key: cord-0748003-t8xy0r58
authors: Akinnuwesi, Boluwaji A.; Fashoto, Stephen G.; Mbunge, Elliot; Odumabo, Adedoyin; Metfula, Andile S.; Mashwama, Petros; Uzoka, Faith-Michael; Owolabi, Olumide; Okpeku, Moses; Amusa, Oluwaseun O.
title: Application of intelligence-based computational techniques for classification and early differential diagnosis of COVID-19 disease
date: 2021-12-09
journal: Data Science and Management
DOI: 10.1016/j.dsm.2021.12.001
sha: cc77479060f7991349611c619555da3d7ab30801
doc_id: 748003
cord_uid: t8xy0r58

Clinical methods are used for diagnosing COVID-19 infected patients, but reports posit that, several people who were initially tested positive of COVID-19, and who had some underlying diseases, turned out having negative results, after further tests. Therefore, the performance of clinical methods is not always guaranteed. Moreover, chest X-ray image data of COVID-19 infected patients are mostly used in the computational models for COVI-19 diagnosis while the use of common symptoms such as “Fever, Cough, Fatigue, Muscle aches, Headache etc”, in computational models is not yet reported. In this study, we employ seven classification algorithms to empirically test and verify their efficacy when applied to diagnose COVID-19 using the aforementioned symptoms. We experimented with logistic regression (LR), support vector machine (SVM), naïve Byes (NB), decision tree (DT), multilayer perceptron (MLP), fuzzy cognitive map (FCM) and Deep neural network (DNN) algorithms. The techniques were subjected to random under-sampling and over-sampling. Our results showed that with class imbalance, MLP and DNN outperform others but without class imbalance MLP, FCM and DNN outperform others with the use of random under-sampling but DNN has the best performance with the use random oversampling. This study identified MLP, FCM and DNN as better classifiers over LR, NB, DT and SVM, that healthcare software system developers can adopt to develop intelligence-based expert systems which both medical personnel and patients can use for differential diagnosis of COVID-19 based on the aforementioned symptoms however, the test of performance must not be limited to the traditional performance metrics.

The weekly epidemiological reports on COVID-19 from the World Health Organization (WHO) shows that COVID-19's confirmed cumulative cases are above 240 million with over 4.9 million deaths, as of October 23, 2021, globally (World-Health-Organization, 2020a). New cases are always reported due to the advent of different SARS-CoV-2 (Severe acute respiratory syndrome coronavirus 2) variants despite the ongoing vaccination programme across several countries. SARS-CoV-2 variants such as Cluster 5, SARS-CoV-2 VOC 202012/01, 501Y.V2, and P.1, have been reported in countries such as Denmark, United Kingdom, Brazil, Nigeria, and South Africa. There are possibilities that the variants could weaken the strength of body immune system that is built naturally or due to vaccination and hence reduce neutralization of the effects of the virus in human (World-Health-Organization, 2020a , 2020b . . Hence, the spread of the disease and there is a need to increase the diagnostic capacity and systematic sequencing of SARS-CoV-2. As a result, many clinical methods have been developed to diagnose SARS-CoV-2. However, there have been reports on false-negative rates that ranged from 5%-40% (Arevalo-Rodriguez et al., 2020; D. R. Long et al., 2020; Weissleder, Lee, Ko, & Pittet, 2020) .

Studies conducted in China and Singapore reported few cases of false-negative results to the tone of 29% and 11% respectively (Fang et al., 2020; Lee et al., 2020) . This implies the clinical methods do not guarantee accurate diagnosis or classification of COVID-19 and it could be attributed to the type and quality of the specimen obtained, the duration of illness at the time of testing, and the specific assay. Thus, the need arises for scientists to complement the clinical diagnostic methods with computational intelligence methods in order to ensure near to 100% accurate diagnosis and classification.

In addition, COVID-19 and other diseases such as Malaria, HIV/AIDS, Tuberculosis, Flu and Pneumonia have similar symptoms and this make differential diagnosis of COVID-19 at the early stage inevitable, particularly when the infection is very mild and it cannot be established whether the infection is that of Malaria, HIV/AIDS, Tuberculosis, Flu or Pneumonia. Few studies have been reported on the differential diagnosis of COVID-19 using chest CT (Computed Tomography), a medical imaging technique and some of the studies were presented in (C. Long et al., 2020; Murthy et al., 2020; Zeng et al., 2020; Zuo, 2020) . We deduced that the computational models presented in the studies focused more on the use of chest X-ray images and not on other regular symptoms such as: "Fever or chills, Cough, Shortness of breath or difficulty breathing, Fatigue, Muscle or body aches, Headache, loss of taste or smell, Sore throat, Congestion or runny nose, Nausea or vomiting, and Diarrhea". Moreover, the models are not suitable for early diagnosis when the symptom presented are still very mild and chest CT will not be able to show a convincing result in the X-ray images of the lungs and the cardiovascular system. No study known to the authors, has used the aforementioned regular symptoms for the development of intelligence-based systems for early differential diagnosis of COVID-19 using classification algorithms of soft computing, machine learning, deep learning, expert system, and decision support system (DSS) using on non-image-based dataset of COVID-19 infected patients. However, chest X-ray image data of COVID-19 infected patients are mostly used for diagnosis using deep learning algorithm such as convolutional neural network (CNN). Therefore, we were motivated to carry out an experimental study of the various intelligencebased classifiers that could be applied in early differential diagnosis of COVID-19 disease considering the aforementioned regular symptoms and using non-image datasets of the symptoms. Our specific objective is to use the available COVID-19 non-image diagnostic dataset to test the performance of the intelligence-based classifiers (i.e. fuzzy cognitive map (FCM), support vector machine (SVM), logistic regression (LR), multilayer perceptron (MLP), naïve Bayes (NB), decision tree (DT) and deep neural network (DNN)) in terms of accuracy, Fmeasure, recall, precision, balanced accuracy, Mathews correlation coefficient (MCC) and bookmaker informedness (BM) . The goal is to identify the classifier(s) with best performance in early differential diagnosis of COVID-19. This paves way for future work in this area to developing diagnostic systems that could apply the best performing classifier(s) for differential J o u r n a l P r e -p r o o f and early diagnosis of COVID-19 using the aforementioned regular symptoms. In the light of this we have the following research questions (RQ): a. RQ-1: What are the classifiers that are most appropriate for differential diagnosis of COVID-19? b. RQ-2: What is the performance of each of the classifiers identified in (a) above in terms of accuracy, precision, recall, F-measure, Mathews correlation coefficient, balanced accuracy, and bookmaker informedness?

This paper contributes to existing literature by presenting the behaviours of the listed classifiers when applied for COVID-19 diagnosis using non-image based COVID-19 datasets on the regular symptoms (e.g., fever, headache, vomiting, Diarrhea etc) and this is not yet reported in literature as at the time of this report. Earlier studies focus more on image-based chest X-ray data of COVID-19 patients for diagnosis and the mostly used classifier is DNN.

The rest of the paper is organized as follows: Section 2 presents the overview of clinical methods that have been used for diagnosing COVID-19, a detailed discussion on differential diagnosis and a review of related works on the differential diagnosis of COVID-19 disease. The strengths and weaknesses of the diagnostic methods are presented. Section 3 presents materials and methods. Results and discussion are presented in Section 4 wherein detailed discussion and analysis of the aforementioned intelligence-based techniques are done. The conclusion and managerial implication are presented in Section 5.

Several clinical diagnostic methods have been developed for COVID-19. However, the use of intelligence-based computational methods to tackle COVID-19 cannot be underestimated. It will be very helpful to complement the clinical diagnostic methods with other intelligence-based computational methods in order to increase the balanced accuracy, BM and MCC. This section presents a brief description of differential diagnosis and the qualitative features of the clinical and computational methods that have been used for COVID-19 diagnosis vis-à-vis their limitations in the current time.

Differential diagnosis is the defined process that helps to differentiate between diseases with similar symptoms and risk factors. It is a systematic diagnosis process carried out on patients with the view to accurately diagnose a disease that shares the same symptoms with other related diseases and also survive under the same conditions (Mann, 1990; Sand, 2015; Uzoka et al., 2016) . Differential diagnosis is required because rarely do physicians diagnose a disease with certainty, directly from the presentation alone, especially in cases where the symptoms presented relate to many diseases (Jain, 2017) . For example, diseases like HIV/AIDS, Malaria, Flu, Tuberculosis, COVID-19, Ebola virus disease (EVD), Cholera etc. have some similar symptoms. Thus, when patients exhibit one or more of these symptoms, the physicians need to subject them to a differential diagnostic process with the view to establishing the actual disease in the multiple related diseases.

The differential diagnosis process entails weighing the probability of a disease against the probabilities of other related diseases that possibly account for a patients' illness. Differential diagnosis has shown its usefulness in many instances in medicine. It has helped save lives that would have been otherwise lost and we have seen many lives nearly lost because of the ignorance of differential diagnosis. A notable case of a life almost lost due to the ignorance of differential diagnosis is that of a little girl by the name of Isabel as reported in (Rutledge, 2017) . The girl who was diagnosed with chickenpox was misdiagnosed by her family doctor thus, leading to wrong treatment, having her spend about 60 days in the hospital and almost losing her life to multiple organ failures and cardiac arrest, all from a misdiagnosis that could have been otherwise avoided, with differential diagnosis, in order to attain the proper diagnosis, the first time. Though differential diagnosis consumes time, its help in making sure doctors get the right diagnosis and treatment is undeniable.

The cause of COVID-19 is attributed to SARS-CoV-2. Many health complications that are related to the virus have been established in the literature to include the following-failure of respiratory system, complications in cardiac and cardiovascular system, inflammatory complications, thromboembolic diseases, neurological disorders, and multi-organ dysfunctions (Alpdagtas et al., 2020; McIntosh, 2020) . Researchers are working tirelessly on various clinical methods of diagnosing and treating COVID-19, in order to reduce cases of identified health complications. Thus, some of the clinical diagnostic methods employed are chest radiographs, chest computed tomography (Chest CT), Point-of-care lung ultrasonography, clinical suspicion, nucleic acid amplification tests (NAATs), Point-of-care NAATs (Caliendo & Hanson, 2021) .

Computed Tomography diagnosis is generally used for auxiliary diagnosis of the SARS-CoV-2 and the diagnosis is confirmed by positive results of a nucleic acid amplification test (NAAT) of the respiratory tract and or blood specimens using an rRT-PCR (reverse real-time PCR assay) reaction (Dai et al., 2020) . However, it was reported that the rRT-PCR method of diagnosis has limitations. The implication is that when the viral load is low, the detection rate is also low. This consequently leads to the occasional occurrence of false results. Similarly, with the use of this method, the severity and progression of the virus cannot be known. Despite these few disadvantages, the rRT-PCR method of diagnosis has been performing well since the outbreak of the coronavirus. However, the number of people tested is not yet satisfactory. This can be attributed to the fact that this method of diagnosis is relatively costly.

The authors in (Lieberman et al., 2020; Nalla et al., 2020) It was reported that false-negative rates ranged from 5%-40% (Arevalo-Rodriguez et al., 2020; D. R. Long et al., 2020; Weissleder et al., 2020) . The authors in (Fang et al., 2020) reported cases in China where fifty one patients with fever or acute respiratory symptoms, ultimately tested positive using the SARS-CoV-2 RT-PCR test, but in the initial test carried out, 15 patients (29%) had negative results. Similar study in Singapore, where 70 patients were tested positive, the initial nasopharyngeal testing was negative in eight patients (11%). These studies established that few patients were repeatedly negative in the initial testing but they later tested positive after rounds of four or more tests. This implies that there are occurrences of falsenegative results and therefore, some rounds of testing are recommended, in order to confirm the accuracy of the results.

In (Alpdagtas et al., 2020) , various clinical diagnostic methods for COVID-19 were evaluated with emphasis on their pros and cons, vis-à-vis their performances. Immunological and RT-PCR testing methods were identified as the best diagnostic methods for COVID-19. However, the knowledge of these diagnostic methods lies only with professionally trained physicians. Nonprofessional personnel cannot apply it on their own. Neither can a patient. Thus, the authors recommended the production of POC (Point-of-Care) diagnostic devices that can diagnose with little or no support from physicians. In addition, accuracy is not always guaranteed using the diagnostic method, hence there is a need for the development of better methods with a higher level of practicality, accuracy and precision. Similarly, in (Uygun-Can & Acar-Bolat, 2020), RT-PCR and CT testing methods were used to detect COVID-19 in pregnant women and it was established by the authors that the combination of the two testing tools gave an accurate and safe diagnosis. This implies that a combination of two or more clinical methods for COVID-19 diagnosis guarantees better accuracy and precision.

Other examples of clinical diagnostic methods have been used to diagnose COVID-19 and are reported in literature are: Nuclear acid test, Computed tomography, Immunological Examinations, Lung Ultrasound, F-FDG PET/CT (Ardakani, Kanafi, Acharya, Khadem, & Mohammadi, 2020; Mertens et al., 2020; Wan, Luo, Dong, & Zhang, 2020; Xie et al., 2020) ; RT-PCR, POC, Immunoassays for antibody to virus (Giri et al., 2020; He et al., 2020; Wu et al., 2020) ; POC, multiplex assays, CT imaging, genome sequencing, and electron microscopy, PCR Udugama et al., 2020) ; Oligonucleotide-based molecular detection, Point-of-Care immunodiagnostics, radiographical analysis/sensing system, and biosensing prototypes, RT-PCR (Mahapatra & Chandra, 2020) ; Clinical computer-aided diagnosis (CAD) using machine learning algorithms (Ardakani, Acharya, Habibollahi, & Mohammadi, 2021) ; Artificial intelligence (AI) enabled medical imaging (Yuan et al., 2020) ; Molecular-based assay, POC, rRT-PCR (Yang, Wang, Shen, & Cheng, 2020) .

Our deductions are as follows:

a. Most of the studies focused on measuring the accuracy of the clinical methods for COVID-19 diagnosis. b. The studies provided information that helps to guide health professionals to conduct errorfree COVID-19 diagnostic tests. Similarly, the information guides the researchers to identify limitations with each clinical method and hence develop better clinical methods for detecting COVID-19 cases. c. The works affirmed that the use of a single clinical method to detect COVID-19 do not guarantee accurate results as expected. However better performance is always witnessed when two or more of the clinical methods are combined in the process of diagnosing COVID-19 cases. Thus, health personnel were advised to explore the combination of methods in the process of COVID-19 diagnosis. d. Clinical methods mostly used are RT-PCR, POC, Molecular-based assay and Chest CT. e. Test sensitivity is likely influenced by specimen quality, illness duration and specific assay.

The limitations that are common to all the clinical methods are as follow:

a. Accuracy is not always guaranteed with the use of one clinical method for diagnosis.

Therefore, there is a need to carry out an alternate authorized test to confirm test sensitivity and specificity. The combination of two or more methods of testing helps to confirm the accuracy of the test results. b. They are laboratory tests that requires well equipped laboratories. Hence the establishment of the test laboratories is expensive and the test is not cheap and it takes some time. c. The testing machines are relatively expensive (e.g., PCR machines) and few laboratories could afford them. d. Physicians/patients need to travel to a competent laboratory to access the PCR machine and this may take some hours or days depending on location of the patients and the nearest available COVID-19 test laboratory. Hence, results are not received immediately. e. The methods require well trained medical personnel to run the tests in the laboratories, however there is dearth of qualified medical personnel for the laboratories and the few personnel available are overstretched and fatigued due to increasing number of cases of COVID-19. Hence there are possibilities of false-negative results in some cases. f. Performance is not optimal with the use of the clinical diagnostic methods and hence, the need to apply artificial intelligence-based systems using computational algorithms, help to ensure better performance in term of balanced accuracy, precision, specificity and sensitivity, bookmaker informedness (BM), and Mathews correlation coefficient (MCC).

In this sub-section we did a review of past works that have applied computational intelligencebased algorithms for diagnosing COVID-19 with the view to identify the patients' symptom used, performances, the metrics used for measuring the performances and their limitations.

Convolutional Neural Network (CNN) was used in (Mahmud, Rahman, & Fattah, 2020 ) to develop CovXNet model that was used to detect COVID-19 and pneumonia. The model had an accuracy value of 97.4%. Similarly COVIDScreen was developed in (Singh, Pandey, & Babu, 2021) using CNN. The model helped to carry out a differential diagnosis of COVID-19 with an accuracy of 98.67%. In the same light, SVM was applied in (Jin, Dong, Dong, & Ye, 2021) for examining chest X-ray radiograph with the view to carry out differential diagnosis of COVID-19. The accuracy recorded was 98.642%. In China a team of researchers developed LR based model to identify the independent predictors of COVID-19 severity in suspected cases (Xu et al., 2020) ; a similar research was also reported in (Iwendi et al., 2020) where RF (Random forest) model was used for predicting severity of COVID-19 in infected patients. In the same manner, LR was used to predict mortality risk in COVID-19 patients and the accuracy was measured at 70% (Bhandari et al., 2020) . Also, RF algorithm was applied to predict the mortality of COVID-19 patients and the accuracy of 95% was obtained. Also (Fleitas et al., 2020) applied multivariate LR to identify COVID-19 symptoms and infected cases of COVID-19 were detected with specificity of 46%. In (Silahudin & Holidin, 2020 ), an expert system was developed for diagnosing COVID-19 using the Naïve Bayes (NB) technique. Similarly Naïve Bayes decision support system (DSS) was presented in (Awwalu, Umar, Ibrahim, & Nonyelum, 2020) for COVID-19 detection. Fuzzy cognitive map (FCM) was applied in (Groumpos, 2020) to examine the whole spectrum of COVID-19 by considering the causality factors. The authors could not guarantee the performance of the model because real-life data required were not available, however the model was tested using data that were generated from literature. Similarly CNN and LR were used to develop CovNet30 system that was used to automatically diagnose COVID-19 in (Gour & Jain, 2020) . CovNet30 operate with 92.74% classification accuracy and 93.33% sensitivity. Decision tree (DT) and CNN was also used to build a classifier for detecting COVID-19 in (Yoo et al., 2020) and the classifier works with an accuracy of 95%. CoroNet was proposed in (Khan, Shah, & Bhat, 2020; Oh, Park, & Ye, 2020 Our deduction is that the computational models focus more on the use of chest X-ray image data for diagnosing COVID-19 and not on other regular symptoms such as: "Fever or chills, Cough, Shortness of breath or difficulty breathing, Fatigue, Muscle or body aches, Headache, loss of taste or smell, Sore throat, Congestion or runny nose, Nausea or vomiting, and Diarrhea". The models are not useful at the early stage of COVID-19 infection when the aforementioned regular symptoms are mild and it appears that one is infected with flu or malaria. At this stage the chest X-ray images may not present the expected results because the respiratory/cardiovascular system have not been distorted as expected. As of the time of this study, the authors have not found any study that has applied any of the classifiers or combination of the classifiers using the aforementioned regular symptoms for COVID-19 diagnosis.

FCM is a knowledge representation algorithm that makes use of fuzzy graph structure to present the causal relationship between concepts and hence, present the causal values between the concepts. The relation between the concepts helps to calculate the extent to which pairs of concepts impact each other. COVID-19 is associated with several symptoms and risk factors (i.e. concepts) that impact one another. The relationship and strength of pairs of symptoms/risk factors is defined based on the experiential knowledge of the medical doctors which is fuzzy in most cases. The doctors use their discretion, based on previous experiences. FCM can be found useful to solve the fuzziness problem, associated with the classification of COVID-19 patients. The signed and weighted arcs of the FCM graph depict the causal relationship that exists among the symptoms/risk factors and hence illustrate the interconnection between the symptoms/risk factor and how the symptoms/risk factors influence one another (Kosko, 1986; Papageorgiou & Stylios, 2008) .

The following features of FCM as presented in (Papageorgiou & Stylios, 2008) , makes it fit for classification and differential diagnosis of COVID-19: i. FCM is used for the acquisition and representation of causal knowledge, together with the causal knowledge reasoning process. This is needed to identify the causal strength of each of the COVID-19 symptoms relative to other symptoms. ii.

It is a neuro-fuzzy algorithm that can help to solve decision making problems such as the COVID-19 classification problem. The strength of FCM in cognitive decision making during medical diagnosis was presented in (Chandiok & Chaturvedi, 2016) . It was noted that FCM helps to represent the cognitive knowledge required in expert systems that are used for medical diagnosis. This is utilized for rational decision making and aids prediction.

iii.

Similarly, FCM supports differential diagnosis of diseases. The authors in (de Moraes Lopes et al., 2013) made use of FCM in the development of a DSS for differential diagnosis of alterations in urinary elimination. iv.

FCM's weights can be trained and updated using learning algorithms. Hence, it makes it possible to compute the weights of the relationship between all pairs of symptoms of COVID-19.

Support Vector Machine is a supervised machine learning (ML) technique that is applied in the classification and regression problems. Generally, it is best applied in the classification problems (Land & Schaffer, 2020; Pisner & Schnyer, 2020) . SVM algorithm can process datasets with multiple continuous and categorical variables. It labels data (i.e. input and output) for classification. It is a non-probabilistic, binary linear classifier. SVM based models are trained, using labelled data. It is a representation of diverse classes in a hyperplane in a multidimensional space. SVM generates the hyperplane in an iterative form with the view to minimize error. SVM focuses on dividing datasets into different classes in order to find a maximum marginal hyperplane (MMH). Thus, classification provides the basis for training the system for data processing. The support vectors are the data points, closest to the hyperplane while the hyperplane is a decision space that is divided between a set of objects having different classes. The algorithms achieve the best separation of data with the boundary around the hyperplane being maximized and even between both sides. The SVM classification algorithm is fast and dependable and it performs well with a limited amount of data to analyze.

The following are features of SVM that make it fit for classification and differential diagnosis of COVID-19: i. SVM is a binary linear classifier (Noble, 2006; Pisner & Schnyer, 2020) and it is best applied in two-group classification problems. In this light, SVM can be used to build learning models that could accurately classify COVID-19 patients into two groups (i.e., True Positive COVID-19 and True Negative COVID-19 patients) based on the COVID-19 dataset available. ii.

SVM is relatively simple and flexible to address classification problems (Pisner & Schnyer, 2020) . iii.

SVMs distinctly produce balanced predictive performance in cases where the sample data sizes are limited (Pisner & Schnyer, 2020) . This makes it suitable, even with limited COVID-19 datasets.

Decision Tree is applied in classification and regression problems. It embraces supervised learning method. Application of decision tree for classification of COVID-19 cases using X-ray imaging was reported in (Yoo et al., 2020) . Similarly, DT algorithm has been applied in (Atieh et al., 2019) for predicting peri-implant disease. Also, it has been used for the diagnosis of diabetic patients, as presented in (Kamadi, Allam, & Thummala, 2016) . Advantages of the DT are as follows (Gupta, 2017) : it is simple and easy to understand, interpret, and visualize; variable screening and feature selection are implicitly performed; it accommodates numerical and categorical data; it solves multi-output problems; relatively little effort is required for data preparation and the performance of the tree is not influenced by nonlinear relationships that are between parameters. However, DTs have the following shortcomings (Gupta, 2017) : overfitting problem; variance; needs to be lowered using bagging and boosting methods; creation of biased trees if some classes dominate; and optimal decision tree is not guaranteed in cases of greedy algorithms.

Naïve Bayes is an intuitive classification method that uses Bayes' theorem and assumed independence among predictors. Even though complete independence among predictors is impossible in real-life situations (Dinant, 2018; Ray, 2017) . It uses supervised learning method. NB is used with large datasets. It is simple, fast, easy to implement and the classification accuracy is better compared to other classification algorithms, most especially when the assumption of independence predictor holds. There are Gaussian, Multinomial and Bernoulli types of NB. It takes less time for training because small training data is required for NB to estimate the test data. It linearly scales with the number of predictor features and data points. It is applied to binary and multi-class classification problems. NB model was used to diagnose COVID-19 patients in (Mansour et al., 2021) . Similarly, Silahudin and Holidin in (Silahudin & Holidin, 2020) modeled expert systems using the NB technique to diagnose COVID-19 and Awwalu, Umar, Ibrahim, and Nonyelum in (Awwalu et al., 2020 ) developed a DSS for COVID-19 diagnosis using a multinomial NB algorithm.

A MLP is a feedforward ANN that has a minimum of three layers of nodes or neurons (i.e. input layer, hidden layer and output layer). Each neuron makes use of non-linear activation function except the input neurons. MLP embraces backpropagation algorithm for training. The wellarranged network of neurons helps in data modelling with the use of machine learning algorithms and hence facilitate accurate processing vis-à-vis accurate decision making. MLP is trained to carry out a given task and MLP based models help to analyze data and hence recognize patterns within the data. The merits of using MLP are as follows: models complex and non-linear problems; performs very well with large data; it is capable of generalization; its fault tolerance is high; and good for pattern recognition. In the light of aforementioned merits, MLP has been adopted to develop models for diagnosing COVID-19 with remarkable performance. For example MLP was applied in (Salman, Abu-Naser, Alajrami, Abu-Nasser, & Alashqar, 2020) for X-ray images classification to detect COVID-19 in patients. Similarly COVID-19 early vision diagnosis using MLP was proposed in (Hammam, Elmousalami, & Hassanien, 2020) . Also MLP and LR were combined to develop a hybrid model for COVID-19 diagnosis in . The authors established good performance for the MLP based models in terms of accuracy, sensitivity and specificity. However, some limitations are attributed to MLP such as: Computational intensiveness and time consuming; problem of scaling; model performance dependence on quality of training; and hence solution not always guaranteed.

Logistic regression (LR) is class of supervised learning techniques that is applied in classification problems. It is used for predicting the probability of a binary based dependent variable. Thus, the dependent variable has data coded as either 0 (negative/no) or 1 (positive/yes). An example of logistic regression equation is presented in Equation (1) (Brownlee, 2016) : y = e^(b0 + b1*x) / (1 + e^(b0 + b1*x)) Equation (1) where: y: predicted output; b0: bias or intercept term; b1: the coefficient for the single input value (x). Each column in the input data has an associated b coefficient (a constant real value) that must be learned from the training data.

LR is a simple machine learning algorithms and it is well applied in medical diagnostic problems for the detection of diseases. Hence makes it a good algorithm for classification of COVID-19 patients. For example, LR was applied in (Roland, Gurrola, Loftus, Cheung, & Chang, 2020; Shang et al., 2020; Song et al., 2020) to predict the severity of COVID-19 in infected patients. Also LR based system was proposed in (Fink et al., 2020) for validation of the results of COVID-19 diagnostic prediction at the time of admission in the hospital.

Deep learning classifiers such as Recurrent Neural Networks (RNNs), Generative Adversarial Networks (GANs), Radial Basis Function Networks (RBFNs), Self Organizing Maps (SOMs), Deep Belief Networks (DBNs), Restricted Boltzmann Machines (RBMs) and Autoencoders are other algorithms that could be applied for disease diagnosis but we did not consider them in this study because they are not fit for our dataset. Rather they are most commonly applied to analyse visual imagery, time series data, natural-language processing, and machine translation (Biswal, 2021) .

The data were obtained from https://github.com/burakalakuss/COVID-19-Clinical and contains 600 entries of patient records.

In order to overcome overfitting and outliers, there is need for data pre-processing. Data preprocessing were performed using WEKA and Python library Scikit-learn (sklearn). In this study, we experimented with logistic regression (LR), support vector machine (SVM), naïve Byes (NB), decision tree (DT), multilayer perceptron (MLP), fuzzy cognitive map (FCM) and Deep neural network (DNN) algorithms by calibrating their parameters. The algorithms are applied to the resampled data to eliminate class imbalance (520 negative cases and 80 positive cases of COVID-19) using random under-sampling and over-sampling approach. The COVID-19 clinical dataset was divided into a training dataset (80%) and a testing dataset (20%). This study uses the training dataset to train the COVID-19 classification model and uses the testing dataset to measure model performance and to ensure that the results of the classification model is robust. According to (Elujide et al., 2021) , classification problems can be model using single-labelled and multi-label approach. This study employed a single-labelled with binary classification problem. For the deep neural network model, a two-layer neural network was implemented with the two COVID-19 indicators to classify the clinical COVID dataset.

The performance of the classification algorithms is measured based on the following performance metrics: accuracy, precision, recall, Mathew's correlation coefficient (MCC), balanced accuracy and bookmaker informedness (BM).

We did a quantitative analysis of the classifiers. The clinical dataset on COVID-19 contains 600 entries of patient records comprising of an ID, set of symptoms and status (label) features. This section answers: RQ-1: What are the classifiers appropriate for differential diagnosis of COVID-19?; RQ-2: What is the performance of each of the classifiers in terms of accuracy, precision, recall, F-measure, Mathews correlation coefficient, balanced accuracy, and bookmaker informedness?

We empirically tested and verified the efficacy of the following classification algorithms: LR, SVM, NB, DT, MLP, FCM and DNN. The algorithms were subjected to random under-sampling and over-sampling. The classification algorithms were subjected to random under-sampling in order to convert class imbalance to balanced class on the target variable (520 negative cases and 80 positive cases of COVID-19). We evaluated the performances of the classification algorithms using the following metrics: accuracy, precision, recall, F-measure, and Mathews Correlation Coefficient (MCC) while Balanced accuracy, bookmaker informedness (BM), MCC, accuracy, precision, recall, F-measure are considered for FCM based on percentage split (80% for training and 20% for testing). The COVID-19 dataset experiments were performed using python and R.

According to the authors in (Chicco & Jurman, 2020) , most data scientists and machine learning experts use confusion matrix to evaluate binary classification. Few studies recently reported that it is not adequate to base the performance evaluation results on class imbalance dataset on accuracy and F1 score, but also on MCC result because it puts the ratio between the positive and negative into consideration. The closer MCC is to +1, the better the binary classification but the closer MCC is to -1, the worse the binary classification. The authors in (Chicco, Tötsch, & Jurman, 2021) claimed that MCC is a more reliable performance metric to put into consideration over balanced accuracy, bookmaker informedness (BM) and markedness (MK) in evaluating binary classification.

We compared the results on the use of random under-sampling, integrated with the classification algorithms based on the use of percentage split. As shown in Table 1 , with class imbalance, MLP outperforms LR, NB, DT and SVM in accuracy, precision, recall, F-measure and MCC. Similarly, DNN has the best performance in accuracy, recall and F-measure. As shown in Table  2 , without class imbalance, MLP outperforms LR, NB, DT and SVM in accuracy, precision, recall, F-measure and MCC while DNN performed best with the use of oversampling. Furthermore, we compared the results on the dataset with a class imbalance on percentage split of 80% for training and 20% for testing with dataset without class imbalance (use of random under-sampling and oversampling are integrated on the classification algorithms to eliminate the class imbalance on the percentage split of 80% for training and 20% for testing) in python and with deep neural network with the following parameters: epoch of 150, batch_size of 10, loss= binary_crossentropy, and adam optimizer. Two-layer was added to a sequential model from Keras, with ReLu and Sigmoid activation functions.

Comparing Tables 1 and 2, it was discovered that the classification algorithms in Table 1 perform better with the traditional performance metrics for evaluation (i.e., accuracy, precision, recall, F-measure) but it was poor in performance with the MCC performance metric. The performance in table 1 is due to overfitting. While on Table 2 , their performances dropped for accuracy, precision, recall and F-measure. Therefore, our findings show that without class imbalance, the rate of improvement of performance with MCC as a performance metric is very high for all the classifiers while there is a drop in performances for other metrics. This aligns with the claims of Chicco & Jurman in (Chicco & Jurman, 2020) that MCC is the best performance metric to determine the best classifier, without class imbalance. We noticed in Table 1 , that all the performance metrics for machine learning algorithms was higher than 80% in the MLP and DNN and this makes them better classifier over others but the MCC is very far from 100% while in Table 2 , DNN is the best classifier even though the MCC value is close to 100%.

In addition, this report also summarizes the development of a FCM model that analyses COVID-19 symptoms and determines status as either positive or negative. The FCM model was developed with the twenty (20) concepts of the COVID-19 dataset. Weights for each relationship between concepts were extracted from a weight matrix which was in term extracted and trained from the dataset, using a simplified version of the Hebbian Learning algorithm given by equation 

The weight matrix and imported data frame of the data were then pre-processed by scaling column values in the [0, 1] as per the requirement of the parameters of the inference function of the FCM model. The dataset was split into 80% training set and 20% testing set, i.e., 480 records were used in developing and training the weight matrix and 120 used to test the dataset with class imbalance, while another experiment was also carried out on the dataset without class imbalance after applying the random under-sampling on it. The inference result from running the fcm.infer function in twenty-five (25) iterations were stored in an array. These results were compared with the actual labels of the test data. Sigmoid function was used in this study. Table  3 shows the comparative summary of the results. We compared the results on the fuzzy cognitive map for the dataset with class imbalance on percentage split of 80% for training and 20% for testing with dataset without class imbalance (use of random under-sampling integrated on the classification algorithms to eliminate the class imbalance on the percentage split of 80% for training and 20% for testing). It was discovered that the FCM without class imbalance outperforms the FCM with the class imbalance in all the performance evaluation metrics, except in MCC. Our findings, as presented in Table 3 , does not support the claims of Chicco, Totsch & Jurman (Chicco et al., 2021) which state that MCC performs better than balanced accuracy and bookmaker informedness. Our results in Table 3 reveal that balanced accuracy outperforms BM and MCC.

The increasing number of global cases of COVID-19 infection is alarming and the large number of infected people that flood the hospitals on daily basis has caused an overstretching of the available medical facilities as well as the healthcare workers. The patients require reliable and accurate diagnosis and treatments. There are reports that with the use of the clinical methods for COVID-19 diagnosis, some people who were initially tested positive for COVID-19 and who were having some underlying diseases, turned out having negative results after series of further tests. This implies that true positive and true negative result is not always guaranteed.

In this study, having analyzed the quality attributes of the clinical methods for diagnosing COVID-19, we further employed some classifiers which constitute intelligence-based methods, which could be applied for early diagnosis of COVID-19. The classifiers are considered to complement the clinical diagnosis methods. The classifiers employed for the experiment are Logistic Regression, Support Vector Machine, Naïve Bayes, Decision Tree, Multilayer perceptron, Fuzzy Cognitive Map and Deep Neural Network. We did experiments with our data to determine the best classification technique in terms of accuracy, precision, recall, F-measure, Mathews correlation coefficient, balanced accuracy, and bookmaker informedness, based on percentage split (80% for training and 20% for testing). Our results showed that with class imbalance, MLP and DNN outperform LR, NB, DT, SVM and FCM in accuracy, precision, recall, F-measure and MCC (See Tables 1) but without class imbalance, MLP, FCM and DNN outperform LR, NB, DT and SVM in Accuracy, Precision, Recall, F-measure, MCC, BM, and Balanced Accuracy (See Tables 2 and 3) .

It is worth noting that with class imbalance, their performances are poor using the MCC performance metric, but without class imbalance, MCC values increases for all the classifiers while there is a drop in performances for other metrics. This agrees with the claims of Chicco & Jurman in (Chicco & Jurman, 2020 ) that MCC is a good performance metric to determine the best classifier without class imbalance. We also discovered that without class imbalance, FCM performs better for all the performance metrics, compared to the performance with class imbalance. The only exception is with MCC, whose value dropped without class imbalance (See Table 3 ). Thus, MCC performs better than balanced accuracy and bookmaker informedness.

Considering the relatively good performance values of the tested classification algorithms (i.e., MLP, LR, NB, DT, SVM, FCM and DNN), intelligence-based systems can be developed for COVID-19 diagnosis using these algorithms, but our results present MLP, FCM and DNN as better algorithms that healthcare system developers should adopt for early differential diagnosis of COVI-19. Moreover, the test of performance must not be limited to the classic performance metrics for evaluation (i.e., accuracy, precision, recall, F-measure), other performance metrics such as MCC, balanced accuracy and bookmaker informedness must also be considered. In the light of this, our future works will be considering the development and implementation of a COVID-19 smart medical diagnostic system (i.e., C-19-SmartMed) that will be able to differentially diagnose patients for COVID-19 based on the common symptom (i.e., Fever or chills, Cough, Shortness of breath or difficulty breathing, Fatigue, Muscle or body aches, Headache, loss of taste or smell, Sore throat, Congestion or runny nose, Nausea or vomiting, and Diarrhea) and using classifiers such as MLP, FCM and DNN. The system is envisaged to complement the effort of the physician.

Comparison of deep learning approaches to predict COVID-19 infection

Evaluation of current diagnostic methods for COVID-19

COVID-19 detection from xray images using different artificial intelligence hybrid models

The 12 deadliest viruses on Earth

COVIDiag: A clinical CAD system to diagnose COVID-19 pneumonia based on CT findings

Application of deep learning technique to manage COVID-19 in routine clinical practice using CT images: Results of 10 convolutional neural networks

False-negative results of initial RT-PCR assays for COVID-19: a systematic review

Predicting peri-implant disease: Chi-square automatic interaction detection (CHAID) decision tree analysis of risk indicators

A Multinomial Naïve Bayes Decision Support System for Covid-19 Detection

Logistic regression analysis to predict mortality risk in COVID-19 patients from routine hematologic parameters

Top 10 Deep Learning Algorithms You Should Know in 2021

Logistic Regression for Machine Learning

Coronavirus disease 2019 (COVID-19): Diagnosis. Retrieved

MH-COVIDNet: Diagnosis of COVID-19 using deep neural networks and metaheuristic-based feature selection on X-ray images

Cognitive decision support system for medical diagnosis

The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation

The Matthews correlation coefficient (MCC) is more reliable than balanced accuracy, bookmaker informedness, and markedness in two-class confusion matrix evaluation

CT imaging and differential diagnosis of COVID-19

Fuzzy cognitive map in differential diagnosis of alterations in urinary elimination: a nursing approach

What's so naive about naive Bayes

Application of deep and machine learning techniques for multi-label classification performance on psychotic disorder diseases

Sensitivity of chest CT for COVID-19: comparison to RT-PCR

False Positive Results with BD SARS-CoV-2 Reagents for the BD Max System -Letter to Clinical Laboratory Staff and Health Care Providers

Development and internal validation of a diagnostic prediction model for COVID-19 at time of admission to hospital

Understanding the value of clinical symptoms of COVID-19. A logistic regression model. medRxiv, Preprint Retrieved

Review of analytical performance of COVID-19 detection methods

Stacked convolutional neural network for diagnosis of covid-19 disease from x-ray images. Preprints, 2020

A new Mathematical Modell for COVID-19: A Fuzzy Cognitive Map Approach for Coronavirus Diseases

Decision Trees in Machine Learning

Stacking deep learning for early COVID-19 vision diagnosis

Diagnosis and detection of infected tissue of COVID-19 patients based on lung X-ray image using convolutional neural network approaches

Diagnostic performance between CT and initial real-time RT-PCR for clinically suspected 2019 coronavirus disease (COVID-19) patients outside Wuhan

COVID-19 patient health prediction using boosted random forest algorithm. Frontiers in public health

The key role of differential diagnosis in diagnosis

A Hybrid Ensemble Model for Differential Diagnosis between COVID-19 and Common Viral Pneumonia by Chest X-ray Radiograph

A computational intelligence technique for the effective diagnosis of diabetic patients using principal component analysis (PCA) and modified fuzzy SLIQ decision tree approach

CoroNet: A deep neural network for detection and diagnosis of COVID-19 from chest x-ray images. Computer methods and programs in biomedicine

Fuzzy cognitive maps. International journal of man-machine studies

The support vector machine

Testing for SARS-CoV-2: Can We Stop at 2? Clinical Infectious Diseases

Comparison of commercially available and laboratory-developed assays for in vitro detection of SARS-CoV-2 in clinical laboratories

Diagnosis of the Coronavirus disease (COVID-19): rRT-PCR or CT?

Occurrence and timing of subsequent SARS-CoV-2 RT-PCR positivity among initially negative patients

Utility of chest CT in diagnosis of COVID-19 pneumonia

Clinically practiced and commercially viable nanobio engineered analytical methods for COVID-19 diagnosis

CovXNet: A multi-dilation convolutional neural network for automatic COVID-19 and other pneumonia detection from chest X-ray images with transferable multi-receptive feature optimization

Differential diagnosis and classification of apathy

Accurate detection of Covid-19 patients based on Feature Correlated Naïve Bayes (FCNB) classification strategy

Ethics for integrating emerging technologies to contain COVID-19 in Zimbabwe

Coronavirus disease 2019 (COVID-19): Clinical features

Development and Potential Usefulness of the COVID-19 Ag Respi-Strip Diagnostic Assay in a Pandemic Context. Frontiers in medicine

Artificial neural network and logistic regression modelling to characterize COVID-19 infected patients in local areas of Iran

Deep neural network to detect COVID-19: one architecture for both CT Scans and Chest X-rays

Differential diagnosis of acute ocular pain: Teleophthalmology during COVID-19 pandemic-A perspective

Comparative performance of SARS-CoV-2 detection assays using seven different primerprobe sets and one assay kit

What is a support vector machine?

Deep learning covid-19 features on cxr using limited training data sets

Fuzzy cognitive maps. Handbook of granular computing

Support vector machine

6 Easy Steps to Learn Naive Bayes Algorithm with codes in Python and R. Retrieved

Smell and taste symptombased predictive model for COVID-19 diagnosis

A Review on Isabel Pro

Covid-19 detection using artificial intelligence

Classification, diagnosis, and differential diagnosis of multiple sclerosis

Detection of coronavirus disease (COVID-19) based on deep features and support vector machine. Preprints, 2020

Detecting COVID-19 patients based on fuzzy inference engine and Deep Neural Network

The value of clinical parameters in predicting the severity of COVID-19

Model Expert System for Diagnosis of Covid-19 Using Naïve Bayes Classifier

COVIDScreen: Explainable deep learning framework for differential diagnosis of COVID-19 using chest X-Rays

Analysis of prediction and early warning indexes of patients with COVID-19. Expert review of respiratory medicine

A prediction model based on machine learning for diagnosing the early COVID-19 patients. medRxiv

Diagnosing COVID-19: the disease and tools for detection

Clinical Properties and Diagnostic Methods of COVID-19 Infection in Pregnancies: Meta-Analysis

A framework for early differential diagnosis of tropical confusable diseases using the fuzzy cognitive map engine. World Academy of Science, Engineering and Technology

Current practice and potential strategy in diagnosing COVID-19

COVID-19 diagnostics in context

COVID-19 Weekly Epidemiological Update

Emergencies preparedness, response: SARS-CoV-2 Variants

The diagnostic methods in the COVID-19 pandemic, today and in the future

Characteristics of patients with coronavirus disease (COVID-19) confirmed using an IgM-IgG antibody test

Application of ordinal logistic regression analysis to identify the determinants of illness severity of COVID-19 in China

Point-of-Care RNA-Based Diagnostic Device for COVID-19

Deep learning-based decision-tree classifier for COVID-19 diagnosis from chest X-ray imaging

Current and perspective diagnostic techniques for COVID-19

Differential Diagnosis of COVID-19 Pneumonia in Cancer Patients Received Radiotherapy

Contribution of CT Features in the Diagnosis of COVID-19

We acknowledge the contributions of all the medical practitioners that supported us in the course of this research. We express our gratitude to them.