An Improved COVID-19 Forecasting by Infectious Disease Modelling Using Machine Learning

Abstract

The mechanisms of data analytics and machine learning can allow for a profound conceptualization of viruses (such as pathogen transmission rate and behavior). Consequently, such models have been widely employed to provide rapid and accurate viral spread forecasts to public health officials. Nevertheless, the capability of these algorithms to predict outbreaks is not capable of long-term predictions. Thus, the development of superior models is crucial to strengthen disease prevention strategies and long-term COVID-19 forecasting accuracy. This paper provides a comparative analysis of COVID-19 forecasting models, including the Deep Learning (DL) approach and its examination of the circulation and transmission of COVID-19 in the Kingdom of Saudi Arabia (KSA), Kuwait, Bahrain, and the UAE.

1. Introduction

An atypical coronavirus, COVID-19, was classified in Wuhan, China, in December 2019. As of 8 September 2021, the newly identified virus had become liable for hundreds of thousands of confirmed cases and over 4,610,056 deaths globally [1]. In addition to anti-viral preventative methods such as wearing masks and avoiding crowds, the widespread distribution of SARS-CoV-2 vaccines is fundamental to mitigating viral spread as many vaccines have demonstrated an efficacy of over 95% in inhibiting SARS-CoV-2 symptoms. Around 545,624 COVID-19 infections and 8598 coronavirus-related casualties were disclosed KSA alone. Its swift and ceaseless transmission has induced a continuous pandemic, and as the number of cases increases, the urgent need to “flatten the curve” has become fundamental to evade the calamitous burdening of health institutions throughout the affected nations. Therefore, the critical need to elucidate the pathogenic mechanisms and epidemiological components of COVID-19 have led mathematical modeling and simulation to become the primary techniques for forecasting outbreaks. Predicting COVID-19 spread must be considered as it has the potential to aid the time-sensitive resolutions of healthcare personnel, strengthening the advancement of prevention strategies. In the literature, different mathematical epidemic surveillance models [2,3,4] have been presented related to COVID-19 diseases to forecast and predict disease occurrence patterns and trends under different situations. These patterns enable policies at the government level to stop the spread of disease and decrease the chance of infection. Choi et al. [5] applied the Susceptible–Infectious–Recovered (SIR) model to forecast the SARS disease cases and deaths in Canada. Their finding estimated the fast spread of the SARS disease; every five days, the infection spread was about 1.5 to 3 per infection. Peng et al. [6] employed a more comprehensive mathematical model for Susceptible-Exposed-Infectious-Removed (SEIR) by considering new variables (a quarantined state with a recovery state) using the public data of NHC of COVID-19.

Furthermore, for the present prognosis of disease flareups globally, clinical data from the Johns Hopkins dashboard has consistently held the attention of numerous modeling simulation experts [7]. For instance, Goncalves D. Balcan et al. [8] reported that the primary reproduction number (R0; transmissibility of a virus), mortality, forecast per day of infection, and recovery rates are forecasted by employing the Susceptible–Infectious–Recovered–Dead (SIRD) model. In addition, the fatality and recovery case ratios were predicted with 90% confidence intervals, and the outbreak expansion was predicted for three weeks. Nevertheless, these models display much ambiguity in the long-term prediction of disease flareups due to the absence of necessary data. Thus, developing mechanisms using artificial intelligence (AI), such as the ML and DL approaches, is fundamental for analyzing clinical data during extended periods.

As AI can help overcome the problems created due to the COVID-19 pandemic more efficiently by employing data mining techniques. AI exhibited a significant impact from different sides to tackle the COVID-19 issues. The system reviews show that the statistical and machine learning (ML) models have been widely used for various tasks such as COVID-19 forecasting, predicting, etc. Now, ML techniques are used worldwide for predictions because of their performance. However, machine learning (ML) techniques. However, a few challenges of using ML techniques are data availability, hyper tuning of parameters, and selecting the best ML model for prediction. Therefore, researchers have made predictions and chosen the best ML model based on the available datasets. Various ML models are mentioned in the literature, but the deep learning (DL) models give better performances for forecasting the COVID-19 cases [9]. The DL approach establishes the localized transmissibility (R0) of a virus from a given dataset and grants the prolonged investigation of data to understand the mechanism of COVID-19 transference and the apprehension of susceptible districts and emerging hotspots. This study explores various forecasting approaches ranging from the mathematical model to the DL models in investigating COVID-19 outbreak trends to understand further its proliferation patterns and the recognition of susceptible communities and evolving hotspots.

2. Related Work

This segment contains a review of data analyses that have investigated COVID-19 transmission trends. Various forecasting techniques (statistical, analytical, and mathematical) have been discussed in the literature by applying different data sources for COVID-19 analysis. In addition, they considered different parameters of the existing dataset. In past pandemics, researchers have applied traditional mathematical approaches [9] to forecast the spread of disease and mortality rate in a specified period. These traditional methods were effective and gave a better performance in terms of forecasting. Due to this reason, in the current COVID-19 pandemic, most researchers have used these traditional mathematical models to predict the spread and mortality rates [10,11,12,13]. For example, Sameni [14] used the well-known and simple mathematical model known as the SIR model in his study to find the pattern of COVID-19. Their finding show that countries that have taken concrete measures and other facilities (quarantine, etc.) to stop the spread of COVID-19 disease have a positive impact on the mortality rate. However, this mathematical model shows a lower performance due to the dynamic nature of the COVID-19 dataset. Yuan et al. [15] introduced Boltzmann’s function-based analysis. Their finding shows that this model gives better forecasting about the spread of disease and helps the government body take proper actions at the right time. Dowd et al. [16] applied mathematical modeling using age and gender parameters on mortality rate. The COVID-19 virus mainly affects the elderly age group (above the age of 65). Therefore, policies such as quarantine and social distancing can assist in stopping the spread of the virus.

GLEaM, a computational model for visualizing viral spread (Balcan et al. [8]), can effectively evaluate lifelike scenarios compared to data. As the modeling system’s development progresses, transportation and interaction layers depend on newly evolving epidemic strains. The GLEaM model capitalizes on one’s portability to construct epidemic simulations through the assimilation of unlike processes. Moreover, health personnel may reveal clinical data much later, impeding epidemiologic progression. Real-time data contained in informal media permit the advancement of epidemic forecasts. For example, data from Twitter and HealthMap for Haitian cholera outbreaks were analyzed for the first 100 days of 2010 and compared to government-authorized case statistics. The trends of two-week cases of both unofficial data origins correlated with the official case [17]. Similarly, in the example of Hurricane Tomas, guesses of the reproductive numbers were found to range from 1.04 to 1.51 (informal) and 1.06 to 1.73 (official) [18]. Thus, fundamentally comparable data were achieved significantly earlier than the dispensation of official case counts.

Therefore, unofficial media can be utilized as a cost-effective point of supply for prompt assessments of fluctuations in transmission trends. In addition, the Susceptible–Exposed–Infected–Recovered (SEIR) models were employed to compose a simulation inquiry to examine viral spread tendencies and predict new cases in 51 countries. Epidemiological parameters separated the nations into four aggregations that represented the following risk levels: high risk, medium risk, low risk, and efficacious control.

Likewise, Approximate Bayesian Computation (ABC) algorithms can also forecast infectious disease trends [19]. Three viral spread case studies highlight the user-based components that heighten processing rates and reliability. Furthermore, Pujari et al. suggested a method of providing swift and authentic epidemic forecasts for public health officials in India. Although COVID-19 transmission has been thoroughly examined, most related algorithms are either too computationally expensive, too granular to be dependable, or too fine-grained to be competent. Therefore, a compound approach of the Susceptible–Infected–Recovered (SIR) model examines transportation trends to incorporate various intra-city populations and intercity coupling, permitting accelerated and reliable COVID-19 forecasting in India. Through its use, India’s urban population was predicted to encounter COVID-19 in 90 days of the pandemic under any more minor conditions than the implementation of severe cautionary measures. The report concluded that the combination of public transportation, a small number of COVID-19-infected individuals, and the incredibly contagious essence of COVID-19 is satisfactory for the rapid transmission of the epidemic. The researchers [20] state that the diligent guidance of domiciliary transportation structures is crucial despite the restrictions on international travel.

Furthermore, Chae et al. developed a model that exploits distinct deep learning techniques as the elemental composition and several input variables. The prime task was to draft an enhanced adaptation of existing models. The specifications of these algorithms were optimized, allowing an increase in functionality in both Deep Neural Networks (DNN) and Long Short-Term Memory (LSTM) models compared to traditional prediction models by an average of 24% and 19% when predicting infectious diseases. An extensive examination of these deep learning algorithms uncovered balance in the DNN model’s functioning, while the LSTM model consolidated on certainty during the rise of a contagious disease. It is thus accepted that these forecasting frameworks can curtail hindrances in broadcasting institutions and permit prompt feedback, viral transmission monitoring, and the supervision of expenditures [21].

Table 1 below provides a summarization of models used and the pros and cons of each related work.

3. Materials and Methods

Using the dataset provided by Kaggle [24]. This segment contains a review o, we assessed the efficiency of various traditional epidemiologic forecasting frameworks, ML and DL, for anticipating the spread of contagious ailments, as shown in Figure 1. This dataset contains the most up-to-date COVID-19 statistics chronologically ordered, e.g., the number of reported cases, deaths, and recoveries.

The analysis process mainly includes data collection, data preprocessing, model training, and prediction. We utilize various python packages for the analysis process, including Pandas, Numpy, SKlearn, Tensorflow, Keras, and more. The model performance is directly related to the inputted data/information quality. Therefore, we employed preprocessing steps (i.e., data cleansing, data transformation) to improve the quality of any prediction models. By checking the missing values using Pandas, we have summed all of the null values in each column in the dataset. Then, the blank spaces were filled with “No Province/State”. Irrelevant values and useless pieces of data were removed from the dataset. The columns “Province/State” and “Last Update” were dropped. To convert the date type, date information was loaded as a string rather than chronologically.

Furthermore, the Pandas data frame columns containing dates were converted into date/time objects using the date/time object in Python with Pandas. Next, the index was converted to a date column. As the index is like an address, any data point across the data frame or series can be accessed through its use. Then, the index was changed to the “Observation Date”. To add new features and to expand the dataset, more relevant features were added. To understand the daily trend, columns for daily confirmed cases, recoveries, and deaths were created to calculate the difference between the current value and the previous day’s value. Moreover, to filter by a country or use observation date, confirmed cases, death cases, and recovered attributes, the growth of each country’s different types of cases until 19 January 2021, has been visualized. The subsequent sections describe the specifics of each of the forecasting frameworks.

3.1. Mathematical Models

In the mathematical models, we used the Susceptible–Infected–Recovered (SIR) model [4], a straightforward and fundamental epidemiologic model for the spread of disease prediction. This model is mainly used to model infectious diseases and more effectively predict the infectious disease spread from human to human, where recovery is ongoing struggle. In addition, this model is used if individuals gain lasting immunity at the end of the disease. The SIR model has a different version, but the basic SIR model has three components: susceptible (S), infectious (I) and recovered (R). The SIR framework allows for a population of size N in which S(t) individuals may be infected at time t. Infected members of the population may pass on the disease to those who are vulnerable. R(t) represents the sum of those who have contracted the disease at some point (including death cases) and is permanent. The differential equation of the SIR model is

(1)$\frac{dS}{dt}=-\beta SI, \frac{dI}{dt}=\beta SI-\mu I, \frac{dR}{dt}= \mu I$

The β represents the infection rate, and the average infectious period is 1/μ days. The R0 is the primary reproduction number, showing the expected number of individuals infected from a single infection carrier is introduced into a population of N susceptible hosts.

3.2. Machine Learning Models

ML is one of the essential tools for the classification task and is employed to forecast the possible confirmed cases and mortality numbers. Currently, various research has been carried out to forecast the COVID-19 cases using different ML models [9]. From the pool of machine learning (ML) models, we tested various well-known regression and forecasting models such as linear regression [25], polynomial regression [26], and support vector regression [27]. A brief description of each model is given in the subsequent sections.

3.2.1. Linear Regression Model

The linear regression model [28] draws the best fit line across datasets using the independent (features) and dependent variables (labels). Linear regression is a technique that models the interconnection of two variables. This is conducted by applying a linear equation to perceived data points. The first of the two variables are an explanatory variable. The second variable is a dependent variable. Through adjusting a linear equation to the observed data, the framework models the affiliation of the two variables. The equation of the linear regression line is shown below in Equation (2).

(2)$Y=a+bX$

Y is the dependent variable and X is the explanatory variable. The intercept (the value of y when x = 0) is a, and the slope of the line is b.

3.2.2. Polynomial Regression (PR)

A linear regression model gives the best results when the relationship between the data is linear, will not draw a best-fit line, and fails in the case of non-linear data. The PR overcomes this problem, which supports finding the curvilinear association between independent and dependent variables. PR is a modified form of linear regression where the linear regression is transformed into PR due to the non-linear relationship between dependent and independent variables. PR represents the relationship between the independent variable and dependent variable while the dependent variable is compared to the independent variable with an nth degree. The dependent and independent variables do not need to have a linear relationship. If we increase the degree of the independent variable term, the equation will mirror the line. For example, the polynomial regression equation, Equation (3), with an nth degree may be depicted as the following:

(3)$Y=b_0+a_{1}x+a_2{x}^2 +a_3{x}^3 +\cdots +a_n{x}^n$

3.2.3. Support Vector Regression (SVR)

The main objective in most linear regression models is to minimize the sum of squared errors. However, if we are interested in error reduction to some acceptable range, we do not care about the error value. In such cases, the SVR helps us in defining the errors’ acceptable range and finding a fitting line to the inputted data. SVR is an ideal choice for linear and non-linear regression curve fittings and forecasting. SVR is established on the various components of the support vector machine (SVM). Support vectors are marks in an n-dimensional feature space. They precisely disconnect the marks around the hyperplane that is nearer to the generated hyperplane. The resulting equation is

(4)$y=wX+b$

Here, w represents weight, b represents the intercept at X = 0, and epsilon represents the tolerance margin.

3.3. Deep Learning Models

For forecasting infectious disease and time-series data analysis, Deep Learning (DL) such as LSTM, Bi-LSTM, and ARIMA performed well as compared to ML models [29,30,31]. DL models performed better in classification modeling and forecasting than traditional time series models and networks. DL model can automatically learn the temporal and contextual trends from the inputted data. Due to the ubiquitous nature of COVID-19 data, we employed the LTSM and Bi-LSTM DL models, to predict the temporal trend in data and detect the missing patterns, enhancing the forecasting performance.

3.3.1. Long Short-Term Memory (LSTM)

The Artificial Neural Networks (ANNs) can handle the multifaceted associations but cannot detect long-term dependences in time-service data [32], and the prediction performance primarily depends on the dataset features. On the other hand, the Recurrent Neural Network (RNN) model is the one type of ANN that can resolve the handle temporal dependencies using the previously hidden state of data and better predict the short-term time-series data, unable to predict the long-term dependencies. The LSTM [33] model resolves the RNN with sophisticated gated memory in the network for effective sequence prediction. The LSTM is a sophisticated gated memory created to address the vanishing gradient issues that limit a simple RNN. Desirable features of the LSTM include their prolonged ability to detect long-term dependences and to process time-series data. Its hidden units in each layer in the memory block allow the information flow between the input and output ports. The sigmoid function determines which information should be forgotten, saved, and passed the next hidden state within the input and output gates. This model performance is highly reliant on tuning the hyper-parameters for better performance [34]. We implemented the LSTM model using TensorFlow 2.0 of the python package. We used only a set of parameters (units, activation function, return sequence, and dropout) of LSTM from the list of 23 parameters available in the LSTM layer and kept the rest of the parameters with default values. We used 200 epochs with batch size of 64 and verbose of 2, with validation split as 0.2.

3.3.2. Bidirectional LSTM

The extended version of LSTM is Bi-directional LSTM (Bi-LSTM), which saves both directional state (forward and backward) information using two LSTMs models. Forward direction LSTM takes the input in the forwarding direction, and the other in a backward direction, which eventually collects the contextual information within the network and forecast more effectively. To resolve the shortcomings of the LSTM model, which was able to keep the information of the previous step but unable to keep the information of future one. The bidirectional recurrent neural networks have been proposed, containing two LSTM layers for both forward and backward with same output. In Bi-LSTM architecture, for an input sequence, the forward and backward direction weighted vector are calculated, and the final sequence is made by combining both directional vectors as the final output. Commonly accepted RNNs advance inputs in a singular course while disregarding the data acquired ahead of real time. This problem is solved by applying the bidirectional topology of LSTM. The Bi-LSTM removes the entire temporal information taking into account information before and after real-time. Moreover, hidden neurons of the standard RNN are divided into the states of forwarding and backward. This refers to when neurons of the forwarding states are not joined to the backward states and vice versa. Without the backward states, the architecture is similar to the commonly accepted unidirectional RNN. Furthermore, in this framework, more time delays are not necessary as taken in the standard RNN. The same setting was used for Bi-LSTM, as described for the LSTM model.

4. Results

4.1. Evaluation Criteria

In this study, various performance criteria have been used to evaluate the execution of our models, as shown in Table 2.

4.2. Mathematical Models (Susceptible–Infected–Recovered (SIR))

The susceptible population in the SIR model lowers and reaches zero. For COVID-19, Figure 2, Figure 3, Figure 4 and Figure 5 reveal that the SIR model does not mirror the one of the infected cases.

The fitting values of the SIR model with actual data with different performance metrics are reported in Table 3.

4.3. Results of the Machine Learning Models

This subsection trains the algorithms’ linear regression, PR, and SVR for each country individually. For confirmed case predictions and recovered case predictions, two attributes were used: observation date and the cumulative number of confirmed cases or recovered cases. Furthermore, for the prediction of death cases, two attributes, observation date and the cumulative number of death cases in the existing data of each country, were used. As a result, the ML model fits the actual data with different performance metrics, as reported in Table 4, Table 5 and Table 6. The regression results of the confirmed cumulative recovered and cumulative death cases are illustrated in the figures below. Figure 6 shows the fit of the ML models for KSA infected cases, and the polynomial regression model is considered the optimal ML model in fitting the actual data.

Figure 7 shows the fit of the ML models for Bahrain infected cases; the linear regression model is considered the optimal ML model in fitting the actual data.

Figure 8 shows the fit of the ML models for Kuwait infected cases; the polynomial regression model is considered the optimal ML model in fitting the actual data.

Figure 9 shows the fit of the ML models for UAE infected cases; the polynomial regression model is considered the optimal ML model in fitting the actual data.

Figure 10 shows the fit of the ML models for Saudi Arabian recovered cases; the polynomial regression model is considered the optimal ML model in fitting the actual data.

Figure 11 shows the fit of the ML models for Bahrain recovered cases; the polynomial regression model is considered the optimal ML model in fitting the actual data.

Figure 12 shows the fit of the ML models for Kuwait recovered cases; the polynomial regression model is considered the optimal ML model in fitting the actual data.

Figure 13 shows the fit of the ML models for UAE recovered cases; the polynomial regression model is considered the optimal ML model in fitting the actual data.

Figure 14 shows the fit of the ML models for Saudi Arabian death cases; the polynomial regression model is considered the optimal ML model in fitting the actual data.

Figure 15 shows the fit of the ML models for Bahrain death cases; the linear regression model is considered the optimal ML model in fitting the actual data.

Figure 16 shows the fit of the ML models for Kuwait death cases; the polynomial regression model is considered the optimal ML model in fitting the actual data.

Figure 17 shows the fit of the ML models for UAE death cases; the polynomial regression model is considered the optimal ML model in fitting the actual data.

Table 4, Table 5 and Table 6 show performance in terms of RMSE, MAE, and R2 measures based on actual and predicted cases.

4.4. Results of the Deep Learning Models

The following sections provide the results of both LSTM and Bi-LSTM models, and Table 7 describes the values of the different parameters used in both models.

4.4.1. LSTM

The LSTM resulted in fitting the actual data reported in Table 7. The results of the confirmed cumulative cases, death cases, and recovered cases are illustrated in the figures below. Figure 18 shows the fit of the LSTM model for Saudi Arabia’s infected cases.

Figure 19 shows the fit of the LSTM model for Saudi Arabia’s recovered cases.

Figure 20 shows the fit of the LSTM model for Saudi Arabia’s death cases.

The LSTM model fit is shown in Figure 21 for Bahrain’s infected cases.

The LSTM model fit is shown in Figure 22 for Bahrain’s recovered cases.

The LSTM model fit is shown in Figure 23 for Bahrain’s death cases.

The LSTM model fit is shown in Figure 24 for Kuwait’s infected cases.

The LSTM model fit is shown in Figure 25 for Kuwait’s recovered cases.

The LSTM model fit is shown in Figure 26 for Kuwait’s death cases.

The LSTM model fit is shown in Figure 27 for UAE’s infected cases.

The LSTM model fit is shown in Figure 28 for UAE’s recovered cases.

The LSTM model fit is shown in Figure 29 for UAE’s death cases.

4.4.2. Bi-LSTM

The Bi-LSTM resulted in fitting the actual data reported in Table 4. The results of the confirmed cumulative cases, death cases, and recovered cases are illustrated in the figures below. Figure 30 shows the fit of the Bi-LSTM model for Saudi Arabia’s infected cases.

Figure 31 shows the fit of the Bi-LSTM model for Saudi Arabia’s death cases.

Figure 32 shows the fit of the Bi-LSTM model for Saudi Arabia’s death cases.

Figure 33 shows the fit of the Bi-LSTM model for Bahrain’s infected cases.

Figure 34 shows the fit of the Bi-LSTM model for Bahrain’s recovered cases.

Figure 35 shows the fit of the Bi-LSTM model for Bahrain’s death cases.

Figure 36 shows the fit of the Bi-LSTM model for Kuwait’s infected cases.

Figure 37 shows the fit of the Bi-LSTM model for Kuwait’s recovered cases.

Figure 38 shows the fit of the Bi-LSTM model for Kuwait’s death cases.

Figure 39 shows the fit of the Bi-LSTM model for UAE’s infected cases.

Figure 40 shows the fit of the Bi-LSTM model for UAE’s recovered cases.

Figure 41 shows the fit of the Bi-LSTM model for UAE’s death cases.

Table 8, Table 9 and Table 10 compare the performance of DL models (LSTM, Bi-LSTM) in terms of RMSE, MAE, and R² error measures.

4.5. Future Prediction

Our testing shows that the model is somewhat reasonable. Therefore, we can move on to predicting the future or forecasting. Feeding the past n days’ values into the model to predict tomorrow’s value, we obtain tomorrow’s value as output. To find the day after tomorrow’s value, we feed in past n-1 days’ values along with tomorrow’s value and the model outputs the day after tomorrow’s value. Bi-LSTM outperforms as compared to mentioned models. We finally selected the Bi-LSTM model for forecasting confirmed cases, death cases, and recovered cases for KSA.

Figure 42 shows the infected forecasting after 30 days, the blue color represents the actual infected cases, and the red color represents the prediction cases.

Figure 43 shows the death forecasting after 30 days, the green color represents the actual death cases, and the red color represents the prediction cases.

Figure 44 shows the recovered forecasting after 30 days, the blue color represents the actual recovered cases, and the red color represents the prediction cases.

5. Discussion

This section provides a tabular and graphical analysis of the results. The comparison covers the ML and DL algorithms. Among the ML models, prediction errors of performance measures are plotted in bar charts for comparison. Figure 45 displays the polynomial regression to be the best model for predicting the COVID-19 confirmed cases. Compared to other Saudi Arabia models, SIR and SVR were determined to be the least favorable.

Figure 46 displays linear regression to be the most favorable model in Bahrain for predicting COVID-19 confirmed cases. SIR and SVR are the least favorable models.

Figure 47 shows that polynomial regression is the most favorable model to predict COVID-19 confirmed cases compared to the other ML models in Kuwait. Conversely, SIR and linear regression are the least favorable models.

Figure 48 reveals polynomial regression as the most favorable model for forecasting COVID-19 confirmed cases in the UAE. In contrast, SIR and linear regression are the least favorable models.

The traditional mathematical model and ML models show different behavior for target countries due to data variability. However, every country applied its policies to counteract the spread of the diseases. In the case of Bahrain, linear regression gives the best results due to linear relationship independent and dependent variables, while polynomial regression gives the best results because we found the non-linear relationship within the dependent and independent variables. So, both traditional mathematical and ML models are not the best choice due to the nature of the current COVID-19 data.

Bi-LSTM performs quite well among DL models, and its predicted values entirely overlap several actual cases. Therefore, the performance of Bi-LSTM denotes it as the best model with the highest accuracy. Comparative analysis of the performance metrics for DL models is provided below. Figure 49 displays the BILSTM model, represented by blue, to be the most advantageous in predicting COVID-19 confirmed cases in Saudi Arabia compared to the LSTM model, represented by orange.

Figure 50 showcases the Bi-LSTM model, represented by blue, as the best model for predicting COVID-19 confirmed cases in Bahrain compared to the LSTM model, represented by the orange color.

Figure 51 shows the Bi-LSTM model, represented by blue, as the best model for forecasting COVID-19 confirmed cases in Kuwait compared to the LSTM model, represented by the orange color.

Figure 52 shows that the Bi-LSTM model, represented by blue, is the best model for forecasting COVID-19 confirmed cases in UAE compared to the LSTM model, represented by the orange color.

6. Conclusions

The advancement of algorithms is necessary to enhance the prediction accuracy of COVID-19 and minimize the virus’s effect on the health and life of many people around the world. This paper highlights the mechanisms of the coronavirus forecasts, namely, the ML and DL approach. The united employment of informal datasets regarding the circulation and transmission of COVID-19 in Saudi Arabia, the Kingdom of Bahrain, Kuwait, and the UAE. Additionally, the following algorithms were validated, and the optimal model was selected based on error rate in predictions: linear regression, PR, SVR, SIR, LSTM, and Bi-LSTM. The proposed models should be easy to perform and provide reliable conclusions in real time. The DL models have demonstrated crucial enhancements when dealing with time-series data in various applications. The DL methods promise time-series forecasting, such as the self-regulating comprehension of temporal dependence, and the administration of temporal complexes Bi-LSTM outperforms among DL models, and its forecast values entirely overlap significant numbers of actual cases. Therefore, it performs with the highest accuracy and is most recommended for disease forecasting. We will perform more analysis in the future, such as identifying the peak value of COVID-19 cases against each country and identifying the cause of that peak. We will check the model behavior and performance by splitting the data into chunks (peak vs. non-peak) cases. Therefore, we can get an idea of which model performs best in these two cases. This will allow us to design a framework for the best model selection. The peak value analysis also helps us to derive new features by casual analysis, which can be used for forecasting.

Author Contributions

Conceptualization, H.F.A. and J.H.; Data curation, H.K. and J.H.; Formal analysis, H.K.; Investigation, H.F.A.; Supervision, H.F.A. and J.H.; Validation, H.F.A. and J.H.; Visualization, H.K.; Writing—original draft, H.K. and Z.A.; Writing—review and editing, H.F.A., Z.A., A.A. and J.H. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

We used the Novel Corona Virus 2019 Dataset [11] in this study.

Conflicts of Interest

The authors declare no conflict of interest.

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Figures and Tables

Figure 1. The proposed framework for COVID-19 forecasting.

Figure 2. The fit of the SIR model of infected cases for KSA.

Figure 3. The fit of the SIR model of infected cases for Bahrain.

Figure 4. The fit of the SIR model of infected cases for Kuwait.

Figure 5. The fit of the SIR model of infected cases for UAE.

Figure 6. The ML forecasting for KSA Confirmed cases (Saudi Arabia).

Figure 7. The ML forecasting for confirmed cases (BAHRAIN).

Figure 8. The ML forecasting for confirmed cases (KUWAIT).

Figure 9. The ML forecasting for confirmed cases (UAE).

Figure 10. The ML forecasting for recovered cases (Saudi Arabia).

Figure 11. The ML forecasting for recovered cases (Bahrain).

Figure 12. The ML forecasting for recovered cases (Kuwait).

Figure 13. The ML forecasting for recovered cases (UAE).

Figure 14. The ML forecasting for death cases (Saudi Arabia).

Figure 15. The ML forecasting for death cases (Bahrain).

Figure 16. The ML forecasting for death cases (Kuwait).

Figure 17. The ML forecasting for death cases (UAE).

Figure 18. Infected cases (Saudi Arabia).

Figure 19. Recovered cases (Saudi Arabia).

Figure 20. Death cases (Saudi Arabia).

Figure 21. Infected cases (Bahrain).

Figure 22. Recovered cases (Bahrain).

Figure 23. Death cases (Bahrain).

Figure 24. Infected cases (Kuwait).

Figure 25. Recovered cases (Kuwait).

Figure 26. Death cases (Kuwait).

Figure 27. Infected cases (UAE).

Figure 28. Recovered cases (UAE).

Figure 29. Death cases (UAE).

Figure 30. Infected cases (Saudi Arabia).

Figure 31. Recovered cases (Saudi Arabia).

Figure 32. Death cases (Saudi Arabia).

Figure 33. Infected cases (Bahrain).

Figure 34. Recovered cases (Bahrain).

Figure 35. Death cases (Bahrain).

Figure 36. Infected cases (Kuwait).

Figure 37. Recovered cases (Kuwait).

Figure 38. Death cases (Kuwait).

Figure 39. Infected cases (UAE).

Figure 40. Recovered cases (UAE).

Figure 41. Death cases (UAE).

Figure 42. Infected forecasting after 30 days.

Figure 43. Death forecasting after 30 days.

Figure 44. Recovered forecasting after 30 days.

Figure 45. Comparison of ML models in terms of error measures (Saudi Arabia) in the terms of RMSE, MAE, and R2.

Figure 46. Comparison of ML models in terms of error measures (Bahrain) in the terms of RMSE, MAE, and R2.

Figure 47. Comparison of ML models in terms of different error measures (Kuwait) in the terms of RMSE, MAE, and R2.

Figure 48. Comparison of ML models in terms of error measures (UAE) in the terms of RMSE, MAE, and R2.

Figure 49. Comparison of DL models in terms of error measures (Saudi Arabia) in the terms of RMSE, MAE, and R2.

Figure 50. Comparison of DL models in terms of error measures (Bahrain) in the terms of RMSE, MAE, and R2.

Figure 51. Comparison of DL models in terms of error measures (Kuwait) in the terms of RMSE, MAE, and R2.

Figure 52. Comparison of DL models in terms of different error measures (UAE) in the terms of RMSE, MAE, and R2.

Table 1

Comparison of existing related mathematical modeling, ML, and DL techniques for COVID-19 case forecasting.

Paper	Model and Techniques	Pros	Cons
Balcan, Duygu et al. [8]	GLEaM model	GLEaM has been proved to be very flexible; the scalability of GLEaM also has the advantage of making possible the use of statistical methods to fit epidemic parameters which are usually not known in the case of new emerging diseases. Provide an accurate description of the spreading pattern.	It required high-performance computing.
Chunara, Rumi et al. [17]	Disclosures fromTwitter, HealthMap, and Official reports of cholera cases	Unofficial data and official data used together for timely estimates of epidemic dynamics as a cost-effective source. The sources of the statistics mentioned are made publically accessible in real time Assessments from these sources can be calculated before official sources,	The third peak of the case was not apprehended by the unofficial sources. This may have been due to local disease oscillations without the initial gain or loss of media attention. Additional investigation is required to access unofficial news sources to be an effective measure of mortality in other epidemics
Liu, Qinghe et al. [22]	SEIR model	The SEIR model has the strength of being more biologically realistic.	SEIR model has the drawback of having two additional unknown parameters: the latent period and the initial latent population. The analysis is based on that the official statistics are accurate in a unified environment. However, there may be deviations in the norms of statistical data in various countries, which will also lead to errors between conclusions and real situations.
Minter, Amanda, and Renata Retkute. [19]	ABC Algorithms	ABC algorithms may be selected for coarse or complex data. It is a vigorous framework offers flexibility with its application The algorithms can be applied to forecast viral spread trends when used by users in a time-efficient approach.	The ABC framework often results in an inefficient sampling of the parameter space.
Pujari, Bhalchandra S., and Snehal M. Shekatkar. [20]	Hybrid approach (SIR and transportation network)	The model is also capable of simultaneously handling cities with different populations. The model allows for quick and accurate COVID-19 predictions.	Despite the comprehensive usage of the SIR model, demographic specifics and spatial heterogeneity cannot be investigated. The network-based models are computationally inefficient
Volkova, Svitlana et al. [23]	Neural Network model(historical and social networks data)	It generate accurate and vigorous forecasts. Assists in overriding an accepted 2-week lag-time of Centers for Disease Control (CDC) case counts.	They need more tweets per geolocation to improve the model performance in terms of lower error rate and the Pearson score increases the number of locations associated with a higher number of tweets.
Chae, Sangwon, Sungjun Kwon, and Donghyun Lee [21]	Various deep learning techniques: DNN LSTM ARIMA	DNN and LSTM models perform better than ARIMA. The DNN model performed stably, and the LSTM model revealed more authentic results during high viral transmission rates. Deep learning methods are superior in their scalability. Deep learning methods are exemplary analytical approaches for increasing and decreasing transmission rates effectively well.	The ARIMA model was adequate if the number of infectious diseases cases was regular and had no increasing or decreasing trends.

Table 2

Evaluation criteria.

Criterion	Formula
Root mean square error	$$\mathrm{RMSE}=\sqrt{{{\displaystyle \sum }}_{}^{}\frac{{\left(y_{pred}-y_{actual}\right)}^2}{N}}$$
Mean Absolute Error	$$\mathrm{MAE}=\frac{1}{n}{{\displaystyle \sum }}_{i=1}^n\left|yi-\widehat{yi} \right|$$
Mean Squared Error	$$\mathrm{MSE}=\frac{1}{n}{{\displaystyle \sum }}_{i=1}^n{\left(yi-\widehat{yi}\right)}^2$$
R-squared	$${\mathrm{R}}^2=1-\frac{\sum \left(yi-\widehat{yi}\right){}^2}{\sum \left(yi-\underset{¯}{yi}\right){}^2}$$

Table 3

Performance evaluation of SIR model.

Countries	RMSE	MAE	R2
Saudi Arabia	170,033.74	135,132.88	−0.45
Bahrain	56,518.49	44,547.75	−1.64
Kuwait	44,103.03	4969.99	0.35
UAE	102,177.12	74,681.14	−1.15

Table 4

Comparison of machine learning models (linear regression, PL, SVR) in terms of different error measures.

Countries	Models	Confirmed Cases
		RMSE	MAE	R2
Saudi Arabia	Liener Regression	48,514.63	48,512.25	−103.62
	PR	3961.99	2754.49	0.30
	SVR	119,299.83	117,811.78	−631.6
Bahrain	Liener Regression	6307.73	6299.25	−17.55
	PR	13,489.85	11,589.41	−83.85
	SVR	54,273.51	54,779.55	−1398.2
Kuwait	Liener Regression	7842.52	7826.75	−10.24
	PR	1350.10	1274.91	0.66
	SVR	88,824.25	88,037.89	−1441.0
UAE	Liener Regression	61,598.57	60,376.96	−15.23
	PR	5095.28	4553.26	0.88
	SVR	56,183.45	56,085.20	−12.50

Table 5

Comparison of machine learning models (linear regression, PL, SVR) in different error measures.

Countries	Models	Death Cases
		RMSE	MAE	R2
Saudi Arabia	Liener Regression	709.106	702.039	−803.79
	PR	551.60	467.63	−485.99
	SVR	4014.08	3970.29	−25,788.01
Bahrain	Liener Regression	33.25	32.97	−130.4
	PR	114.30	97.44	−1551.54
	SVR	273.28	270.84	−8874.03
Kuwait	Liener Regression	62.36	60.98	−221.2
	PR	71.18	59.38	−288.5
	SVR	495.07	489.64	−14,003.1
UAE	Liener Regression	49.42	47.21	−2.83
	PR	18.25	13.19	0.47
	SVR	144.48	143.46	−31.76

Table 6

Comparison of machine learning models (linear regression, PL, SVR) in terms of different error measures.

Countries	Models	Recovered Cases
		RMSE	MAE	R2
Saudi Arabia	Liener Regression	83,255.07	82,986.14	-11,066.8
	PR	8083.33	6329.03	-103.33
	SVR	218,510.45	216,320.95	-76,239.4
Bahrain	Liener Regression	-130.4	97,633.80	97,618.07
	PR	−1551.54	20,214.07	17,260.02
	SVR	−8874.03	59,873.90	59,318.44
Kuwait	Liener Regression	152,385.09	152,360.41	−132,6795452.6
	PR	4302.93	3504.65	−8.71
	SVR	90,132.26	89,181.64	−4260.35
UAE	Liener Regression	34,702.00	34,291.25	−24.12
	PR	7770.93	6535.61	−0.25
	SVR	23,674.16	23,442.58	−10.69

Table 7

DL models with parameters and their values.

Models	Parameters	Values
LSTM/Bi-LSTM	Layers	3
	Optimizer	Adam
	Batch size	64
	Epochs	200
	Timestep	10

Table 8

Comparison of error measures of the DL models (LSTM, Bi-LSTM) for confirmed cases.

Countries	Models	Confirmed Cases
		RMSE	MAE	R2
Saudi Arabia	LSTM	0.034	0.033	−3.38
	Bi-LSTM	0.026	0.026	−1.69
Bahrain	LSTM	0.051	0.050	−0.05
	Bi-LSTM	0.051	0.050	−0.05
Kuwait	LSTM	0.012	0.011	0.96
	Bi-LSTM	0.031	0.028	0.76
UAE	LSTM	0.205	0.193	−0.96
	Bi-LSTM	0.023	0.0176	0.97

Table 9

Comparison of error measures of the DL models (LSTM, Bi-LSTM) for death cases.

Countries	Models	Death Cases
		RMSE	MAE	R2
Saudi Arabia	LSTM	0.02	0.02	0.87
	Bi-LSTM	0.018	0.017	0.87
Bahrain	LSTM	0.05	0.05	−1.01
	Bi-LSTM	0.05	0.05	−1.01
Kuwait	LSTM	0.004	0.003	0.99
	Bi-LSTM	0.023	0.02	0.87
UAE	LSTM	0.264	0.25	−4.77
	Bi-LSTM	0.0148	0.0125	0.98

Table 10

Comparison of error measures of the DL models (LSTM, Bi-LSTM) for recovered cases.

Countries	Models	Recovered Cases
		RMSE	MAE	R2
Saudi Arabia	LSTM	0.02	0.02	0.21
	BI-LSTM	0.02	0.02	0.21
Bahrain	LSTM	0.045	0.044	0.24
	BI-LSTM	0.045	0.044	0.24
Kuwait	LSTM	0.007	0.005	0.99
	BI-LSTM	0.019	0.016	0.93
UAE	LSTM	0.194	0.17	−1.17
	BI-LSTM	0.008	0.006	0.99

References

1. COVID Live Update: 222,777,304 Cases and 4,600,073 Deaths from the Coronavirus—Worldometer. Available online: https://www.worldometers.info/coronavirus/ (accessed on 8 September 2021).

2. Cooper, I.; Mondal, A.; Antonopoulos, C.G. A SIR Model Assumption for the Spread of COVID-19 in Different Communities. Chaos Solitons Fractals; 2020; 139, 110057. [DOI: https://dx.doi.org/10.1016/j.chaos.2020.110057]

3. Toda, A.A. Susceptible-Infected-Recovered (SIR) Dynamics of COVID-19 and Economic Impact. arXiv; 2020; arXiv: 2003.11221

4. Shah, S.T.A.; Mansoor, M.; Mirza, A.F. Predicting COVID-19 Spread in Pakistan Using the SIR Model. J. Pure Appl. Microbiol.; 2020; 14, pp. 1423-1430. [DOI: https://dx.doi.org/10.22207/JPAM.14.2.40]

5. Choi, B.C.K.; Pak, A.W.P. A Simple Approximate Mathematical Model to Predict the Number of Severe Acute Respiratory Syndrome Cases and Deaths. J. Epidemiol. Community Health; 2003; 57, pp. 831-835. [DOI: https://dx.doi.org/10.1136/jech.57.10.831]

6. Peng, L.; Yang, W.; Zhang, D.; Zhuge, C.; Hong, L. Epidemic Analysis of COVID-19 in China by Dynamical Modeling. medRxiv; 2020; pp. 22-24. [DOI: https://dx.doi.org/10.1101/2020.02.16.20023465]

7. Punn, N.S.; Sonbhadra, S.K.; Agarwal, S. COVID-19 Epidemic Analysis Using Machine Learning and Deep Learning Algorithms. medRxiv; 2020; 1, pp. 132-145. [DOI: https://dx.doi.org/10.1101/2020.04.08.20057679]

8. Balcan, D.; Gonçalves, B.; Hu, H.; Ramasco, J.J.; Colizza, V.; Vespignani, A. Modeling the Spatial Spread of Infectious Diseases: The Global Epidemic and Mobility Computational Model. J. Comput. Sci.; 2010; 1, pp. 132-145. [DOI: https://dx.doi.org/10.1016/j.jocs.2010.07.002]

9. Kwekha-Rashid, A.S.; Abduljabbar, H.N.; Alhayani, B. Coronavirus Disease (COVID-19) Cases Analysis Using Machine-Learning Applications. Appl. Nanosci.; 2021; pp. 1-13. [DOI: https://dx.doi.org/10.1007/s13204-021-01868-7]

10. Liu, Z.; Magal, P.; Webb, G. Predicting the Number of Reported and Unreported Cases for the COVID-19 Epidemics in China, South Korea, Italy, France, Germany and United Kingdom. J. Theor. Biol.; 2021; 509, 110501. [DOI: https://dx.doi.org/10.1016/j.jtbi.2020.110501]

11. Weber, A.; Ianelli, F.; Goncalves, S. Trend Analysis of the COVID-19 Pandemic in China and the Rest of the World. medRxiv; 2020; [DOI: https://dx.doi.org/10.1101/2020.03.19.20037192]

12. Botha, A.E.; Dednam, W. A Simple Iterative Map Forecast of the COVID-19 Pandemic. arXiv; 2020; arXiv: 2003.10532

13. Wang, H.; Zhang, Y.; Lu, S.; Wang, S. Tracking and Forecasting Milepost Moments of the Epidemic in the Early-Outbreak: Framework and Applications to the COVID-19. F1000Research; 2020; 9, 333. [DOI: https://dx.doi.org/10.12688/f1000research.23107.2]

14. Sameni, R. Mathematical Modeling of Epidemic Diseases; A Case Study of the COVID-19 Coronavirus. arXiv; 2020; arXiv: 2003.11371

15. Gao, Y.; Zhang, Z.; Yao, W.; Ying, Q.; Long, C.; Fu, X. Forecasting the cumulative number of COVID-19 deaths in China: A Boltzmann function-based modeling study. Infect. Control Hosp. Epidemiol.; 2020; 41, pp. 841-843. [DOI: https://dx.doi.org/10.1017/ice.2020.101] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32238216]

16. Dowd, J.B.; Andriano, L.; Brazel, D.M.; Rotondi, V.; Block, P.; Ding, X.; Liu, Y.; Mills, M.C. Demographic Science Aids in Understanding the Spread and Fatality Rates of COVID-19. Proc. Natl. Acad. Sci. USA; 2020; 117, pp. 9696-9698. [DOI: https://dx.doi.org/10.1073/pnas.2004911117]

17. Chunara, R.; Andrews, J.R.; Brownstein, J.S. Social and News Media Enable Estimation of Epidemiological Patterns Early in the 2010 Haitian Cholera Outbreak. Am. J. Trop. Med. Hyg.; 2012; 8, pp. 39-45. [DOI: https://dx.doi.org/10.4269/ajtmh.2012.11-0597]

18. Kass-Hout, T.A.; Alhinnawi, H. Social Media in Public Health. Br. Med. Bull.; 2017; 108, pp. 5-24. [DOI: https://dx.doi.org/10.1093/bmb/ldt028] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/24103335]

19. Elsevier Enhanced Reader. Approximate Bayesian Computation for Infectious Disease Modelling. Available online: https://reader.elsvier.com/reader/sd/pii/S175543651930026X?tken=56814D8EF7975027F99E5E5D06F393367D6542C07858DBABF920908D2DFC6B149A039B177850A0D0A2717DDD247D5B04&originRegion=us-east-1&originCreation=20210908052126 (accessed on 8 September 2021).

20. Pujari, B.S.; Shekatkar, S. Multi-City Modeling of Epidemics Using Spatial Networks: Application to 2019-nCov (COVID-19) Coronavirus in India. medRxiv; 2020; [DOI: https://dx.doi.org/10.1101/2020.03.13.20035386]

21. Chae, S.; Kwon, S.; Lee, D. Predicting Infectious Disease Using Deep Learning and Big Data. Int. J. Environ. Res. Public Health; 2018; 15, 1596. [DOI: https://dx.doi.org/10.3390/ijerph15081596] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30060525]

22. Liu, Q.; Liu, Z.; Zhu, J.; Zhu, Y.; Li, D.; Gao, Z.; Zhou, L.; Tang, Y.; Zhang, X.; Yang, J. et al. Assessing the Global Tendency of COVID-19 Outbreak. medRxiv; 2020; [DOI: https://dx.doi.org/10.1101/2020.03.18.20038224]

23. Volkova, S.; Ayton, E.; Porterfield, K.; Corley, C.D. Forecasting Influenza-Like Illness Dynamics for Military Populations Using Neural Networks and Social Media. PLoS ONE; 2017; 12, e0188941. [DOI: https://dx.doi.org/10.1371/JOURNAL.PONE.0188941]

24. Kaggle. Novel Corona Virus 2019 Dataset. Available online: https://www.kaggle.com/sudalairajkumar/novel-corona-virus-2019-dataset (accessed on 8 September 2021).

25. Yao, W.; Li, L. A New Regression Model: Modal Linear Regression. Scand. J. Stat.; 2014; 41, pp. 656-671. [DOI: https://dx.doi.org/10.1111/sjos.12054]

26. Ostertagová, E. Modelling Using Polynomial Regression. Procedia Eng.; 2012; 48, pp. 500-506. [DOI: https://dx.doi.org/10.1016/j.proeng.2012.09.545]

27. Parbat, D.; Chakraborty, M. A Python Based Support Vector Regression Model for Prediction of COVID19 Cases in India. Chaos Solitons Fractals; 2020; 138, 109942. [DOI: https://dx.doi.org/10.1016/j.chaos.2020.109942]

28. Yan, X.; Gang Su, X. Linear Regression Analysis: Theory and Computing; World Scientific: Kansas City, MO, USA, 2009; pp. 1-328. [DOI: https://dx.doi.org/10.1142/6986]

29. Tian, Y.; Luthra, I.; Zhang, X. Forecasting COVID-19 Cases Using Machine Learning Models. medRxiv; 2020; [DOI: https://dx.doi.org/10.1101/2020.07.02.20145474]

30. Nazmoon Nabi, K.; Toki Tahmid, M.; Rafi, A.; Ehsanul Kader, M.; Asif Haider, M. Forecasting COVID-19 Cases: A Comparative Analysis between Recurrent and Convolutional Neural Networks. Results Phys.; 2021; 24, 104137. [DOI: https://dx.doi.org/10.1016/j.rinp.2021.104137]

31. Kırbaş, İ.; Sözen, A.; Tuncer, A.D.; Kazancıoğlu, F.Ş. Comparative Analysis and Forecasting of COVID-19 Cases in Various European Countries with ARIMA, NARNN and LSTM Approaches. Chaos Solitons Fractals; 2020; 138, 110015. [DOI: https://dx.doi.org/10.1016/j.chaos.2020.110015]

32. Devaraj, J.; Madurai Elavarasan, R.; Pugazhendhi, R.; Shafiullah, G.M.; Ganesan, S.; Jeysree, A.K.; Khan, I.A.; Hossain, E. Forecasting of COVID-19 Cases Using Deep Learning Models: Is It Reliable and Practically Significant?. Results Phys.; 2021; 21, 103817. [DOI: https://dx.doi.org/10.1016/j.rinp.2021.103817]

33. Zeroual, A.; Harrou, F.; Dairi, A.; Sun, Y. Deep Learning Methods for Forecasting COVID-19 Time-SERIES Data: A Comparative Study. Chaos Solitons Fractals; 2020; 140, 110121. [DOI: https://dx.doi.org/10.1016/j.chaos.2020.110121] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32834633]

34. Yadav, A.; Jha, C.K.; Sharan, A. Optimizing LSTM for Time Series Prediction in Indian Stock Market. Procedia Comput. Sci.; 2020; 167, pp. 2091-2100. [DOI: https://dx.doi.org/10.1016/j.procs.2020.03.257]