Symptom based COVID-19 test recommendation system using machine learning technique

1.Introduction

The International Committee on Taxonomy of Viruses (ICTV) labeled Severe Acute Response Syndrome coronavirus (SARS-CoV2) induced coronavirus disease (COVID-19) [1] pandemic initially started in Wuhan, China since January 2020 and spread all over the world. The COVID-19 is belonging to Coronaviridae family, and size is 65–125 nm diameter creates severe respiratory problems and ease of spread through air and droplets of human while coughing.

Currently, many developed countries in the world are severely affected by this virus due to the lack of authorized information about this virus. At the same time, ignorance, unpreparedness, and reluctance of public authorities of different countries increased the severity of this virus. The dynamic gene mutative nature of the COVID-19 virus enables it to adapt to quick changes in the environment and helps to sustain in different weather conditions. This peculiar behavior of the virus poses many challenges to the entire world. The advent of this pandemic exposes many drawbacks in the medical and healthcare system. It exposes many vulnerabilities especially related to the readiness of societies to such kinds of health emergencies. The World Health Organization (WHO) releases several guidelines to face the current situation from time to time. Many countries following either the 3T (Test, Treat, Track) or 5T (Test, Trace, Treat, Teamwork, Track, and monitoring) formula to contain the virus. Many countries have announced Lock-down to reduce the spread of the virus.

In India till 22 March 2020 around 360 people have been infected with COVID-19 and 7 were dead. At present, according to the statics of the Indian government as of 19 Jan 2021 exactly 1,05,81,837 has been infected with COVID-19, the death toll was 1,52,556, the death rate is 1.44% and discharged cases (cured) after treatment is 96.66%. This analysis shows that India is on the safer side because of strict implementation of lock-down till 31 May 2020 and subsequent unlock rules till date. On the other hand, In Andhrapradesh (AP) state 886245 people have been infected with COVID-19, 877443 were cured, active cases are 1660, the death toll was 7142, the death rate is 0.8%, and discharge rate is 99% as of 19 Jan 2021.

Mucahid Barstugan et al. [2] proposed the machine learning-based classification of COVID-19 CT images which applies feature extraction methods such as Grey Level Co-occurrence Matrix (GLCM), Local Directional Pattern (LDP), Grey Level Run Length Matrix (GLRLM), Grey-Level Size Zone Matrix (GLSZM), and Discrete Wavelet Transform (DWT) to improve the performance in classification. The Support Vector Machines (SVM) algorithm is used for the effective classification of infected regions of CT abdominal images.

Zifeng Yang et al. [3] proposed the Modified SEIR and AI prediction of the trend of the epidemic of COVID-19 in China under public health interventions which integrates machine learning techniques with Susceptible-Exposed-Infectious-Removed (SEIR) model to predict the progression of COVID-19 epidemic in china. The domestic migration and the most recent COVID-19 epidemiological data are applied to the modified SEIR model to predict the status of coronavirus in china. In Japan, according to COVID-19 policy, people with high fever for more than four days are recommended for COVID-19 test, and the people with mild COVID-19 symptoms are restricted to homes. This situation hides the deaths due to COVID-19.

The society coexisting with COVID-19 [4] discusses how badly Japan country affected due to the delay in COVID-19 testing. Hence, a recommender system for selecting a potential subject or candidate for COVID 19 testing is proposed. Machine Learning algorithms are used to classify a subject into any of the five classes such as No-symptom, General symptoms, Low-risk, Medium-risk, High-risk. The people labeled as Low-risk, Medium-risk, and High-risk are recommended for the COVID-19 test. This proposed work considers five-cold, cough, fever, breath, and comorbodities-ill-health symptoms and age as the other important factor.

The rest of this paper is organized as follows, Section 2 discusses past literature related to different disease recommender systems. Section 3 explains the working approach of the proposed work. Performance evaluation of classification techniques discussed in section 4. Implementation of the symptom-based recommended system presented in section 5. Finally, conclusion and discussion for future scope in Section 6.

2.Related work

Patil et al. [5] discussed suitable machine learning algorithms used to predict, decision making, and analyze medical data related to diabetes. The authors experimented on PIMA Indian diabetic data set. The proposed method applies different pre-processing techniques, future (Forward and Backward) selection methods, and then applied five different classifiers (ADABoost, Decision Tree, XGBoost, Voting Classifier, and Stacking Classifier) to produce five different trained models on the PIMA Indian data set. Finally, it is concluded that the results of ADABoost and Random Forest are best out of all five classifiers in terms of accuracy and other metrics.

Mumtaz Ali et al. [6] proposed a neutrosophic recommended system for medical diagnosis based on algebraic neutrosophic measures which identify or recommend disease of a patient based on three issues such as symptoms of a patient, the probable symptoms of a respective disease, and identification of disease based on symptoms possessed by a patient.

Chen et al. [7] proposed Disease Diagnosis and Treatment Recommendation System is an ontology-based Diabetes Medication Recommendation system which creates knowledge on different attributes – the nature of the diabetic drug, dispensing type of drug, and side effects – of each drug/medicine and ontology information on symptoms of a patient to advise/recommend a potential prescription to the diabetic diseased person. Phanich et al. [8] proposed the Food Recommendation System (FRS) which suggests a healthy and nutritional diet plan for diabetic patients. The food clustering analysis is performed by FRS to recommend food with less sugar and fat content. In recent times, the importance and usage of natural herbs for chronic diseases are increased and the Choi et al. [9] proposed HerDing which is a herb recommended system useful for treating a chronic disease like leukaemia, diabetics, heart diseases, etc. The HerDing recommends respective natural products (herbs) based on symptoms of a patient.

In the literature, recommended systems [5, 6, 7, 8, 9] using different machine learning techniques were proposed. In addition to this, recently, Srinivasa Rao et al. [10] proposed a machine learning-based web survey through mobile phone for the improvement in the identification of possible COVID-19 cases and to reduce the spread of the virus among susceptible people who are under quarantine. The survey intention is to collect travel history and symptoms from a person who is under COVID-19 investigation. Similarly, AI techniques are applied on such preliminary data for early identification of people with COVID-19 and non-COVID-19 symptoms as well as to classify them into no-risk, minimal-risk, moderate-risk, and high-risk groups. The identified high-risk people can be recommended for isolation. In line with the proposed web survey through mobile phone [10], in this paper, we proposed COVID-19 test recommendation system using machine learning techniques on house-to-house survey data collected by the health department of Andhra pradesh (AP) state government to identify the people infected with General and High-risk symptoms. This proposed work recommends people for COVID-19 test based on disease class – No-Symptoms, General Symptoms, Low-risk, Medium-risk, and High-risk – identified using machine learning methods like KNN, DT, Naïve Bayes, RF, SGD, and SVM. The proposed symptom-based test recommendation system gives more than 99% accuracy in recommending COVID-19 test to the people who suffer from defined ill-health symptoms described in Algorithms 3.1 and 3.2. These algorithms are prepared based on inputs of the Superintendent of Physical Health Center (Rajam) mentioned in Table 2 and representation of each symptom is as shown in Fig. 1.

Figure 1.

The Symptom and it’s representation.

In this piece of work, the proposed Algorithms 3.1 and  3.2 are applied on house-to-house surveyed data to build a synthesized data set and then, data balancing methods applied to this synthesized data set to make it balanced. Now, the machine learning techniques like KNN, DT, Naïve Bayes, RF, SGD, and SVM are applied on a both synthesized and balanced data set to identify different classes based on age, comorbidity, and other symptoms (which are defined by the health department of AP) of people suffering from various diseases.

Figure 2.

The flow of proposed work.

3.Symptom based COVID-19 test recommendation system using machine learning techniques

The recent statistics reveal that spread of the COVID-19 virus transformed into a pandemic form epidemic. The mankind of the entire world is fighting against the COVID-19 pandemic. The World Health Organization (WHO) [11, 12] strictly recommends more tests to admit COVID-19 infected people in COVID-Hospitals to decrease community spread of COVID-19. It also recommends different testing priorities and strategies [12] for early detection of infected people as well as best utilization of available medical – testing infrastructure, medical staff, etc. – resources. The proposed COVID-19 testing recommender system is very useful to detect and separate probable high-risk category people. The flow of proposed work is as shown in Fig. 2.

3.2Implementation of rules to create synthesized data set

The original data set shown in Table 1 is imbalanced and not suitable for the learning process because of the below reasons

• • Huge set of people without symptoms in the given data set.

• • The number of people in each house is not known since it populates secretariat wise data.

• • The populated data shows how many people under one secretariat have symptoms in each category such as the total number of persons either male or female who has general symptoms (Gs) and High-Risk Factor (described in Table 2).

• • The symptoms – cold, cough, fever, or breathing – of a patient is unknown in each category.

In this connection, the below steps are implemented to extract features based on the data given in Tables 1 and 2 to address the above-said reasons.

• Step 1: According to the Indian context, it is assumed that the total number of people who dwell in each house is four. So, the number of houses multiplied by four is the total population in each secretariat. As a result, the total surveyed population Sp is the sum of Secretariat wise houses multiplied by four. That is 28326 × 4 = 113304.

• Step 2: According to the inputs of PHC superintendent a subject does not have any symptoms treated as No-Symptoms (Ns), subject with Cough, Cold, Fever are classified as general symptoms (Gs), the Age and Comorbidities (chronic diseases) are considered as two other important features along with Gs to decide risk factors (Rf) such as Low-risk (rl), Medium-risk (rm), and High-risk (rh). The subject who have 𝑎𝑔𝑒>60 with Gs and breathing problems considered as rl, the subject who have 𝑎𝑔𝑒<60 with Gs, breathing and Comorbidities considered as rm, and the subject who have 𝑎𝑔𝑒>60 with Gs, breathing and Comorbidities considered as rh.

• Step 3: The features or symptoms such as Cold, Cough, Fever, Breathing, Chronic (Comorbidities), and Age are identified according to step 2

• Step 4: The value 1 (one) of a corresponding feature represents that the people suffering from that symptom and value 0 (zero) means not suffering from the corresponding symptom as shown in Fig. 1.

• Step 5: Among Sp, 321 people suffering from general symptoms, 123 are classified as a Low-risk category, 67 are in Medium-risk and 77 are in the High-risk category.

• Step 6: Based on the data set shown in Table 1 the symptoms of a person who is under Gs, Low-risk, Medium-risk, and High-risk category is unknown. In this context, it can be assumed that the person who may suffer from Gs may suffer from either only cold, only cough, only fever, only cold and cough, only cold and fever, only cough and fever, or all three symptoms. Hence, each person who suffers from Gs multiplied by 7 to cover all possibilities.

• Step 7: In the same way as in step 6 the people who suffer from Rf are multiplied by 16 since Rf considers 6 features such as cold, cough, fever, breathing, age, and chronic. Among these symptoms the age and chronic are fixed with either 0 (zero) or 1 (one) based on the category of risk factor Rf and the remaining four features keep changing.

• Step 8: As per step 6&7 multiply the total people under Gs category with 7, Low-risk (123) category with 16, Medium-risk (77) category with 16 and High-risk (67) category with 16. So, the Sp is added with 321 × 7 (gs), 123 × 16 (rl), 77 × 16 (rm) and 67 × 16 (rh). Hence, the total size of each feature is 1 ×Tp where Tp=Sp+gs+r⁢fl+r⁢fm+r⁢fh. The remaining 113192 entries are considered as No-Symptoms.

[b]

To initialize symptoms under each class for all possible cases. The class number(0,1,2,3,and ⁢4), starting index and end index of corresponding class considered as input Each feature is filled with either 1 or 0 to represent all five classes as described in step 8 of Section 3.2. set_gs_rf (class, start, end)Initialize array a[]←{0,0,0,0}Initialize counter c←0i←𝑠𝑡𝑎𝑟𝑡endc==0a[0]←0; a[1]←0; a[2]←0; a[3]←0; j←0; c>0; j++a[j]←c%⁢2;c←c/2; Cold[i]=a[0]; Cough[i]=a[1]; Fever[i]=a[2];class==0Recommendation[i]=0 (No-Symptoms); class==1Breath[i]=a[3]; Age[i]=0;Chronic[i]=0;Recommendation[i]=1 (Gs); class==2Breath[i]=a[3]; Age[i]=2;Chronic[i]=1;Recommendation[i]=2 (rl); class==3Breath[i]=a[3]; Age[i]=1;Chronic[i]=0;Recommendation[i]=3 (rm); class==4Breath[i]=a[3]; Age[i]=1;Chronic[i]=1;Recommendation[i]=4 (rh); c++class==0 and c>7c=0 class≠0 and c>15c=0 An index of each feature consists a value either 0 or 1 to indicate corresponding symptom. Initialization of all possibilities according to step 6 and 7 as in Section 3.2

Table 3

The class labels of data sets 𝒅𝒔𝟏 and 𝒅𝒔𝟐 before and after train-test split

Classes	[rgb]0.2,0.2,0.2𝒅𝒔𝟏	[rgb]0.2,0.2,0.2𝒅𝒔𝟐		[rgb]0.2,0.2,0.2After splitting[rgb]0.2,0.2,0.2of 𝒅𝒔𝟏		[rgb]0.2,0.2,0.2After splitting[rgb]0.2,0.2,0.2of 𝒅𝒔𝟐
		RUS	SMOTE	Train (70%)	Test (30%)	Train (70%)	Test (30%)
0	113192	1072	2247	79200	33992	749	323
1	2247	2247	2247	1572	675	1590	657
2	1968	1968	2247	1397	571	1359	609
3	1072	1072	2247	757	315	764	308
4	1232	1232	2247	871	361	851	381

Figure 3.

The data balancing methods

These steps are used to synthesize the data set according to the data given in Table 1 by extracting the necessary features based on inputs of PHC superintendent (Rajam) given in Table 2 followed by providing proper initialization of extracted features to make them suitable for classification as described in Algorithms 3.1 and 3.2. Finally, prepare the data set ds1 by combining all initialized features. Based on a literature study it is observed that the training and prediction accuracy of the proposed model is effectively evaluated with a balanced data set. Hence, the data balancing techniques discussed in Section 3.3 are applied on a resultant data set ds1 to generate even more balanced data set ds2. The different machine learning methods for classification applied on both data sets (ds1 & ds2) and its results are analyzed in Section 4.

4.Evaluation of synthesized data set and classification models

The performance of machine learning models is evaluated using accuracy, sensitivity, specificity, precision, rate of misclassification, and F1-score metrics [19]. The calculation of evaluation metrics uses numerical values of true positive (TP), true negative (TN), false positive (FP), and false-negative (FN) which are obtained from confusion matrices of a respective classification model is as shown below:

Figure 4.

(a) General symptom, (b) Low-risk, (c) Medium-risk, (d) High-risk.

Four steps are performed in this piece of work. At first, the COVID-19 survey data shown in Table 1 is not suitable for learning as described in Section 3.2. Hence, the Algorithms 3.1 and  3.2 are applied on Table 1 to create synthesized data set ds1. Second, the data balancing methods – described in Section 3.3 – such as random under-sampling followed by SOMTE data balancing methods applied on ds1 to create balanced data set ds2. Third, six classification models such as KNN, DT, Naïve Bayes, RF, SGD and SVM are applied on both the data sets ds1 and ds2. Finally, The less data sensitive and data biased KNN learning model is integrated with the GUI of the COVID-19 Test Recommender system.

In first task, the data set ds1 is created as described in Algorithm 3.1. According to the step-1&8 of Algorithm 3.1 the total number of houses is 28326 based on data shown in Table 2 multiplied with four to fix the Sp. Then, the people of gs,rl,rm, and rh added to Sp. As a result, the Tp in ds1 is 119823. The data samples of each class of ds1 shown in Table 3.

In the second task, the RUS data balancing method makes data samples of a maximum category is equal to the minority class of the respective data set by deleting the required number of entries from the maximum category. On the other hand, the data oversampling method SMOTE makes the number of data samples of all classes is equal by adding the required number of extra data samples to each class which are less than the maximum data entries class of the same data set. The RUS followed by SMOTE applied on ds1 to create ds2. In this process, the RUS reduces data entries (113192) of a maximum category (No Symptoms) to data entries (1072) of the minority category (Medium-risk). As a result, 1072 data entries in No-Symptoms category, 2247 samples in General symptoms, 1968 data entries in rl, 1072 data samples in rm and 1232 entries in rh as shown in Table 3. Next, the minority data oversampling method SMOTE is applied to the result of RUS to make an equal number of example data samples to all five categories. The result of SMOTE is stored in ds2 which is shown in Table 3.

In the third task, The training and testing accuracy of classification models like KNN, DT, Naïve Bayes, RF, SGD, and SVM is applied on both ds1 and ds2 with 70:30 percent of train and test ratio. Then, the accuracy and response times of training and testing of each classification model on both the data sets are shown in Table 4. Moreover, the evaluation metrics of all these classification models on both ds1 and ds2 are also evaluated resulting in sensitivity or true positive rate is equal to 1.0, the specificity or true negative rate is equal to 0, precision is 1.0, rate of misclassification is 0, and F1-score is 1.0.

From the results as shown in Table 4 KNN classifier performs well on both the datasets when compared with other classifiers. It is a multi-class classifier suits to the datasets ds1 and ds2 which consists five classes, consumes less time for both training and prediction, and it performs well when the size of a dataset is small as well as training data is larger than the number of features. Hence, it is saved and integrated with the COVID-19 Recommended system as shown in Fig. 4 as the final task.

5.The symptom based recommended system

The results discussed in Section 4 shows the performance of all six classification methods such as KNN, DT, Naïve Bayes, RF, SGD, and SVM gives better results on both the data sets. However, the KNN outperformed other classifiers with both training and testing data. Henceforth, it is saved and used in the GUI of a recommended system to predict the status of a person based on collected symptoms. The proposed recommended system collects symptoms of Cold, Cough, Fever, Breath, Chronic (Comorbidities), and Age from the user as shown in Fig. 4. This information is used to recommend whether the COVID-19 test is required or not as well as the level of severity in collected symptoms. The proposed recommended system is useful for both web and mobile-based survey. The symptoms of a subject are entered through the GUI shown in Fig. 4. A health assistant can enter either 0 or 1 in a specific text field of the GUI. As shown in Fig. 1 the value 0 (Zero) indicates the no presence of a symptom, and the value 1 (One) indicates the presence of a symptom.

The machine learning algorithms are applied on the synthesized dataset where the algorithms are adopted from the sklearn API. All the modules in the proposed work such as creation of synthesized data, data balancing, classification, saving of model and GUI are implemented in python. The source code of this work is available at https://github.com/KLakshmanarao/covid19 for reference.

6.Conclusion and future work

We have evaluated our proposed model with six classification algorithms. From the experimentation results, it is observed that the evaluation metrics – Sensitivity, Specificity, Precision, Rate of misclassification, and F1-Score – and train-test accuracy are good and similar on both the data sets ds1 and ds2. Consequently, it is proved that the generation of synthesized data set ds1 through Algorithms 3.1 and 3.2 is balanced. It is also observed that the integration of the COVID-19 test recommender system with the KNN model achieves 100% accuracy in predicting the category of a subject based on its symptoms. In addition to this, the proposed COVID-19 Test recommended system using machine learning techniques will be very helpful to public health systems to decide whether a subject needs a COVID-19 test or not. In this proposed model, all the classes are recommended for the COVID-19 test except for the subject with general symptoms. Because of the scarcity of COVID-19 test kits and other medical resources, the proposed recommender system will be very useful for optimal utilization of medical resources since only required people will be recommended for COVID-19 testing. If a subject which is recommended for COVID-19 test can be quarantined to reduce the spread of the virus. The citizen can also use the proposed system and test themselves on their own. The proposed work can be extended in such a manner that the integration of additional symptoms related to COVID-19 can be taken into consideration to make the prediction more active and robust.