1. Introduction

Infectious diseases occur due to a number of pathogens that transmit easily from person to person, animal to animal, or person to animal. There are different ways that lead these pathogens to spread, and their transmission is fast ^[1]. It is essential to detect these infectious diseases at the initial phase for their prevention and control.

An impenetrable viral pneumonia was reported in Wuhan, China, in December 2019. The World Health Organization (WHO) termed this virus Novel Corona virus (2019-nCoV) on January 12, 2020, and Coronavirus Disease 2019 has occurred due to this virus. This virus is the 7$^{\mathrm{th}}$ known species of coronavirus due to which human beings have suffered. Various symptoms are found in patients who suffer from 2019-nCoV, such as fever, mild cough, or pneumonia, which causes death ^[2].

The mortality rate of COVID-19 is approximately 2 to 4%, although these are very initial percentages, and variation could be found when more information is available. The 2019 novel coronavirus is an uncompromising and unexpected virus. Initially, this virus is capable of camouflaging itself ^[3]. This virus can occur in an asymptomatic manner, in which the patient can be seen as healthy, or it can be visible through various observable symptoms that a person suffers from in a common respiratory disease. It also has a long latency period. A report from Lancet states that the median latency time of 2019-nCoV can be 20 days and may reach up to 37 days for some patients. Furthermore, the transmission of this virus occurs in diverse ways. Similarity can be seen in the spreading process of coronavirus and several other infectious diseases ^[4].

At present, there is no precise treatment available for COVID-19, but treatment of a number of signs/symptoms is possible. This treatment is provided in accordance with the clinical position of the patient. Furthermore, supplementary care is proven efficient for infected patients. This supplementary care or self-protection consists of hygienic eating routines, maintenance of basic hand and respiratory sanitation, and the avoidance of close interaction with other people who have signs of a respiratory disease like cough and sneezing.

During lockdown, factories are closed, schools are suspended, and people are isolated in their own homes, from which daily life is troublesome as COVID-19 is spread extensively. Thus, the prediction and analysis of the progression of COVID-19 has become very crucial ^[5]. Prior studies have made use of various predictive models, such as regression prediction models, Markov chain models, Bayesian networks, and machine learning (ML) schemes to evaluate the occurrence, transmission, and variation of communicable diseases. A number of earlier studies were conducted on the basis of research regarding Influenza, HIV, and SARS.

There are 4 modules in a COVID-19 prediction model. These modules are feature selection, network search, local trend prediction, and fuzzy rule-based risk assessment ^[6]. Most researchers have used COVID-19 dataset1, which has dates, countries, confirmed cases, recovered cases, and total deaths for empirical evaluation of their proposed COVID-19 prediction techniques. In the initial phase, it is assumed that there is no connection between the attributes and the predictive variable. This data has three major issues related to categorizing the risk of a country. The number of cases is represented with ${\kappa}$, the number of deaths is ${\delta}$, and ${\rho}$ shows the number of recovered patients. ${\kappa}$ ${-}$ (${\rho}$ + ${\delta}$) is used to compute the active cases ${\alpha}$.

A backward elimination scheme is implemented to choose the attributes ^[7]. Ordinary least squares (OLS) regression is employed to compute ${\rho}$ of all features through ${\alpha}$. The attributes are selected using a threshold. In the second phase, $\varphi _{1},\,\,\,\varphi _{2},\ldots \ldots ,\,\,\,\varphi _{n}$ are assumed as different hyper-parameters of a learning algorithm, and ${\mathrm{\mu}}$$_{1}$, ${\mathrm{\mu}}$$_{2}$, {\ldots}.., ${\mathrm{\mu}}$$_{\mathrm{m}}$ are the domains of the parameters. The training set (D$_{\mathrm{train}}$) and test set (D$_{\mathrm{test}}$) are two categories of the dataset D. The hyper-parameter space is denoted with $\theta $= $\mathrm{\mu}_{1}\times \mathrm{\mu}_{2}\times \ldots .\times \mathrm{\mu}_{m}.$ Training data is used for training on $\varphi \in \theta .$ The test error E(${\varphi}$, D$_{\mathrm{train}}$, D$_{\mathrm{test}}$) is considered as the error on D$_{\mathrm{test}}$ of the parameter ${\varphi}$. The optimization of a hyper-parameter is computed using Eq. (1) for the dataset (D) with minimization.

Generally, the hyper-parameter search is very costly because there is a necessity of training and evaluating the dataset for every combination of parameters ^[8]. The risk of a country is forecasted using the prediction of ${\delta}$, ${\kappa}$, and ${\rho}$. Initially, the death rate (Eq. (2)), new case rate, and recovery rate are computed using Eqs. (3)-(5), respectively. A Gaussian fuzzy membership function is utilized to reveal the risk measurement of these parameters. Rules are entailed in estimating the risk of the final class.

An Artificial Neural Network (ANN) is an idea that is influenced by a biological nervous system and is used to process information like the brain. The novel structure of the information processing system is the major component. In ANNs, programming knowledge is useful to plan a data structure that can play the role of a neuron ^[9]. This data structure is known as a node. An educational algorithm is employed to train the network among these nodes.

The nodes consist of two states: active and inactive states. Every edge includes a weight. The next inactive node is activated through positive weights, and the next connected node is inactivated through negative weights. Long Short Term Memory (LSTM) is a modified version of a recurrent neural network (RNN). An RNN is capable of handling a sequence with a recurrent hidden state that is activated each time on the basis of the preceding time ^[10]. With a set of inputs x = (x$_{1}$, x$_{2}$, {\ldots}., x$_{\mathrm{t}}$), the hidden state $h_{t}$of an RNN is estimated using Eq. (5).

In Eq. (5), $v$ denotes a non-linear function. The LSTM generates an output$y$output $y=\left(y_{1},y_{2},\ldots .y_{t}\right)$. The hidden states are updated using Eq. (6):

In Eq. (6), $g$ represents a bounded function. A universal RNN predicts the conditional probability of every input state using Eq. (7).

LSTM is flexible and estimates dependencies of various time scales. LSTM is a popular variant of an RNN. LSTM makes use of gate and memory cells for predicting sequences. Network-based algorithms consider the communication between various parts. Nevertheless, the contact network G is hidden (and subsequently, the infection probability matrix B) and must be inferred from the epidemic outburst ^[11].

A version of a network inference-based prediction algorithm (NIPA) has been implemented to predict the spread of COVID-19 in Hubei, China, and Italy. There are mainly two steps included in NIPA. The first step infers the original infection matrix from the epidemic outburst. The next step makes use of the infection matrix B and the estimated curing rates$~ \delta _{i}$ for node $i$ for outbreak prediction by iterating the Susceptible-Infected-Removed (SIR) model on the estimated infection matrix B. NIPA effectively predicted the spread of COVID-19 in the Chinese area of Hubei, but the hidden contamination matrix B could not be induced.

2. Literature Review

Massimo A. Achterberg et al. ^[12] described that researchers belonging to different scientific disciplines made attempts for predicting the spread of COVID-19 ^[12]. NIPA was investigated for predicting COVID-19, in which ML and SIR models were integrated. The interaction among various regions was taken in account in this algorithm. The evaluation revealed that the original NIPA provided the finest performance to predict the spread of COVID-19. Superior accuracy was obtained from the presented algorithm in comparison with another algorithm. Additionally, network-based prediction was better, and the presented algorithm outperformed an LSTM network with regard to predictive error.

Md Masud Rana et al. ^[13] suggested an innovative -19 stochastic model for securing communication systems to transfer and store epidemic data ^[13]. The observable and non-observable COVID-19 conditions included infection, cases, recovery, and deaths, which were taken into consideration to construct and authenticate the dynamic COVID-19 model. In order to transmit the local observation information to a central control center, a secure communication system based on IoT was put forward. COVID-19 was predicted by constructing and deploying an optimal signal processing algorithm. The outcomes demonstrated that the suggested approach was adaptable for predicting the phases of COVID-19 in a short period of time. As a result, the constructed simulator and analysis provided a significant resource to forecast the state and proactive action of COVID-19.

Michał Wieczorek et al. ^[14] discussed the research outcomes to design a neural network framework to forecast the spread of COVID-19 ^[14]. The proposed predictive framework is planned on the basis of a traditional approach along with a deep architecture whose learning is accomplished through a Nadam training model. The training phase is carried out using official data taken from governmental and open repositories. The predictive outcomes are evaluated for countries as well as for regions for presenting the possibly wide spectrum of values related to forecasting the COVID-19 spread. The outcomes indicated that a higher accuracy of 99% is provided through the proposed framework.

Chamara Sandeepa et al. ^[15] discussed the tracking of the social interaction of users and forecast the infection possibility on the basis of social interactions ^[15]. Initially, a social interaction tracking system was built depending upon Bluetooth Low Energy (BLE) and the Global Positioning System (GPS). An algorithm was established to forecast the probability of COVID-19 infection on the basis of the collected data. At last, the deployment of a prototype of the system was accomplished by means of a mobile app and a web monitoring tool. Furthermore, a graph-based model was utilized to carry out the simulation for the analysis of the behavior of the established algorithm. The simulation showed that self-isolation is essential to slow down the development of this disease.

Hossein Abbasimehr et al. ^[16] recommended three hybrid approaches based on time series techniques to predict COVID-19. Three deep learning (DL) models known as multi-head attention, LSTM, and a convolutional neural network (CNN) were integrated with the Bayesian optimization algorithm ^[16]. The performance of prediction was improved in an automatic manner using a Bayesian optimization technique. The epidemic data collected from the Coronavirus Resource Centre of Johns Hopkins University (JHU) was employed for performing the experiments. The experimental outcomes validated that the recommended DL frameworks are better than the benchmark approach.

Refat Khan Pathan et al. ^[17] intended to use an LSTM algorithm based on an RNN for forecasting the future mutation rate of COVID-19 ^[17]. The nucleotide mutation and codon mutation were verified at a separate level by processing the gathered data. The Root Mean Square Error (RMSE) obtained in testing using LSTM was computed as 0.06 and 0.04 for the training phase. The training and testing phases assisted in predicting the nucleotide mutation rate of 400 patients at a future time. This exhibited that the intended algorithm was capable of forecasting the day-basis mutation rates in the availability of more data of patient data in updated time.

Furqan Rustam et al. ^[18] presented a study based on ML approaches to predict the number of upcoming patients who suffer from COVID-19 ^[18]. Generally, this study made implementation of 4 basis predictive techniques in the prediction of the threatening factors of COVID-19: linear regression (LR), least absolute shrinkage and selection operator (LASSO), support vector machine (SVM), and exponential smoothing (ES). The outcomes generated in this study presented that the proposed system is efficient for the recent scenario of the COVID-19 pandemic. Additionally, ES provided efficient performance among other techniques to predict the fresh confirmed cases, death cases, and recovery rates.

Logan Ryan et al. [190 emphasized constructing and computing ML-based algorithm for forecasting the mortality in COVID-19 and ventilating patients in a mechanical way ^[19]. The MIMIC dataset, which includes the data of 53,001 ICU patients, was used to conduct the retrospective study. The constructed algorithm was proven as a valuable predictive tool that assisted in predicting the mortality of a patient at clinically useful time points. This algorithm successfully predicted the morality for mechanically ventilated patients and the patients who suffered from pneumonia and COVID-19 with high accuracy.

K. Shyam Sunder Reddy et al. ^[20] presented a neural network-based technique for the forecasting of COVID-19 cases ^[20]. A dataset collected from India was utilized to implement the LSTM based on RNN to predict COVID-19 cases. The performance of the presented network was analyzed, and comparison was accomplished with parameter mitigated variants of LSTM. The concealed unit signals, bias, and input signal were removed to achieve those variants. The performance of the presented approach was computed with regard to the mean square value.

Ertuğrul Karaçuha et al. ^[21] discussed the utilization of a former Deep Assessment (DA) technique to model a dataset initially ^[21]. Two techniques were implemented to predict COVID-19: a DA technique and LSTM.

Subsequently, a Gaussian predictive model was suggested for forecasting the short-term future of the pandemic. The quantification of the model was also performed. The suggested model was compared with the SIR model. Finally, wavelet-based de-noising and correlation coefficients were employed to analyze the impact of history on memory vectors. The outcomes exhibited that the suggested Deep Assessment technique provided average errors of 0.6671% for confirmed cases, 0.6957% for a recovered case, and 0.5756% for death cases.

In Table 1, a comparison of various proposed COVID-19 prediction techniques is systematically presented. COVID-19 prediction has various steps, like pre-processing, feature extraction, and classification. Various researchers proposed such techniques for COVID-19 prediction. The techniques proposed by the researchers are based on ML. The ML technique predicts COVID-19 and improves accuracy. The techniques are compared based on outcomes, advantages, and disadvantages.

3. Conclusion and Future Scope

In this work, the various predictive algorithms proposed in the literature for the purpose of the analysis of COVID-19 were compared based on their outcomes, pros, and cons. These COVID-19 techniques are based on ML and optimization techniques. It is concluded that ML methods outperformed optimization algorithms. In the future, hybrid ML methods can be designed for COVID-19 prediction for better prediction results.