Deep learning model for forecasting COVID-19 outbreak in Egypt

Abstract

The World Health Organization has declared COVID-19 as a global pandemic in early 2020. A comprehensive understanding of the epidemiological characteristics of this virus is crucial to limit its spreading. Therefore, this research applies artificial intelligence-based models to predict the prevalence of the COVID-19 outbreak in Egypt. These models are long short-term memory network (LSTM), convolutional neural network, and multilayer perceptron neural network. They are trained and validated using the dataset records from 14 February 2020 to 15 August 2020. The results of the models are evaluated using the determination coefficient and root mean square error. The LSTM model exhibits the best performance in forecasting the cumulative infections for one week and one month ahead. Finally, the LSTM model with the optimal parameter values is applied to forecast the spread of this epidemic for one month ahead using the data from 14 February 2020 to 30 June 2021. The total size of infections, recoveries, and deaths is estimated to be 285,939, 234,747, and 17,251 cases on 31 July 2021. This study could assist the decision-makers in developing and monitoring policies to confront this disease.

1. Introduction

COVID-19 is a novel coronavirus strain, where the letters CO, VI, and D stand for corona, virus, and disease, respectively (WHO, 2020a). This disease is characterized by vigorous progression and high infection risk (Guo et al., 2020). It is transmitted through respiratory droplets (sneezing and coughing) of infected persons and contaminated surfaces (WHO, 2020b). Besides, it could be transmitted from diseased carriers without any symptoms to contacted persons (Cao et al., 2020). This virus is infectious because its incubation period ranges between one day and two weeks. Additionally, it can persist on surfaces for up to nine days at room temperature, accelerating its transmission (van Doremalen et al., 2020). Therefore, this infectious disease is a critical topic for medical society (Saba and Elsheikh, 2020).

The global epidemic disease COVID-19 is first reported publicly from Wuhan, China in late December 2019 (WHO, 2020c). It has spread around the world, and the World Health Organization (WHO) declared it a pandemic in March 2020 (WHO, 2020d). By 2 July 2021, the confirmed infected cases reached 182,319,261 cases, including 3,954,324 deaths worldwide (WHO, 2021). Therefore, this virus is considered the most critical crisis since the Second World War (Boccaletti et al., 2020). The observed cases are based on the testing results, so the number of actual infections is higher than that of reported infections. It is pertinent to mention that under-testing leads to errors in the forecasting model (Farooq and Bazaz, 2021). As a result, predicting the number of infected persons is critical (Saba and Elsheikh, 2020).

The coronaviruses have been identified as severe pathogens of respiratory diseases (Cascella et al., 2020). COVID-19 manifests itself in a variety of ways, including fever, cough, fatigue, and breathing problems (Guo et al., 2020). The virus causes pneumonia similar to that caused by severe acute respiratory syndrome coronavirus and Middle East respiratory syndrome coronavirus (Li et al., 2020a). Besides, it can cause organ failure, leading to people physiological deterioration and death (Gibson et al., 2020; Jiang et al., 2020). It may also lead to severe complications for the elderly and patients with comorbidities (Li et al., 2020b, c; Tian et al., 2020; Zhang et al., 2020a). This virus has a mortality rate of about 2–3 % (Rodriguez-Morales et al., 2020).

Egypt declared the first infection and death on 14 February 2020 and 20 March 2020, respectively (Amar et al., 2020; Saba and Elsheikh, 2020). By 2 July 2021, Egypt has confirmed nearly 281,722 cases, including 16,215 deaths (WHO, 2021). Egypt ranked 9th with respect to the infection fatality ratio. Second, concerning the recovery rate, the ranking is 187th, and the ratio is equivalent to 75.5 % compared to countries and regions worldwide. Third, concerning the total infections per million people, the ranking is 164th (2702 cases/million). Finally, among 222 regions and nations around the world, the number of infected people ranks 65th (Egyptian ministry of health and population, 2021).

The Egyptian government has designed and implemented strategies to control and minimize the impact of this pandemic on human health and economy. It enforced a lockdown from mid-March until the end of June 2020, when most industrial, commercial, and transportation activities were banned. Furthermore, other industries were operating with low productivity to minimize contact between workers. An exception was the businesses required to provide basic needs and operations (Abou El-Magd and Zanaty, 2020). The government declared an emergency in some ports due to the COVID-19 contagion. Besides, it delegated isolation departments in hospitals to deal with the infected cases (Amar et al., 2020).

The government launched several campaigns to raise public awareness about the global epidemiological situation (UN Information Centre in Cairo, 2020). In middle and low-income countries, community awareness plays a substantial role in preventing the infection spread because health systems cannot respond efficiently to outbreaks (Egypt today staff, 2020). Parallel to these efforts, a free-of-charge vaccination is being provided to all the citizens, with priority given to medical personnel and individuals with chronic diseases. A total of 4,282,641 doses have been provided, with only 780,460 people have been fully vaccinated (i.e., 0.8 % of the total population) till 28 June 2021 (Our World in Data, 2021a). In this regard, the government has made agreements to receive vaccinations from different sources in order to meet its immunization needs (Egyptian streets, 2021).

It is crucial to develop accurate outbreak prediction models to explore how this disease is spreading. These models open the way for governments and other legislative bodies to suggest new policies and assess their effectiveness (Remuzzi and Remuzzi, 2020). The global COVID-19 pandemic exhibits a different behavior from other recent outbreaks, which proves that the ability of standard models to deliver reliable results is questionable (Ardabili et al., 2020; Ivanov, 2020; Koolhof et al., 2020). The model uncertainty is further attributed to the numerous variables involved in the spread, population behavior complexity in various regions, and differences in containment strategies (Darwish et al., 2020).

2. Related work

Many scholars have studied the COVID-19 spread in Egypt from various perspectives: a) impact of partial and complete lockdown on the air quality and health environment (Abou El-Magd and Zanaty, 2020), b) public awareness and attitude towards the COVID-19 (Abdelhafiz et al., 2020), c) psychological impact of COVID-19 on healthcare professionals and the public (Elkholy et al., 2020; AboKresha et al., 2021; Arafa et al., 2021), d) influence of COVID-19 pandemic on education in medical schools (Shehata et al., 2020), e) skeleton of the health services post COVID-19 (Kamel, 2020). This research is concerned with applying time-series analysis to estimate the number of COVID-19 cases. It provides insights into the spread pattern of this disease to determine the best policies. Besides, it is critical for establishing resilience against this pandemic. Resilience provides a paradigm for orchestrating a more effective response to COVID-19 and eventually altering trajectories to reduce the likelihood of future global health crises (Bragatto et al., 2021).

Time series analysis could be described as monitoring a particular phenomenon over time and forecasting its future response. The time-series data could be forecasted using several models. These models could be categorized into three classes (Khan et al., 2020): a) statistical models, b) machine learning models, and c) deep learning models. Numerous research efforts have been exerted to model the spread of this novel coronavirus. For instance, Amar et al. (2020) applied seven regression models (i.e., exponential, logit growth, quadratic, third-degree, fourth-degree, fifth-degree, and sixth-degree) to forecast the number of infected patients in Egypt. The official database from 15 February till 15 June 2020 was used for training the models. These models predicted the COVID-19 outbreak for the next 15 days, one month, and the final epidemic size and time. The results confirmed that the exponential, fourth-degree, fifth-degree, and sixth-degree models yielded the best prediction results for the next 15 days. Meanwhile, the fourth-degree model proved its prediction capability for one month. Using the logit growth regression model, the epidemic peak and final time were estimated to be on 22 June 2020 and 8 September 2020, respectively. Besides, it was predicted that the total epidemic size would be 166,760 cases. However, the reported results were subjected to uncertainty from social and natural (climatic) conditions.

Ardabili et al. (2020) compared the ability of machine learning models to forecast the COVID-19 outbreak in Italy, China, Iran, Germany, and the United States. The results of the multilayer perceptron (MLP) and adaptive neuro-fuzzy inference system models were promising. The study affirmed the outperformance of machine learning models in modeling the COVID-19 outbreak. Moreover, it recommended modeling the mortality rate to estimate the required beds in intensive care units. Finally, it suggested integrating machine learning and susceptible-exposed-infectious-removed (SEIR) models to enhance the accuracy and lead time of standard epidemiological models. Gupta et al. (2020) employed a support vector machine (SVM), prophet forecasting model, and linear regression model for predicting the active, death, and cured rates in India. The prophet forecasting model produced better results compared to SVM and linear regression models because of the following points: 1) it analyzed the dataset trend to determine the growth curve, 2) it used the Fourier transformation series for data analysis based on daily, weekly, or yearly basis, and 3) it automatically identified the change points in data unlike the SVM and linear regression models. However, SVM provided more significant results in comparison to the linear regression model. This is because the linear regression model did not account for the change of data over time parameter.

Saba and Elsheikh (2020) forecasted the spread of COVID-19 in Egypt using nonlinear autoregressive artificial neural networks (NARANN) and autoregressive integrated moving average (ARIMA). The models were trained using the reported dataset from 1 March till 1 May 2020. The models were evaluated using determination coefficient, mean absolute error, root mean square error, deviation ratio, and coefficient of residual mass. The statistical indicators confirmed that the NARANN model had a better prediction ability compared with the ARIMA model. Muhammad et al. (2021) implemented several supervised machine learning algorithms for modeling COVID-19 infection in Mexico. These models included SVM, logistic regression, decision tree, naive Bayes, and artificial neural network. The relationship between features of the dataset was examined using correlation coefficient analysis. The results confirmed that the decision tree model had the highest accuracy of 94.99 %. In comparison, the naive Bayes model had the highest specificity of 94.30 %, and the SVM model had the highest sensitivity of 93.34 %.

Based on the above-referenced literature, classical statistical models fail to capture the nonlinear and complex behavior in the assessed data (Tollenaar and Van der Heijden, 2013). On the contrary, machine learning algorithms could learn the general trend and intricacies of the data, providing high-quality predictive models (Lawson, 2013; Wong et al., 2019; Hui et al., 2020). Therefore, the machine learning models outperform traditional statistical models (Makridakis et al., 2018). However, the machine learning models could not provide reliable results in modeling the complex behavior of COVID-19 data. The reason is that the number of reported cases varies based on multiple factors, including environmental conditions and physical proximity (Khan et al., 2020).

Deep learning models have achieved outperformance in various classification, regression, and time-series analyses (Khan et al., 2020). Farooq and Bazaz (2021) developed a deep neural network model for modeling the COVID-19 spread in India’s five worst-affected states. The model parameters were updated using an adaptive online incremental learning technique. This technique eliminated the need to rebuild the model from scratch when receiving a new dataset. Therefore, it can overcome the computational complexities arising from data size, run-time duration, transmission dynamics change, and change in infection control mechanisms or government policies. The results affirmed that the reported cases would significantly increase in the coming weeks without flattening the infection curves during this period. This necessitates that policymakers should take appropriate decisions to combat the disease spread. Jin et al. (2021) investigated the spatial-temporal features of the COVID-19 growth in mainland China and Hubei Province. Additionally, a correlation study of the parameters affecting the pandemic severity was carried out. The daily fluctuations in the number of confirmed cases were modeled using logical growth, polynomial, and fully connected neural network models. The experimental findings affirmed the outperformance of the polynomial model (degree = 18) for fitting the daily number of confirmed cases. Furthermore, the deep learning models have exhibited outstanding performance in detecting COVID-19 cases based on chest radiology data. For instance, Aishwarya and Kumar (2021) reviewed and analyzed the application of various machine and deep learning techniques to test the presence of COVID-19 using chest X-rays and computed tomography images. For the deep learning models, it was found that the CheXNet network architecture yielded the most accurate detection results. Furthermore, support vector regression and stacking-ensemble learning models were determined to be the best machine learning models. The proposed methodology could aid doctors and physicians in detecting COVID-19 instances in a short time.

The previous COVID-19 outbreak prediction models demonstrated sound results. However, they have not tackled the following aspects: 1) predicting the total number of infections, recoveries, and deaths in Egypt, 2) examining the application of deep learning models to forecast the epidemic size in Egypt, and 3) comparing the performance of machine learning approaches and deep learning techniques using the official reported data till June 2021. In order to address the aforementioned limitations, this research employs two deep learning algorithms to predict the COVID-19 outbreak in Egypt. These algorithms are the long short-term memory (LSTM) and convolutional neural network (CNN). The performance of these models is evaluated against that of the MLP neural network. These forecasting models are trained and tested using the data reported from 14 February to 15 August 2020. These models are applied to predict the cumulative infected cases for the next seven and thirty days. Different statistical indices are applied to decide the optimal parameters of the best forecasting model. The proposed model with the optimal parameter values is trained using the data reported from 14 February 2020 to 30 June 2021. This model is utilized to forecast the cumulative infections, recoveries, and deaths for one month ahead. The proposed model could help researchers, health officials, and policymakers forecast the epidemic size and develop policies to defeat this crisis.

3. Proposed framework

The objective of this research is to forecast the prevalence of the COVID-19 outbreak in Egypt. Fig. 1 depicts the mechanism of the proposed framework in making forcast. In order to achieve this main objective, the following stapes are carried out: 1) collecting the COVID-19 data, 2) preparing and pre-processing the data before being imported to neural networks, 3) splitting the datasets into training and testing datasets, 4) employing deep and machine learning models, 5) evaluating the performance of models using assessment metrics, and 6) forecasting using the optimum model. The features of the proposed framework are discussed in the next sub-sections.

Fig. 1.

Research framework schematic diagram.

3.1. Data preparation

It can be observed that the dataset has lengthy sequences, causing the following issues (Khan et al., 2020): 1) consuming a long time to train and demanding a lot of memory, and 2) getting a poor learned model caused by back-propagating lengthy sequences. As a result, it is crucial to prepare and pre-process the data before being imported to neural networks. Fig. 2 illustrates the data preparation procedure. The lengthy sequence ($S$) is partitioned into sub-sequences ($\hat{S}$) of fixed-length ($L$), where $S=\{x_t,x_{t+1},\dots x_T\}$ and $\hat{S}=\{x_t,x_{t+1},\dots x_{t+L}\}$. In this research, the length of the sub-sequence is equal to 20 consecutive days. The data is then pre-processed in the following phase. Data normalization and standardization are two techniques used in the data preparation phase (Ali and Faraj, 2014). In this research, data standardization is utilized as a scaling approach to set the mean and standard deviation to 0 and 1, respectively. After data standardization, the neural networks conduct time-series analysis by splitting the data in each sub-sequence into input and output observations. For one-step prediction, the first twenty data points of each sub-sequence are modeled as inputs and the twenty-first data point is regarded as an output.

Fig. 2.

Data preparation pipeline.

3.2. Dataset splitting

The dataset is further prepared such that the first 90 % and 10 % of the datasets are utilized for training and testing purposes, respectively. The training set is used to train and improve the models, whereas the test set is used to assess the performance of the models (Zhang et al., 2020b).

3.3. Long short-term memory network

LSTM is a deep recurrent neural network that gained its popularity from its internal memory mechanism (Chung et al., 2014; Donahue et al., 2015; Lipton et al., 2015; Le et al., 2019). It has a high capability to learn dependencies and process large quantities of data for a long time (Elsheikh et al., 2021a). This network has a chain architecture consisting of several modules. These modules are used to store processed data from previous stages (Sherstinsky, 2020). The structures of these modules differ from those of traditional recurrent neural networks. The structure of a typical LSTM consists of several memory modules called cells. Hidden and cell states are transferred from the previous cell to the next one. The data is processed as it moves in the forward direction through the cell state. The memory blocks remember things through three primary gate controllers; input, forget, and output gates. The input gate adds information to the cell state while the forget gate filters the unrequired or unimportant information from the cell state. Finally, the output gate displays the output from the current cell state. These gates are responsible for enhancing the network capabilities by controlling the memorizing process and avoiding dependency problems (Elsheikh et al., 2021b). As depicted in Fig. 3 , the proposed LSTM architecture comprises an input layer, an LSTM layer, a dropout layer, a fully connected layer, and an output layer.

Fig. 3.

Proposed LSTM architecture.

3.4. Convolutional neural network

CNN is another class of deep learning models that gained popularity in features and patterns from the time series analysis (Khan et al., 2020). It differs from a recurrent network as it is a feedforward network that filters spatial data, unlike a recurrent network that feeds data back to itself. CNN has achieved good performance in forecasting time series data (Acharya et al., 2017; Zhao et al., 2017). Its structure is inspired by biological neurology. The neurons in this network are only connected to the closer neurons, and they all have the exact weights, unlike a simple neural network. The connections among the neurons are simplified to uphold the spatial aspects of the dataset. As shown in Fig. 4 , its architecture comprises convolutional, batch normalization, rectified linear unit (ReLU), pooling, and fully connected layers. The convolutional layer is responsible for extracting features by placing filters over inputs and creating a convolved feature map based on filter stride and size. The pooling layer applies a downsampling operation to reduce the sample size of feature maps, minimizing the computational efforts required for processing data. There are two ways to produce a pooled feature map; max pooling and average pooling. The max-pooling and average pooling take the maximum and average input of a specific convolved feature, respectively. The ReLU layer acts as an activation function to maintain nonlinearity and dimensionality as the data propagates from one layer to another in the network. Finally, the fully connected layer delivers the output by mapping the extracted features (Khan et al., 2020).

Fig. 4.

Proposed CNN architecture.

3.5. Multilayer perceptron neural network

MLP belongs to the feedforward neural network, and it is used to execute predictions for complex datasets (Hui et al., 2020). This method is known for its ease of implementation and shorter training time. Furthermore, it is suitable for modeling diseases because it yields quick results with satisfactory regression performance (Car et al., 2020). It comprises three layers; input layer(s), hidden layer(s), and output layer(s). The input layer consists of neurons representing the dataset inputs (Goodfellow et al., 2016). Besides, the output layer consists of neurons that correspond to the predicted output. The neurons apply an activation function on weighted inputs to forecast the output (Bishop, 2006; Goodfellow et al., 2016). The weights of connections among neurons are random at the initial stage. This network is trained iteratively using an optimization approach, and an error metric is computed based on the difference between observed and predicted values in each training cycle. The weights of neuron connections are adjusted based on the contribution to the total error through the backward propagation process (Bishop, 1999).

3.6. Model evaluation criteria

The forecasting models are assessed using two statistical indices; determination coefficient (R²) and root mean square error (RMSE). These indicators give insights into the model accuracy and goodness of fit for the datasets. Eqs. (1), (2) represent the mathematical formulas of these indices (Saba and Elsheikh, 2020; Elsheikh et al., 2021b).

$$R^2=\frac{{\left(\sum _{i=1}^n\left(\left(x_i-\overline{x}\right)\left(y_i-\overline{y}\right)\right)\right)}^2}{\sum _{i=1}^n{\left(x_i-\overline{x}\right)}^2\times \sum _{i=1}^n{\left(y_i-\overline{y}\right)}^2}$$ (1)

$$RMSE=\sqrt{\frac{1}{n}\sum _{i=1}^n{\left(x_i-y_i\right)}^2}$$ (2)

Where; $n$, $x$, $y$, $\overline{x}$, and $\overline{y}$ are the numbers of cases, reported cases, predicted cases, average reported cases, and average predicted cases, respectively.

4. Study area

Egypt lies in the African continent, and its area covers about 1 million square kilometers. It lies between latitude 30° 06′ N and longitude 31° 25′ E. The Mediterranean Sea borders it to the north, Gaza Strip, Israel, and the Red Sea to the east, Sudan to the south, and Libya to the west. It is considered a lower-middle-income country with respect to a World Bank categorization. The total population in Egypt exceeds 100 million people, representing 1.3 % of the total world population and 7.6 % of the Africa population. About 32.5 % of the entire population lives under the national poverty line. The Egyptian health systems have a moderate capacity to confront the coronavirus. The physician and hospital bed densities are 0.79 physicians and 1.6 beds per 1000 persons, respectively. Therefore, the COVID-19 outbreak is a significant challenge for health workers and policymakers. However, the demographic structure of Egypt differs from other European and Asian countries. About 21 % of the Egyptians are between 18 and 29 years old, while only 3.9 % exceed 65 years old. It was found that older people and those suffering debilitating diseases are more susceptible to COVID-19. On the other hand, the Egyptian youth can act as a defensive line to limit the spread of this virus.

5. Data collection

Dataset used in this research is acknowledged by the Egyptian ministry of health and population and it is obtained from the open-source (https://flevy.com/coronavirus). The dataset comprises daily confirmed, death, and recovered cases from 14 February 2020 to 30 June 2021. This dataset spans over 503 days. Fig. 5 shows the number of reported cases during this period. The data statistics are depicted in Table 1 . It is clear from Fig. 5 and Table 1 that the COVID-19 profile is highly variable.

Fig. 5.

Daily number of COVID-19 cases in Egypt.

Table 1.

Descriptive statistics for the daily COVID-19 data in Egypt.

Statistics	Confirmed cases	Death cases	Recovered cases
Minimum	–	–	–
Maximum	1774	97	1716
Median	523	31	411
Mean	559	32	420
Standard deviation	435	22	334

The frequency distribution of the data is illustrated in Fig. 6 . The means of the daily infected cases, deaths, and recoveries are 559, 32, and 420. Additionally, the medians of the daily infections, deaths, and recoveries are 523, 31, and 411, respectively. The data are not normally distributed and are rather positively (right) skewed because the means are greater than the medians.

Fig. 6.

Histogram of daily COVID-19 cases.

Fig. 7 presents an overview of the COVID-19 spread in Egypt. As of 30 June 2021, the cumulative number of confirmed cases is 281,282, with 211,384 recoveries and 16,169 deaths. It could be noted that the virus spread is growing exponentially, which implies the need to apply monitoring and control systems. The future progression of this outbreak is still ambiguous because it spreads randomly. As shown in Fig. 8 , Egypt is considered one of the most COVID-19 hotspots in Africa as it has 2,748.66 cumulative confirmed cases per million people on 30 June 2021. It has been ranked as the fourth African country after South Africa, Morocco, and Tunisia regarding the cumulative confirmed cases. Moreover, regarding COVID-19 resilience ranking, it has been ranked 1st in Africa, 3rd in the Middle East, and 25th globally, as stated by the Egyptian cabinet’s information and decision support center. It is worth mentioning that the COVID resilience index estimates the capacity of nations to battle COVID-19 based on ten indicators: deaths per 1 million, monthly fatality rate, monthly infected cases per 100k, positive test rate, lockdown effectiveness, availability to COVID immunizations, society mobility, growth rate of the gross domestic product in 2020, medical care inclusion, and human advancement index (IDSC, 2020).

Fig. 7.

Cumulative number of COVID-19 cases in Egypt.

Fig. 8.

Cumulative COVID-19 cases per million people in African countries (Our World in Data, 2021b).

Furthermore, the total number of active cases (${ac}_i$), active rate (${ar}_i$), death rate (${dr}_i$), and recovered rate (${rr}_i$) could be calculated using Eqs. (3), (4), (5), (6), respectively (Gupta et al., 2020). Fig. 9 illustrates the active, death, and recovery rates in Egypt.

$${ac}_i={cc}_i-{dc}_i-{rc}_i$$ (3)

$${ar}_i=\left({ac}_i/{cc}_i\right)\times 100$$ (4)

$${dr}_i=\left({dc}_i/{cc}_i\right)\times 100$$ (5)

$${rr}_i=\left({rc}_i/{cc}_i\right)\times 100$$ (6)

Where; ${cc}_i$ represents confirmed cases, ${dc}_i$ represents death cases, and ${rc}_i$ represents recovered cases for the ith day.

Fig. 9.

Active, death, and recovery rates due to COVID-19 in Egypt.

6. Results and discussion

The LSTM, CNN, and MLP models are employed to predict the cumulative number of infections in Egypt. The general parameter setting of the deep learning models is depicted in Table 2 . The data reported from 14 February to 15 August 2020 is divided into 90 % for the training phase (166 readings from 14 February 2020 to 28 July 2020) and 10 % for the testing phase (18 readings from 29 July 2020 to 15 August 2020). The trained models are applied to forecast the total number of infections for a short period (i.e., one week ahead). These models are developed using MATLAB R2019a, under the same environment of Intel i7−2.20 GHz, 8 GBs of RAM, and Windows 10.

Table 2.

General parameter setting of the deep learning models.

Parameter	Value
Optimal time delay	20
Maximum number of training epochs	200
Minimum batch size	256
Execution environment	Central processing unit (CPU)
Optimization algorithm	Adam

The CNN architecture comprises three blocks, where each block is composed of convolutional, batch normalization, and rectified linear unit (ReLU) layers. The convolutional layers of the first, second, and third blocks comprise 16, 32, and 64 filters, respectively. For the three blocks, the sizes of filters, stride, and padding are 2, 1, and 1, respectively along all edges of the layer. Two max-pooling layers are utilized after the first and second blocks. The first pooling layer has pool size 2 and stride 2 while a second pooling layer with pool size 2 and stride 1 is utilized. For the LSTM model, the number of hidden layers, the number of hidden units in each layer, and the dropout value in each layer are assumed to be 1, 120, and 0.5, respectively. Regarding the MLP model, the number of feedforward layers is three, while the training function is scaled conjugate gradient. The number of neurons in the first, second, and third layers is 3, 5, 10, respectively. It shall be noted that the maximum number of training iterations for the prediction models is 200. The time series plot of the reported and forecasted data is illustrated in Fig. 10 . An excellent agreement is observed between the forecasted results of the LSTM network (i.e., 97,396 cases) and the reported official data (i.e., 97,237 cases) on 22 August 2020.

Fig. 10.

Cumulative number of confirmed and forecasted cases for one week ahead using artificial intelligence-based models.

Different assessment criteria are applied to determine the optimal model as shown in Fig. 11 . The R² of the predicted data using LSTM is higher than that of CNN and MLP models by 0.017 % and 0.19 %, respectively. Furthermore, the RMSE of the predicted data using LSTM is less than that of CNN and MLP models by 47.06 % and 76.65 %, respectively. It can be observed that the MLP model has failed to obtain acceptable forecasting results. Additionally, the LSTM model outperforms the remaining models for one week ahead forecasting based on the results of applied metrics.

Fig. 11.

Assessment criteria for one week ahead forecasting using artificial intelligence-based models.

Furthermore, the forecasting performance of the models is assessed to determine the accumulated cases for a long period (i.e., one month ahead) using the data reported from 14 February to 15 August 2020. The time series plot of the reported and forecasted data is illustrated in Fig. 12 . It could be noted that there exists a good agreement between the forecasted results of the LSTM network (i.e., 98,346 cases) and the reported official data (i.e., 101,177 cases) on 14 September 2020.

Fig. 12.

Cumulative number of one month ahead confirmed and forecasted cases using artificial intelligence-based models.

The performance of the forecasting models is evaluated using different assessment criteria as shown in Fig. 13 . The R² of the predicted data using LSTM is higher than that of CNN and MLP models by 0.11 % and 0.16 %, respectively. Furthermore, the RMSE of the predicted data using LSTM is less than that of CNN and MLP models by 67.56 % and 67.94 %, respectively. Regarding the results of the applied metrics, the LSTM model outperforms the remaining deep learning and machine learning models for one month ahead forecasting.

Fig. 13.

Assessment criteria for one month ahead forecasting using artificial intelligence-based models.

The LSTM performance is impacted by the values of many hyperparameters such as the number of hidden layers, the number of hidden units in each layer, and the drop-out ratio. In this research, the dropout value in each layer is fixed as 0.5, while the remaining parameters are determined by trial and error. The number of hidden units of 50, 100, and 200 is used while changing the number of hidden layers to be 1, 2, 3, and 4. As illustrated in Fig. 14 , the prediction accuracy of the different architectures of the LSTM model is evaluated using two assessment criteria. It is found that the network with 50 hidden units and 3 hidden layers has achieved the highest forecasting accuracy among all the other networks. This network is associated with the highest R² value of 0.9999 and the lowest RMSE value of 346.12. Therefore, it is suggested to set the network parameters at 50 hidden units and 3 hidden layers to maximize the forecasting performance of the LSTM network.

Fig. 14.

Assessment criteria for different architectures of LSTM network.

The results of the proposed LSTM model are compared against those reported in the literature. The suggested model yields an R² value of 0.999, compared to 0.989 (Elsheikh et al., 2021b) and 0.94 (Car et al., 2020). Besides, it is associated with an RMSE value of 346.12, substantially lower than the reported value of 473.454 (Elsheikh et al., 2021b). It can be concluded that the proposed model enhances the robustness of modeling COVID-19 infection cases.

The proposed model with the optimal parameters is trained using 90 % of the data reported from 14 February 2020 to 30 June 2021. This model is utilized to forecast the rate of COVID-19 spread for one month ahead (i.e., from 30 June 2021 to 31 July 2021). Fig. 15 demonstrates the behavior of the proposed LSTM model in terms of the learning loss curve. The deep learning model learns quite quickly and achieves a static loss curve after 400 epochs.

Fig. 15.

Infection model loss curve for LSTM model.

As plotted in Fig. 16 , the infection curve would increase at a slow pace in the next month. It is estimated that the total number of infections, recoveries, and deaths on the last day will reach 285,939, 234,747, and 17,251 cases, respectively. This implies that the cumulative active cases will be 33,941 cases on 31 July 2021. This confirms that the third wave of the epidemic has begun to recede, and the number of coronavirus cases in Egypt is declining. The number of daily infections will decrease from 251 cases on 30 June 2021 to 147 cases on 31 July 2021. This follows the average infection toll of 621 cases/day throughout June 2021. Moreover, the recovery rates from the COVID-19 at isolation hospitals would increase rapidly, achieving more than 80 % in late July 2021. This can be attributed to the public awareness and vaccination campaigns, which have dramatically reduced infection estimates. The results of the forecasting model could assist policymakers and healthcare providers in confronting this epidemic in Egypt.

Fig. 16.

Cumulative number of the reported and predicted cases.

7. Conclusion

The World Health Organization has declared COVID-19 as a global health pandemic by March 2020. The application of artificial intelligence-based methods has been emphasized as an alternative for clinical methods to model the spread of this infectious disease. This intelligent method could reduce the burden on healthcare systems and revive the economic sector. This research applied two deep learning models to predict the spread of the COVID-19 outbreak in Egypt. These models were the long short-term memory network (LSTM) and convolutional neural network (CNN). The models were trained and tested using 90 % and 10 % of the data reported by the Egyptian ministry of health and population from 14 February to 15 August 2020. The trained models were applied to forecast the total number of infections for one week ahead. The forecasting performance of these models was assessed against that of multilayer perceptron (MLP) using determination coefficient (R²) and root mean squared error (RMSE). The LSTM model obtained an R² value of 9.999E-01, higher than that of CNN and MLP models by 0.017 % and 0.19 %, respectively. Furthermore, it obtained an RMSE value of 436.06, less than that of CNN and MLP models by 47.06 % and 76.65 %, respectively. The forecasting performance of the models was assessed to determine the accumulated cases for one month ahead. The R² of the predicted data using LSTM was higher than that of CNN and MLP models by 0.11 % and 0.16 %, respectively. Besides, the RMSE of the predicted data using LSTM was less than that of CNN and MLP models by 67.56 % and 67.94 %, respectively. Therefore, the LSTM model outperformed the remaining deep learning and machine learning models for one week and one month-ahead forecasting. The reason is that feedback connections characterize the LSTM network to propagate the data in the backward pass, improving the prediction accuracy. Besides, it could capture the nonlinear pattern in the input data over time.

The optimal parameters of the LSTM network were determined and the optimum number of hidden layers and the number of hidden units in each layer were 3 and 50, respectively. The network with the optimal parameters was trained using 90 % of the data from 14 February 2020 to 30 June 2021. The model was applied to forecast the cumulative number of infections, recoveries, and deaths for one month ahead. The forecasted numbers of infected cases, recovered cases, and deaths on 31 July 2021 were 285,939, 234,747, and 17,251 cases, respectively. It is worth mentioning that the reliability of the provided results relies on data quality because under-testing could lead to fewer reported cases and more forecasting errors. Besides, the performance of the models is limited to the presence of noise in the daily data. In such cases, it is recommended to apply the moving average technique for smoothing out variations and separating noise from trends in time series data. However, this research could help researchers, public health officials, and policymakers forecast and track the COVID-19 outbreak in Egypt. Obtaining accurate predictions will improve the resilience ability of the society against COVID-19 disease. The resilience strategy cannot prevent external events from jeopardizing the health system. However, it can improve the system’s ability to overcome unanticipated crises and become stronger over time. The major contributions of this research could be summarized as follows:

• ▪Estimating the overall number of infections, recoveries, and fatalities in Egypt
• ▪Investigating the application of deep learning models using the official reported data till June 2021
• ▪Comparing the performance of machine learning approaches and deep learning techniques to anticipate the epidemic size in Egypt

This research could be extended in the future to enable modeling the impact of non-pharmaceutical interventions (e.g., mask-wearing, social distancing, and lock-down effectiveness) on controlling the COVID-19 spread. Also, it can study the effects of vaccination campaigns for COVID-19 prevention using system dynamics technique.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.