Artificial Intelligence-Driven Ensemble Model for Predicting Mortality Due to COVID-19 in East Africa

East Africa was not exempt from the devastating effects of COVID-19, which led to the nearly complete cessation of social and economic activities worldwide. The objective of this study was to predict mortality due to COVID-19 using an artificial intelligence-driven ensemble model in East Africa. The dataset, which spans two years, was divided into training and verification datasets. To predict the mortality, three steps were conducted, which included a sensitivity analysis, the modelling of four single AI-driven models, and development of four ensemble models. Four dominant input variables were selected to conduct the single models. Hence, the coefficients of determination of ANFIS, FFNN, SVM, and MLR were 0.9273, 0.8586, 0.8490, and 0.7956, respectively. The non-linear ensemble approaches performed better than the linear approaches, and the ANFIS ensemble was the best-performing ensemble approach that boosted the predicting performance of the single AI-driven models. This fact revealed the promising capability of ensemble models for predicting the daily mortality due to COVID-19 in other parts of the globe.


Introduction
"Artificial Intelligence could be the saviour of the COVID-19 pandemic in the coming year; we just need to prove it." The Lancet Digital Health, 2021 Most pandemics in the 20th and 21st centuries were caused either by the coronavirus or the influenza virus. The coronavirus disease 2019 (COVID-19) is one of the 21st-century pandemics and highly contagious infections caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV2) [1,2]. The WHO, on 11 February 2020, named this outbreak COVID-19, used as a shorthand for coronavirus disease 2019. Again, the WHO, on 30 January 2020, affirmed this outbreak as a "Public health emergency of international concern" and finally as a "Pandemic" on 11 March 2020 [3].
In terms of mortality, COVID-19 has caused more than 6.5 million deaths (6,559,902 as of 8 October 2022) globally, with a case fatality rate of 2.04% [4]. This number proves that the pandemic is much different, in terms of global crises, compared with previous flu pandemics, such as the Spanish flu (in 1918), the Asian flu (1957)(1958), the Hong Kong flu (1968)(1969)(1970), and the swine flu (2009)(2010). The nature of this pandemic made COVID-19 the first global public health issue that had a brutal impact on the global economy, which triggered a near to total shutdown of social and economic activities. Finally, the pandemic has shrunk the global economy by nearly 3 per cent according to the prediction of the International Monetary Fund (IMF) [5].

Data Source and Attribute Selection
This study used COVID-19-related data collected daily over 24 months, from April 2020 to April 2022, in the region. These data were public data from the "Our World in Data (OWID)" team and the COVID-19 data warehouse at John Hopkins University (JHU), collected by the Center for Systems Science and Engineering (CSSE), which is open to researchers. The data were retrieved from: https://github.com/owid/COVID-19-data/tree/ master/public/data (accessed on 25 June 2022). The OWID, in its statement under the license section, explains that "All visualizations, data, and code produced by 'Our World in Data' is completely open access under the Creative Commons BY license. You have the permission to use, distribute, and reproduce these in any medium, provided the source and authors are credited" [22,23]. Ethical approval was not necessary, as this dataset does not include personal information and is public, as approved by the CSSE and JHU.

Data Preprocessing and Analyses
In the 'Our World in Data (OWID)' COVID-19 database, many variables are available, but ten variables were selected because of the completeness of their data and their relationship with mortality. Datasets were retrieved for each country in the East Africa region independently, and we calculated the average values to represent the region with a single variable for mortality and other input variables. It is known that data collected daily is non-linear by nature. Hence, the first step that we took was to normalize each variable in the dataset.
The second activity that we undertook was to select the dominant input variables through a non-linear sensitivity analysis called the coefficient of determination (DC). This analysis was conducted using an artificial neural network (ANN), applying one target and one input variable to predict the estimated values and calculate the coefficient of determination so as to verify the correlation of each input variable with the target variable.
The dataset from these countries was classified into two training (70%) and testing (30%) sub-datasets for the development of AI-driven single and ensemble models. For all the developed AI-driven models, the target variable was the daily number of new deaths due to COVID-19 in the region, and the input variables were the new daily number of cases in the region, the positive rate, the number of people vaccinated, hospital beds/1000 patients, and so on. In Table 1, the list of all the variables used in this study and their explanations are presented. In the data processing step, the data normalization was calculated using Microsoft Excel. However, the menu-based MATLAB (Version 20) was applied to conduct the sensitivity analysis, the single black-box AI-driven models, and AI-driven ensemble models.

Proposed Methods
In this study, we modelled three AI-driven models, including an adaptive neurofuzzy inference system (ANFIS), feedforward neural network (FFNN), support vector machine (SVM), and one conventional data-driven model, multiple linear regression (MLR), to predict mortality due to COVID-19 in East Africa. In addition, we classified a training dataset and a test dataset after normalizing the inputs. Figure 1 shows that three stages were conducted to carry out the given study. Firstly, the selection of dominant inputs for the prediction of COVID-19 mortality in the region was conducted to rank and select the most influential input variables for the modelling. In the second stage, four AI-driven black-box models (ANFIS, SVM, FFNN, and MLR) were applied independently to predict the COVID-19 mortality. Thirdly, as a final stage, four ensemble approaches, namely the ANFIS ensemble (ANFISE), neural network ensemble (NNE), weighted average ensemble (WAE), and simple average ensemble (SAE), were constructed. In the ensemble stage, the estimated output of every single model was used as an input for the AI-driven ensemble process. Then, the predicted mortality based on the ensemble model was compared with the predicted results from each of the black-box models in the second stage.
Diagnostics 2022, 12, x FOR PEER REVIEW 4 o In the data processing step, the data normalization was calculated using Micro Excel. However, the menu-based MATLAB (Version 20) was applied to conduct the sitivity analysis, the single black-box AI-driven models, and AI-driven ensemble mod

Proposed Methods
In this study, we modelled three AI-driven models, including an adaptive ne fuzzy inference system (ANFIS), feedforward neural network (FFNN), support vector chine (SVM), and one conventional data-driven model, multiple linear regression (M to predict mortality due to COVID-19 in East Africa. In addition, we classified a train dataset and a test dataset after normalizing the inputs. Figure 1 shows that three sta were conducted to carry out the given study. Firstly, the selection of dominant inputs the prediction of COVID-19 mortality in the region was conducted to rank and select most influential input variables for the modelling. In the second stage, four AI-dri black-box models (ANFIS, SVM, FFNN, and MLR) were applied independently to pre the COVID-19 mortality. Thirdly, as a final stage, four ensemble approaches, namely ANFIS ensemble (ANFISE), neural network ensemble (NNE), weighted average ensem (WAE), and simple average ensemble (SAE), were constructed. In the ensemble stage estimated output of every single model was used as an input for the AI-driven ensem process. Then, the predicted mortality based on the ensemble model was compared w the predicted results from each of the black-box models in the second stage. The artificial neural network (ANN) is one of the most significant AI-driven mod because it can build links between the target and input variables by training the ne network without having comprehensive information on the entire data set [24]. T model is a self-learning simulation function that demonstrates the capacity to model forecast complicated processes. This capability makes ANN a more practical and effic model in different domains of science, such as biomedical technology, engineering, a culture, and business [25].
Because of its simplicity and favored ability to react to various challenges with considering the past information regarding the process, this study used the feedforw neural network (FFNN), employing propagation algorithms. FFNN is formed of lin The artificial neural network (ANN) is one of the most significant AI-driven models, because it can build links between the target and input variables by training the neural network without having comprehensive information on the entire data set [24]. This model is a self-learning simulation function that demonstrates the capacity to model and forecast complicated processes. This capability makes ANN a more practical and efficient model in different domains of science, such as biomedical technology, engineering, agriculture, and business [25].
Because of its simplicity and favored ability to react to various challenges without considering the past information regarding the process, this study used the feedforward neural network (FFNN), employing propagation algorithms. FFNN is formed of linked pieces called 'nodes' that have unit properties of information, such as learning, nonlinearity, noise tolerance, generalization capability, and so on, and it has three layers (see Figure 2), including the input, the hidden, and the output layers. As a result, the input variables provided to the input layers' 'neurons' are transmitted forwards, and the activation function, a non-linear function, is employed to construct the output vector.
• Compute    A multi-layer perceptron (MLP) model with a single hidden layer was computed in this study. The formal definition of this model is as follows: the function 'f ' on the fixedsize input 'x', such as f(x) ≈ y for training pairs of (x, y). Alternatively, recurrent neural networks learn sequential data, computing the output 'Ø' on the variable-length input Xk = {x1 . . . Xk} ≈ yk for training pairs of (Xn, Yn) for all 1 ≤ k ≤ n.
In the definition of FFNN with the 'm' layer (or 'm-2 ′ hidden layers) prototype, the output perceptron has an activation function Øo, and the hidden layer perceptron has activation functions Ø. Every perceptron in layer li is connected to every perceptron in layer l i−1. The layers are fully connected, and there is no connection between the perceptrons in the same layer. Hence, According to Brilliant [26], it is computed using the following formulas: First, initialize the input layer l 0 and set the values of the outputs Ø 0 i for nodes in the input layer l 0 in relation to their associated inputs in the vector Second, compute the sum of the products and each output of the hidden layer in the order from l 1 to l m−1 for 'k', progressing from 1 to m−1 , where the MLP uses the denotations below. The w k ij is the weight for perceptron j in the layer l k for the incoming node i, b k i is the bias for the perceptron i in layer l k , h k i is a product of some plus bias for perception i in layer l k , Ø k i is the output for node i in layer l k , r k is several nodes in layer lk, w →k i is the weight vector for perceptron i in layer l k , and Ø →k is the output vector for layer l k .

The Adaptive Neuro-Fuzzy Inference System (ANFIS)
The ANFIS is developed by combining the ability to learn the neural network and its advantage of a rule-based fuzzy inference system, which enables it to integrate a past observation into the process of classification [27,28]. This combination makes ANFIS a good model for overcoming the limitations of individual modelling. Jang JS, in 1997, described the 'defuzzifier', 'fuzzifier', and 'fuzzy' databases as the three parts of a fuzzy system [28].
The ANFIS architecture contains five layers: layer 1 is the input layer, layer 2 is the input membership function (MFs), layer 3 is the association rules, layer 4 is the output membership function, and layer 5 is the model output (see Figure 3). After the construction of the fuzzy system, it specifies the relationship between the fuzzy variables using the 'ifthen' fuzzy rules. The first order of Sugeno's system has the following rules, considering that the FIS contains a single output (f ) and two inputs (x and y): where A and B are membership functions, and p, q, and r are parameters for the outlet functions. Assuming these parameters, the structure of the ANFIS with five layers is as follows:

The Support Vector Machine (SVM)
According to A. M. Kalteh, the support vector machine (SVM) may be utilized f both prediction and classification [11,32]. The type of regression of this model is know as support vector regression. It is used to define regression using SVM and structural ri reduction. Figure 4 depicts the framework of the SVM regression technique that can sim ulate non-linear situations in the real world. The estimation obtained using this regressio is important for estimating a function of the given dataset: where xi is the input vectors, di is the actual values, and n is the total number of the datase Hence, SVM has the following regression function: The Minimization: This is subjected to: Layer 1: every i's node is an adaptive node that has the following node function on this layer: Where Q i 1 is for input x or y, that is, the membership grade. Here, the Gaussian membership function was selected due to the fact that it has the lowest error of prediction.
Layer 2: In this layer, the 'T-Norm' operator connects every rule using the 'AND' operator between the inputs and is presented as: Layer 3: In this layer, the output is the 'Normalized firing strength', and the labelled norm for every neuron is as follows: Layer 4: In this layer, every i's node is an adaptive node and executes the consequence of the rules by considering p 1 , q 1 , and r 1 as an irregular parameter, as follows: In this layer, the overall output is calculated by summing all the incoming signals, as follows:

The Support Vector Machine (SVM)
According to A. M. Kalteh, the support vector machine (SVM) may be utilized for both prediction and classification [11,32]. The type of regression of this model is known as support vector regression. It is used to define regression using SVM and structural risk reduction. Figure 4 depicts the framework of the SVM regression technique that can simulate non-linear situations in the real world. The estimation obtained using this regression is important for estimating a function of the given dataset: where x i is the input vectors, d i is the actual values, and n is the total number of the dataset. Hence, SVM has the following regression function: where φ is a non-linear mapping function, and 'ω 'and 'b' are parameters of the function of the regression that can be determined by assigning positive values for the slack parameters of ξ and ξ* and the minimization of the objective function, considering 'c' as the regularized constant and 1 2 ||w∥ 2 as the weight vector norm, as shown below.
Diagnostics 2022, 12, x FOR PEER REVIEW 8 The optimization problem stated above could be improved to obtain a dual quad problem of the optimization, defining the lag-range multipliers αi and αi*. In additio vector 'w' can be computed by identifying the optimization solution problem, as fol Hence, the SVM regression function is changed to: where b is the bias term and k (xi, xj) is the kernel function that can be expressed as: MLR is commonly used as a statistical modelling technique to observe the line lationships between numerically measured variables. It is a form of linear regression to examine the linear relationship between a single target variable and several inpu The Minimization: This is subjected to: The optimization problem stated above could be improved to obtain a dual quadratic problem of the optimization, defining the lag-range multipliers α i and α i *. In addition, the vector 'w' can be computed by identifying the optimization solution problem, as follows: Hence, the SVM regression function is changed to: where b is the bias term and k (x i , x j ) is the kernel function that can be expressed as: where γ is the parameter of the kernel.

The Multiple Linear Regression (MLR)
MLR is commonly used as a statistical modelling technique to observe the linear relationships between numerically measured variables. It is a form of linear regression used to examine the linear relationship between a single target variable and several input variables. In this technique, the dependent variable (Y) is supposed to be affected by the independent variables (Xi), and the estimated model can be expressed as: where y is the target variable,β 0 is the regression constant, β i are coefficients of the input variables, and x i is input variables.

Ensemble Modelling
In the AI industry, ensemble modelling is computed by combining the estimated predictions of multiple single AI-driven models. This combination can advance the final model's performance and it can provide better predictions than the individual models [33]. To boost the performance of single models, this study used two linear ensemble techniques, the weighted average ensemble (WAE) and simple average ensemble (SAE), and two nonlinear ensemble techniques, the ANFIS ensemble (ANFISE) and neural network ensemble (NNE), were applied (see Figure 5). These ensemble techniques have been applied in various studies for purposes such as the clustering and classifications of medical data, web ranking, economic forecasting, etc. [34][35][36][37][38][39]. Considering this situation, this study also applied the ensemble technique to predict COVID-19 mortality in East Africa.
Diagnostics 2022, 12, x FOR PEER REVIEW 9 two non-linear ensemble techniques, the ANFIS ensemble (ANFISE) and neural net ensemble (NNE), were applied (see Figure 5). These ensemble techniques have bee plied in various studies for purposes such as the clustering and classifications of me data, web ranking, economic forecasting, etc. [34][35][36][37][38][39]. Considering this situation study also applied the ensemble technique to predict COVID-19 mortality in East Af

The Linear Ensemble Approaches
In this approach, the simple average (SA) and weighted average (WA) ensemble niques were applied. In the simple average technique, the arithmetic average of the put, the COVID-19 mortality, of every single AI-driven model is taken as the final dicted mortality in the region. Meanwhile, in the weighted average technique, the pr tion is computed by assigning weights to each output relative to its importance.
The formula for the simple average: COVID-19 is the output of the SA ensemble model, and COVID-19i is the output o

The Linear Ensemble Approaches
In this approach, the simple average (SA) and weighted average (WA) ensemble techniques were applied. In the simple average technique, the arithmetic average of the output, the COVID-19 mortality, of every single AI-driven model is taken as the final predicted mortality in the region. Meanwhile, in the weighted average technique, the prediction is computed by assigning weights to each output relative to its importance.
w i is a weight for the output of ith method and is computed using the performance measure called the determination coefficient (DC) and can be calculated with DC i is the coefficient of determination for the ith model.

The Non-Linear Ensemble Approaches
In this approach, the non-linear averaging was computed by training the single AIdriven non-linear models (FFNN and ANFIS) using the COVID-19 mortality values predicted by these single models, and the neural network ensemble (NNE) and ANFIS ensemble (ANFISE) were applied. In NNE, the non-linear averaging was performed by training different FFNNs by feeding the output of an AI-driven single model as an input. Then, the maximum epoch number and neurons of the hidden layer were determined by trial and error. Meanwhile, in ANFISE, the predicted COVID-19 mortality based on the AI-driven single model is fed to ANFIS for training using a different number of epochs and membership functions (MFs).

Normalization and Evaluation of Models
Both the target and input data should be standardized before training the model at an early stage to reduce dimensions and guarantee that all the variables receive equal attention [40]. The following normalization formula should be performed on the dataset to established the values within the range of 0-1: , and COVID-19max represent the normalized, actual, minimum, and maximum COVID-19 mortality values, respectively. Even though the best model for the validation and training steps is determined by trial and error [41], the root mean square error (RMSE) and determination coefficient (DC) are used to measure the performance and efficiency of the developed models. The DC values range from −1 to 1, and a model value approaching 1 yields better results. In addition, the model with lowest the RMSE is considered to be the best model. The formulas are as follows: where COVID-19 obsi , COVID-19 prei , − (COV ID-19) obs , and N are the observed COVID-19 mortality value, predicted COVID-19 mortality value, average of the observed COVID-19 mortality values, and number of observations, respectively.

Results and Discussions
This study used three AI-driven and one classical mode, namely ANFIS, SVM, FFNN, and MLR, respectively. All the models were trained and tested to model the mortality due to COVID-19 in East Africa. In the results section, the descriptive statistics, sensitivity analysis, single AI-driven black-box modelling, and ensemble modelling of mortality due to COVID-19 are successively presented.

Descriptive Statistics
A line graph was used to present the data on the average daily mortality due to COVID-19 in East Africa ( Figure 6). In this graph, the three largest numbers of mortality cases observed per single day were 979, 966, and 737 deaths on 5 July 2021, 7 March 2022, and 1 October 2021, respectively. However, more than 200 cases per day were registered successively from July to August 2021, and we can conclude that this period was the peak time of mortality in the region. The descriptive statistics in Table 2 present the average, maximum, and min values for the target and all the input variables of the training (70% of the dataset an 584) and verification datasets (30% of the dataset and n = 146) from 01 April 2020 April 2022. The average mortality due to COVID 19 was (61.03 ± 69.1) for the tra dataset, and it was (46.16 ± 83.59) for the verification dataset. The average number o daily cases was (2783.5 ± 2423) for the training and (5724.66 ± 6522.36) for the verifi data, while the rates of confirmed positive cases per day were 0.041 ± 0.022 and 0.052 for the training and verification datasets, respectively. The daily vaccine cov and hospital beds per 1000 people, which are also presented in the table, showed th average daily vaccine coverage was (26234.2 ± 47498.4) and (220514 ± 332466) for th ing and verification data, respectively. The hospital beds/1000 people were (28.25 ± and (20.4 ± 0.5623) for the training and verification datasets, respectively.  The descriptive statistics in Table 2 present the average, maximum, and minimum values for the target and all the input variables of the training (70% of the dataset and n = 584) and verification datasets (30% of the dataset and n = 146) from 1 April 2020 to 1 April 2022. The average mortality due to COVID 19 was (61.03 ± 69.1) for the training dataset, and it was (46.16 ± 83.59) for the verification dataset. The average number of new daily cases was (2783.5 ± 2423) for the training and (5724.66 ± 6522.36) for the verification data, while the rates of confirmed positive cases per day were 0.041 ± 0.022 and 0.05 ± 0.052 for the training and verification datasets, respectively. The daily vaccine coverage and hospital beds per 1000 people, which are also presented in the table, showed that the average daily vaccine coverage was (26,234.2 ± 47,498.4) and (220,514 ± 332,466) for the testing and verification data, respectively. The hospital beds/1000 people were (28.25 ± 3.50) and (20.4 ± 0.5623) for the training and verification datasets, respectively.

Sensitivity Analysis
The careful selection of the most relevant factors to consider as input variables and the correct adjustment of connecting parameters (such as the hidden neurons, number of iterations, and transfer functions) for any AI-driven modelling are crucial steps required to obtain the optimum prediction level. The number of new cases, rate of positive cases, newly vaccinated individuals, number of cardiovascular diseases (CVD), stringency index, GDP per capita in USD, number of smokers, the prevalence of diabetes mellitus (DM), Hospital beds/1000, and population density were selected from the dataset for the sensitivity analysis. Previously, linear sensitivity analytical approaches have been used to select dominant input variables. However, due to the complex non-linear nature of the COVID-19 data, we had to conduct a sensitivity analysis with a non-linear nature. Hence, the nonlinear FFNN was conducted to select the dominant input variables for the modelling of COVID-19 in the area.
The sensitivity analysis conducted in this study is presented in Table 3. Accordingly, the four best dominant input variables with the highest DCs selected were the positive rate (first-ranked), hospital beds/1000 (second-ranked), new cases (third-ranked), and the number of vaccinated individuals (fourth-ranked) based on their chronological order. However, the inputs with the lowest DCs (<0.5) were removed in the modelling process.

Single AI-Driven Black-Box Models
The ANFIS, SVM, FFNN, and MLR were trained and tested for each combination of input variables in the modelling process. Hence, results from each model are presented in Table 4. The FFNN was the first type of AI-driven model used in this study. The Levenberg-Marquardt technique was used to train this model, which had four inputs and one hidden layer, to estimate COVID-19 mortality in East Africa. Identifying the optimal structure (number of hidden neurons) of the model was a key step in obtaining the best results in the FFNN modelling. The possession of too many neurons may result in overfitting, or too few neurons may result in incorrect information. To determine the appropriate structure of the FFNN model, a trial-and-error technique was used. Additionally, this allowed us to analyze the accuracy of the numerous models trained with the variable's hidden number. As a result, the best model structure (x-y-z) with the greatest prediction outcomes was discovered to be six hidden neurons with four inputs and one hidden layer, which was noted as (4-6-1).
The second type of AI-driven model used in this study was the ANFIS, which assumes a fuzzy notion to manage the unpredictable circumstances of complicated data of a non-linear nature. In the modelling process, Sugeno's fuzzy inference system with hybrid algorithms was used to calibrate the parameters of the MFs. The Gaussian, triangular, and trapezoidal MFs were investigated using a trial-and-error approach to produce the best estimation result in predicting mortality due to COVID-19. As a result, the ANFIS model with "Gaussian membership functions" trained over 41 epochs offered better prediction results than the other MFs.
SVM was the third type of AI-driven model used in this study. The kernel of the radial basis function (RBF) was used to generate the SVM model for the combination of all the input variables. RBF was chosen because it has fewer turning parameters and performs better than sigmoidal and polynomial models [42]. Finally, the traditional MLR technique was employed to predict the COVID-19 mortality and to compare the predicted result with those of the other three types of AI-driven models. The linear connection (a-b) between the one target variable and the four input variables was determined using this model and noted as (4-1).
The results of the single black-box models in Table 4 show that the ANFIS was the bestperforming AI-driven model in predicting mortality due to COVID-19, with the highest DC (0.9273) and lowest RMSE (0.000125) at the verification stage. The second, third, and fourth best models, based on their performance, were FFNN, SVM and MLR, respectively. The daily COVID-19 data are non-linear and dynamic by nature. Hence, the non-linear AIdriven model, ANFIS, was found to be the best model for predicting the data. However, according to the calculated DCs, the MLR was the worst-performing model in predicting the mortality data. These results showed that the best models for predicting data of a non-linear and dynamic nature are the non-linear AI-driven models, such as ANFIS and FFNN, while the linear regression estimation approach was a poorly performing model in predicting the mortality due to COVID-19 in the study area.
According to the findings of the single black-box models, provided in Table 4, utilizing the best-predicting model in this study (ANFIS) might improve the performance of the prediction using FFNN, SVM, and MLR by 7%, 8%, and 13%, respectively. In addition to these statistics, scatter plots and Taylor diagrams were created to depict the performances of the single AI-driven models. Figure 7 depicts the correlation between the estimated values obtained from several AI-driven models and the observed value using a scatter plot diagram. In this diagram, the estimated performances of the ANFIS, FFNN, SVM, and MLR models are compared in terms of their predictions of COVID-19 in East Africa. As a result, the ANFIS model indicated fewer spread points in the linking and produce better estimated values than the other models. This might be attributed to ANFIS's capacity to anticipate non-linear data, such as COVID-19 data, as it has a greater DC than the other AI-driven and MLR models. The finding from the diagram supports those of the other analyses and modellings, which showed that the non-linear predicting approaches performed better than the linear predicting approaches. Moreover, the ANFIS was the highest-performing model in predicting COVID-19 in the study area. In this study, we understood that the single black-box models and the scatter plot diagram proved that the non-linear predicting approaches performed better than the linear approaches. More specifically, among all the models, the ANFIS model was the bestperforming predicting approach for the daily COVID-19 data. This finding is similar to the finding of a study conducted on daily suspended sediment load data using the AI ensemble model [43].

Ensemble Models
The AI-driven ensemble model was developed using the estimated outputs from three single AI-driven models (ANFIS, FFNN, and SVM) and one classical regression model (MLR) as the input variables for the ensemble modelling. This model was developed to boost the efficiency of the single models in terms of the prediction capability. Four ensemble approaches (SAE, WAE, ANFISE, and NNE), as novel ensemble processes, were developed to predict COVID-19 in east Africa, and the results are presented in Table 5. Accordingly, the (a-b) structure for the SAE was applied to display the numbers of outputs and inputs used for the prediction. The structure (a, b, c, d) was the structure for the WAE that denoted the weights of the FFNN, ANFIS, SVM, and MLR single models, respectively.  In this study, we understood that the single black-box models and the scatter plot diagram proved that the non-linear predicting approaches performed better than the linear approaches. More specifically, among all the models, the ANFIS model was the bestperforming predicting approach for the daily COVID-19 data. This finding is similar to the finding of a study conducted on daily suspended sediment load data using the AI ensemble model [43].

Ensemble Models
The AI-driven ensemble model was developed using the estimated outputs from three single AI-driven models (ANFIS, FFNN, and SVM) and one classical regression model (MLR) as the input variables for the ensemble modelling. This model was developed to boost the efficiency of the single models in terms of the prediction capability. Four ensemble approaches (SAE, WAE, ANFISE, and NNE), as novel ensemble processes, were developed to predict COVID-19 in east Africa, and the results are presented in Table 5. Accordingly, the (a-b) structure for the SAE was applied to display the numbers of outputs and inputs used for the prediction. The structure (a, b, c, d) was the structure for the WAE that denoted the weights of the FFNN, ANFIS, SVM, and MLR single models, respectively. The ANFISE was best-performing among all the ensemble model development combinations. This is due to its resilience in integrating both the fuzzy concept and the artificial neural network capability, which provided the present ANFIS framework. The NNE, WAE, and SAE were the second-, third-, and fourth-best predictors of COVID-19 in the study area. The weighted ensemble technique outperformed the simple average ensemble approach. This is because the WAE assigns weights to parameters depending on their relevance.
The Levenberg-Marquardt method was used to train the NNE model, as it applied to the FFNN, and the tangent sigmoid activation function was utilized for the hidden and output layers. The study conducted by Sahoo et al. [44] indicated that the FFNN approach has the fastest convergence ability; hence, it was used more often in this study than the other ANN training techniques. We used a trial-and-error method to determine the correct number of hidden layers and the optimal epoch number. In ANFISE, Sugeno's fuzzy inference system, using a hybrid training approach, was used to calibrate the membership function parameters comparable to those of the ANFIS single black-box model. As a result, the ANFISE greatly improved the accuracy of the single models.
The comparison of the prediction performances of the ensemble models and single AIdriven models at the verification and training phases is presented in Table 6. In this table, the NNE boosted the predicting performance of the single models FFNN, ANFIS, SVM, and MLR by 5.6, 2.1, 7.1, and 13.4 per cent, respectively. In addition to this, the ANFISE boosted the performance of FFNN, ANFIS, SVM, and MLR by 13, 6.1, 13.9, and 19.3 per cent, respectively. These numbers show that the capacity for the prediction of COVID-19 was increased in the case of the ensemble models rather than the single models, and these findings were compared to the findings of studies conducted in different fields using AI ensemble models [6,14,35,37]. Hence, these findings showed that ensemble models can be applied to the prediction of COVID-19 in the eastern Africa region more effectively than the single AI-driven models. In addition to this, the findings prove that the non-linear ensemble models are more capable than the linear ensemble models. This might be due to the incapability of linear ensemble approaches to undergo another black-box learning process, unlike the non-linear approaches.  According to Taylor K.E. (2001), we can summarize multiple models' performances in a single diagram that can help us to easily visualize and understand which model performs better [45]. Therefore, four ensemble approaches were assessed using a two-dimensional Taylor diagram, as presented in Figure 8, which coordinates the correlation coefficients (r) and the standard deviations (sd) for both the observed and predicted values of the ensemble models (ANFISE, NNE, WAE, and SAE). The advantage of using the Taylor diagram is that it combines the predicting performances of different models in a single visual display that quantifies the level of resemblance between the observed and the predicted values. It is observed from Figure 8 that the ANFISE was the best approach in predicting COVID-19 in the eastern Africa region, with (r = 0.9852 and sd = 0.0523), and the SAE was the poorestperforming ensemble approach, with (r = 0.9073 and SD = 0.0821).
Diagnostics 2022, 12, x FOR PEER REVIEW visual display that quantifies the level of resemblance between the observed and t dicted values. It is observed from Figure 8 that the ANFISE was the best approach dicting COVID-19 in the eastern Africa region, with (r = 0.9852 and sd = 0.0523), a SAE was the poorest-performing ensemble approach, with (r = 0.9073 and SD = 0.0 In addition to the statistical evidence, the Taylor diagram showed that the 'r' coordinate of the ANFIS ensemble approach was closer to the observed value than t of the ensemble approaches, and we can see that the coordinate for the SAE was fa the observed value compared to the other ensemble approaches. This closeness s that the predicted values obtained from the ANFISE were more closely related to served value. Hence, this proves that this ensemble model has the best prediction bility among the other models.

Conclusions
In this work, the capacity of AI-driven models to predict mortality due to CO in east Africa was investigated. Before predicting the COVID-19 mortality using sin In addition to the statistical evidence, the Taylor diagram showed that the 'r' vs. 'sd' coordinate of the ANFIS ensemble approach was closer to the observed value than the rest of the ensemble approaches, and we can see that the coordinate for the SAE was far from the observed value compared to the other ensemble approaches. This closeness showed that the predicted values obtained from the ANFISE were more closely related to the observed value. Hence, this proves that this ensemble model has the best prediction capability among the other models.

Conclusions
In this work, the capacity of AI-driven models to predict mortality due to COVID-19 in east Africa was investigated. Before predicting the COVID-19 mortality using single AI-driven models, the data were normalized, and a sensitivity analysis was performed to identify the best dominant input variables. When comparing the results from the single AI-driven models, ANFIS outperformed the other models due to its ability to analyze nonlinear, dynamic, and complicated processes using the fuzzy concept and neural network idea. Four ensemble techniques were modelled to improve the performance of the single AI-driven models by aggregating the results from each AI-driven model and using the aggregated result as an input for the ensemble modelling. Because of their capacity to handle unpredictable, non-stationarity, and complicated data, the non-linear ensemble techniques (ANFISE and NNE) outperformed the linear ensemble approaches (SAE and WAE). AN-FISE was the best-performing ensemble technique, improving the prediction performance of the single AI-driven models by 13, 6.1, 13.9, and 19.3 percent, respectively.
Overall, the outcome of this study demonstrated the potential capacity of ensemble models to predict mortality due to COVID-19. The result obtained from the ANFIS ensemble model demonstrated that aggregating the outputs of separate AI-driven models leads to a better prediction than employing them individually. The study's weakness was that it only used black-box models to calculate the COVID-19 mortality. As a result, the use of physically based models in the assembly process should be investigated in future research. Furthermore, due to data limitations, this study used two years of daily COVID-19 mortality data. Therefore it is necessary to test these AI ensemble techniques for further observations in future studies.