multilayer neural network with P input neurons, Q hidden layer neurons and R output neurons in the output layer is written as P-Q-R, as shown in Fig. 2 .

Setting the weights based on training patterns and desired output is the fundamental problem. The basic methodology used in learning the neural network is to start with an untrained network, present a training pattern to the input layer, pass the signals through the network and determine the output at the output layer. The weights cannot be initialized as zero, else learning cannot takes place. Therefore, there is a need to confront the problem by selecting the starting/initial value. For setting weights in a given layer, there is a need to set weights randomly in a uniform distribution which helps in uniform learning. As data standardization gives positive and negative values equally, and on an average, there is also a need of positive and negative weights, hence weights were chosen from an uniform distribution -w \ w \ (w ? w) for some w that are yet to be determined. If w is set to too small, then the linear model will be implemented. Alternatively, if w is set too large, the hidden unit may saturate even before learning originates. Since net activation at a hidden unit (net_j) = ± 1 is the limits to its linear range, value of w was set as such that the net activation at a hidden unit is in the range 0 \ net_j \ ? 1. The method of choosing the values of the weights (training) is an important individual characteristic of different neural networks. Supervised learning is used in this paper. In supervised learning process, every input pattern that is used to train the network is associated with a target or desired output vector. The weights are then adjusted according to the learning algorithm. A teacher is supposed to supervise during the entire learning process to determine the error, when a comparison is made between the network's computed output and the correct likely output. In this type of learning process, the output is binomial element, i.e., either 1 if the input vector belong to the class or 0 if the input vector does not belongs to the class. Multilayer neural networks can be trained to perform a non-linear mapping from an n-dimensional space of input vectors to an m-dimensional output space.

A novel human coronavirus, named as severe respiratory syndrome coronavirus subsequently named as SARS-COV-2, was first reported in Wuhan, China, in December 2019 [1, 2, 5] . COVID-19 is spreading rapidly throughout the world, almost in all developed and developing countries. The first case of COVID -19 was reported in INDIA on January 30, 2020. At present, the nucleic acid test (NAT) can be considered as the most reliable clinical method for the diagnosis of COVID-19. However, the first step for any nation to fight with this disease is to detect it. If doctors already know that the particular individual is at low or high risk of getting COVID -19, then it would be beneficial for them to treat that individual. The aim of this MLNN model is to calculate the low or high risk factor of getting the disease to an individual on the basis of parameters described by the World Health Organization. Common signs and symptoms of COVID-19 infection include fever, coughing and shortness of breath. The virus that causes COVID-19 is mainly transmitted through microdroplets generated from an infected individual when they coughs, sneezes or exhales. When the person comes in contact with the infected person, then there is an urgent need for medical treatment. Hence, identification, tracing and elimination of factors that are responsible for the spreading COVID-19 are very much required [8] .

Following notations are used in this paper.

Artificial Neural network MLNN: Multilayer neural network X i :

input values for the hidden neuron j, j = 1,2…11. Y k :

Net output from input neuron j, j = 1, 2, …11. z_ in :

Net input from hidden neuron. F(x):

Activation function. W ij :

Weights on connection links from neuron i to neuron j. Exp:

Exponential function.

This multilayer neural network-based algorithm is applied to get the high or low risk factor of getting COVID-19. For this, we take eleven basic symptoms as input variables for MLNN and three variables in hidden layer and one output variable. Sigmoid function is used as an activation function to determine the output value for the multilayer neural network (MLNN). Defining various symptoms as input variable as X 1 = Travel history. X 2 = Contact history. X 3 = Fever. X 4 = difficulty in breathing.

X 5 = Sore throat. X 6 = Fatigue. X 7 = Cough. X 8 = Diarrhea. X 9 = Loss of senses of smell and taste. X 10 = Age. X 11 = Other clinical diseases.

Defining input values to above input variables as: 

All the information about the number of neurons in first, second and the third layer, values of the weights on the connection links and the function which are used in the hidden and output layer of the completely trained neural network is used in the present study (Fig. 3, Table 1 ).

The aim of multilayer neural network is to calculate the parameter that describes the high and low risk factor of COVID-19 for an individual.

Step 1: Give values to all input variables x i for all i = 1,2,…,11 accordingly as defined above.

Step 2: Compute the values of weights on the connection links of the hidden neuron as.

• For the first hidden neuron, divide each input value by 10. • For the second hidden neuron, divide each input value by 100. • For the third hidden neuron, divide each input value by 1000.

Step 3: Calculate the activation values of net input y j_in for the first, second and third hidden neurons by using.

Step 4: Calculate the output values y k by using the following functions.

Y k = F(y _in ) k = 1, 2, 3.

where F(x) = 1 1þexpðÀxÞ

Step 5: Input values for the output neurons are obtained by Y K , k = 1, 2,3.

Step 6: Compute the value of weights on the connection links from hidden neurons to output neurons as: w z1 = y 1 /10, w z2 = y 2 /100, w z3 = y 3 /1000. Step 7: Calculate the activation values of net input Z _in for the output neuron by using Z in ¼ P 3 i¼1 y i w iz .

Step 8: Compute the final output values Z k by using the following function.

where F(x) = 1 1þexpðÀxÞ

Step 9: End. 

India is a second most populous country and seventh largest country by area in the world. So far, the cases of COVID-19 pandemic are under control in comparison with USA, Italy, France, Iran, etc. Still, India needs to be more focus towards minimizing the cases and fight against COVID-19 pandemic. In the present research, three different cities, viz. Chandigarh, Panchkula and Mohali, were selected and approximately 180 peoples were randomly selected and were diagnosed on the basis of COVID-19 symptoms described by the World Health Organization. The proposed multilayer neural network model correctly gives everyone's results of high and low risk factor of COVID 19 (Table 2) .

The final output value of the output neuron lies between 0 and 1. The output neuron value can be categorized into two categories, i.e., one is low risk of getting COVID-19 and second is high risk of getting COVID-19. The minimum and maximum values of output neuron from artificial neural network are observed as 0.5070 and 0.5127, respectively. If the output value of the output neuron lies between 0.5070 and 0.5094, then it comes under low risk of COVID 19, and if it lies between 0.5095 and 0.5127, then it come under high risk of COVID 19. Based on some random samples, numerical calculations are shown in Table 3 . Table 3 shows that the output neuron values of randomly taken sample lie within the range and give approximately appropriate result. The study shows the applicability of multilayer neural network (MLNN) to real-world problem. The future directions of this study are to train the different networks based on neural network that will help to cure COVID-19.

Education and the COVID-19 pandemic

An introduction to computing with neural nets. IFFF Access, speed signal process

Parallel distributed processing: explorations in the micro structures of cognition

Table 3 Details of numerical calculations S.no Input values x i , i = 1,2…11 Net input Y j_in , j = 1,2,3 Hidden neuron input value Y k, k = 1

Neural'' computation of decisions in optimization problems

A crew scheduling problem

Passively self-aligned assembly of compact barrel hinges for high-performance, out-ofplane MEMS actuators

Compact, scalable, highresolution, MEMS-enabled tactile displays

SARS and MERS: recent insights into emerging coronavirus

Nonpharmaceutical measures for pandemic influenza in nonhealthcare settings-social distancing measures

Pneumonia of unknown cause-China: disease outbreak news. Geneva: World Health Organization

COVID-19) situation report -52

Clinical characteristics of coronavirus disease 2019 in China

to green zone None of the above