AI-based Power Screening Solution for SARS-CoV2 Infection: A Sociodemographic Survey and COVID-19 Cough Detector

Globally, the confirmed coronavirus (SARS-CoV2) cases are being increasing day by day. Coronavirus (COVID-19) causes an acute infection in the respiratory tract that started spreading in late 2019. Huge datasets of SARS-CoV2 patients can be incorporated and analyzed by machine learning strategies for understanding the pattern of pathological spread and helps to analyze the accuracy and speed of novel therapeutic methodologies, also detect the susceptible people depends on their physiological and genetic aspects. To identify the possible cases faster and rapidly, we propose the Artificial Intelligence (AI) power screening solution for SARS- CoV2 infection that can be deployable through the mobile application. It collects the details of the travel history, symptoms, common signs, gender, age and diagnosis of the cough sound. To examine the sharpness of pathomorphological variations in respiratory tracts induced by SARS-CoV2, that compared to other respiratory illnesses to address this issue. To overcome the shortage of SARS-CoV2 datasets, we apply the transfer learning technique. Multipronged mediator for risk-averse Artificial Intelligence Architecture is induced for minimizing the false diagnosis of risk-stemming from the problem of complex dimensionality. This proposed application provides early detection and prior screening for SARS-CoV2 cases. Huge data points can be processed through AI framework that can examine the users and classify them into “Probably COVID”, “Probably not COVID” and “Result indeterminate”.


Introduction
Novel and emerging pathogens play a crucial issue in public health. The proficiency for scalable, proactive and agile testing has arisen as a significant differentiator in some nation's ability to reverse and handle the pandemic curve. The

Introduction
Novel and emerging pathogens play a crucial issue in public health. The proficiency for scalable, proactive and agile testing has arisen as a significant differentiator in some nation's ability to reverse and handle the pandemic curve. The partnership of researchers and Allen Institute for AI delivered the open-source SARS-CoV2 datasets [1]. These datasets are updated weekly and continuously referred in SARS-CoV2 related research articles to give the fast-track to innovative projects in real-time data. Trace, Test and Treat are the succeeded methodology for leveling the curve of the pandemic in its early stages [2]. However, the pandemic has rapidly spreading all regions in the world. So this methodology is not showing more appropriate [3]. Recent surveys proved that it is a kind of virus that often spreads through undiagnosed coughs [4]. Data demonstrates that 81% of SARS-CoV2 carriers don't develop serious indications to seek medical aid, however they perform as active spreaders [5]. Some people receive symptoms after several days of being affected. They seek medical help. These verdicts attain a new technique based on the "Pre-Screening test, self-isolation, medical care and treatment" for many critical cases.
In this article, we propose the Artificial Intelligence and Machine Learning algorithm to enhance the identification of possible SARS-CoV2 more rapidly using a mobile-based smart application. Moreover, we examine the pathomorphological variations caused by the virus in respiratory tracts. The cough sound of COVID-19 patients is different from other viral and common bacterial infections. Hence the deep study of relevant pathomorphological variations prefers that it can be probable. So, we need to apply the first principle-based methodology [6]. Further, we detect and list the non-SARS-CoV2 respiratory syndromes which are comparatively common and need to analyze a similar cough sound like that of SARS-CoV2 Patients. With the help of the database, this method examines the hypothesis using different kinds of data scrutiny and pre-processing methods.
The various form of data analysis techniques demonstrates that COVID-19 cough does have different attributes while comparing to bronchitis, normal cough and pertussis [7]. To build the perceptiveness from the analysis of cough data and knowledge of the medical domain, we formulate an Artificial Intelligence engine for the initial analysis of COVID-19 from the distinct attributes of cough sounds. This system processes on the front-end cloud-server that acts as the userfriendly application. This application observes the cough sound and transfers that data to the Artificial Intelligence engine . The Artificial Intelligence Engine tracks the sound, and examine whether it is a sound of a cough or not. If the noted audio (recorded sound) is not cough, it instructs the app to indicate to rerecord the cough sound. If the noted audio (recorded sound) is cough, it transfers that cough data to the AI engine diagnosis part. After the analysis, the application gives the outcomes in three possible factors: They are Probably COVID-19, Probably not COVID-19, Result indeterminate. To get reliable results with limited data, the AI risk-averse architecture was proposed. It contains three parallel classifications. Automated mediator consolidated the results of the classifier. In case all the three classifiers do not agree with the terms, the application returns 'Result indeterminate'. It reduces the misdiagnosis rate and offers the 2 nd opinion practice.

Literature Survey
Medical Industry is seeking for novel and effective methodologies to screen and control the SARS-CoV2 pandemic. Artificial Intelligence is one of the technologies which helps to monitor the viral spread and detects the high-risk patients, also helpful in controlling this infection in real-time [8]. By analyzing the previous cases of data, it can also forecast the mortality risk. Artificial Intelligence assists us to fight this virus by medical help, notification, population screening and giving suggestions about infection control [9 -10]. It helps to enhance the treatment, planning and reported results of COVID-19 cases.
An Intelligent platform was built for automatic monitoring and forecast the transmission of SARS-CoV2 cases. A neural network can be developed for visual feature extraction of this disease and this would help for appropriate monitoring and treatment of infected persons [11]. Several surveys [12 -19] suggests that different hidden attributes in cough sounds can be used for Artificial Intelligence-based successful diagnosis of more respiratory problems, so we hypothesize that "The sound of cough can help at least initial analysis of COVID -19 for accomplishing various examinations of its unique hidden attributes which related to other non-COVID -19 Coughs".
According to the guidance of the World Health Organization (WHO), Nucleic Acid Amplification Tests (NAAT) should be used for the confirmation of SARS-CoV2 patients by identifying distinctive orders of viral Ribonucleic Acid (RNA). Currently, this is the standard approach but it is not sufficient to control the pandemic, because for the following reasons.
 Due to temporal and geographical factors, there is limited testing availability.  There is a need for in-person visits to clinics, laboratories, and hospitals. Such visits help us to increase the testing rate of the public. Recent surveys show how SARS-CoV2 is highly stable and highly transmissible. The SARS-CoV2 aerosol stability is up-to 3 hours to 7 days on various surfaces.  The approximate period for existing tests takes more days. In some countries, it expands up to 10 days due to overwhelmed cases in the labs [20 -21]. At the time, the patient is under-examined but the virus might have started to spread to many persons within this period.  The in-person testing methods have only limited protection but they have a high risk to spread the infection.
The United States Food and Drug Administration (FDA), permitted a rapid test that gives outcomes within 15 to 20 minutes for making the test more readily available [22]. The test acts the same as Polymerase Chain Reaction (PCR) by detecting a part of SARS-CoV2 Ribonucleic Acid in Oropharyngeal and nasopharyngeal swab. Recently the FDA permitted the Rapid Molecular dependent test, which brings positive results in 5 minutes and negative results in thirteen minutes [23]. But the FDA alerts that the chance of false-negative results is high [24] while using that test. But, still there is a need for that test in an office visit and to avoid the rupturing of self-isolation and social distancing. Though the newly permitted test requires more solutions for the above-mentioned problems.
To prevent others from this infection, the FDA permitted to collect the samples at home [25]. Though, if a nasal sample of patients is collected, they have to shift it to a specialized lab to process the specific examination on the medical kit. Hence this methodology is also delayed and compromised the quality of the samples (in case the sample stockpiled for too long). Besides, it gives a high probability of errors, while the sample collection is done by the patients, not by the trained health care professionals or doctors. Now two alternative methodologies for SARS-CoV2 infection leveraging analysis of either Computed Tomography (CT) scan images [26 -31] or X-ray images [32 -37] have been developed in literature surveys.
These methodologies are combined with Artificial Intelligence based Image Processing which helps to diagnosis the COVID-19 with higher accuracies. Sometimes it is comparatively better than Reverse Transcription Polymerase Chain Reaction (RT-PCR) based test. Recent surveys display the pooled-sensitivity rate of 94% ( i.e., the confidence interval of 95% cases have 91% to 96%), but low specificity-rate of 37% (i.e., the confidence interval of 95% cases have 26% to 50%) for CT-dependent analysis [38]. Hence, CT-dependent analysis might use to solve the suboptimal PCR test sensitivity rate [39]. Though these methods minimize the problem on radiologists to do the analysis, but still there is a need to visit a well-furnished clinical facility. It is significant because there is a need to examine huge swaths of populations safely, rapidly and cost-efficiently.

Proposed Work
The recent pandemic COVID -19 implies stress on World Health-Care Sector. It is an important need to evaluate innovative patents for attempting to resist the fast-spreading SARS-CoV2. Technologies can provide the rapid detection of COVID-19 cases. To reduce the time for detecting the affected individual, we proposed an innovative method to collect the basic data of an individual such as their travel history, age, gender, common symptoms and signs with the assistance of their mobile application. Also, we collect the data of their cough sound for AI dependent diagnosis. These data analyses can help in the primary screening process and early detection of probable COVID-19 cases. Artificial Intelligence processes these millions of data that examine humans and separate them into safe, low risk, moderate and high-risk groups. It helps to detect the high-risk cases and quarantine them early, thus helps to reduce the viral spread (Table 1).

Data Collection Through Mobile Application
Based on the Center for Disease Control and Prevention Flowchart [40], we established our data collection norms to detect and evaluate novel SARS-CoV2 and we include some additional attributes for enhanced usage of detecting infected cases and resist the spread [41]. Table 1. shows the steps of data collection.

Steps for Identification of Possible Cases:
Let D1, D2, D3, D4 and D5 be the data reported in the data collection system. D2 contains the three datasets that are mentioned in (1) (1) And the 9 sets of Data within D5 are mentioned in (2) (2) Here the set of for I = A,B,C,…..I denote the response of the user based on the existence and non-existence of ith symptom and sign of and denotes the time duration of the specific symptom and sign.
(1) If the Identifier set in (3)  (3) If is similar to one of the features of the dataset in (6) are Then the user will be directed to an NCRC (non-identified respondents or No Health Check Recommended for Coronavirus).
(4) If I2 is similar to one of the features of the dataset , then the user will be directed to NCRC alert.
Did you have close contact with the identified COVID-19 patients in the prior 14 days?
Enter the symptoms and signs listed below a) Cold The above steps help to detect the probable COVID cases and send an alarm to the nearby Healthcare clinic. If a user doesn't have any signs and symptoms of SARS-CoV2 viral infection, then Artificial Intelligence-based health alert can notify them that "Currently there is no COVID -19 risk". Moreover, we collect the cough sound from the user for an AI-based Cough diagnosis that helps to make the result more accurate.

AI-Based Cough Diagnosis for COVID-19
Cough is the frequent and common symptom over a lot of diseases caused by viral or bacterial respiratory infection that not only belongs to the SARS-CoV2 or COVID -19 [42 -44]. Cough can also be caused by many non-respiratory conditions. Table 2. shows the medical condition of non-COVID-19 but that can be the reason for the cough. By using the cough sound, trained physicians execute various analyses among respiratory infections such as asthma, laryngitis, pneumonia and Tracheitis [45 -50]. The location and nature of the primary irritant in the respiratory tract are slightly different, which leads to produce the unique cough sound. Sometimes, it is difficult for AI to differentiate the unique latent attributes in the COVID-19 patient's cough sound. The risk exists in AI-based Cough Diagnosis for the COVID-19 tool to obscure the cough that occurred by any identified disease shown in Table 1 with the COVID-19 cough. To estimate this risk, a brute forcedependent technique is used for collecting cough datasets from huge patients for each of the status mentioned in Table  2. This massive dataset can help to train the powerful AI engine to distinguish the COVID-19 cough from other coughs which mentioned in Table 2. This methodology is difficult at this pandemic movement because it is hard to incorporate all the data which is time consuming. So, we developed a methodology of Domain-cognizant AI-design. It means that the Proposed Artificial Intelligence Engine doesn't simply depend on blind big data churning. But it depends on deep domain knowledge of medicinal scientists, researchers and scholars trained in infectious and respiratory illness to evaluate and short down the hypothesis testing scope. Also helps to reduce the data amount required to test for our hypothesis. The aim is to examine the pathomorphological variations occurred by SARS-CoV2 are unique from another cough mentioned in Table 2. Artificial Intelligence-based Cough Diagnosis for the COVID-19 tool will pick the idiosyncratic attribute of COVID -19 coughs and give a consistent diagnosis.

Solution for Data Description and Practical Possibility
Collecting such massive data in this pandemic is difficult at this time because COVID -19 pandemic requires a fast response. To attain good results in this pandemic situation, we put domain knowledge instead of looking for big data. We select some of the indistinct cough causing conditions from Table 2 to complicate our Artificial Intelligence Engine due to the same pathomorphological variations in respiratory tracts as SARS-CoV2 and the same sign in cough. The shortlist consists of asthma, bronchitis, croup, bronchiolitis, pneumonia and influenza. With the help of previous surveys, we use verdicts to minimize the scope of differential analysis and data collection campaign to the respiratory infections. The unanalysed coughs (bronchitis and pertussis) have a distinct feature.
To make the active mobile application for COVID-19 in a public place or with the existence of different background noises, we design an AI dependent cough detector. This tool of cough detector turns as a filter before the cough analysis and an ability to separate the cough sound from environmental noises. The open-source Environmental Sound Classification -50 (ESC-50) datasets are used to train and validate this cough detector [51]. This ESC -50 dataset gives the massive collection of human, social and environmental sounds. It is classified into fifty classes. One of the classifications is coughs. In our work, we took 1900 cough sounds and 3600 environmental sounds (non -cough sounds) for the training and validating process. And we took the datasets from Virufy open-source cough Dataset for COVID -19 cough analysis. These datasets are original and collected at the hospital. Also, we took datasets from Kaggle open-source data and Hernanmd open-source Cough Datasets.
In the solutions of Artificial Intelligence, the theoretical viability does not assure the result based on the data quality. Moreover, the machine learning algorithm is used. The audio files (recorded cough sound) used in our method are 16bit uncompressed Pulse Code Modulation format with a sampling rate of 44.1 kHz and 3-sec fixed length. For more process, we convert the audio samples of cough sound into the Mel scale. The Mel scale is a pitch classification technique. It is intended to create alterations in frequency (i.e., spectrogram with more closely replicate sound variations). Mel spectrogram is used over the typical frequency because it generates lower resolution in higher frequencies and vice versa, also it is related to the frequency bands of uniformed space in standard spectrogram [52 -53]. In lower frequencies, cough sounds contain higher energy, thus the Mel spectrogram is suitable for recognizing the cough sound. Here (7), we transform the frequency Fq into Mel Scale Ms as: * To evaluate the Cepstral Coefficients, Cepstral analysis was performed on the Mel Spectrum of cough audio (recorded sound) samples is called Mel Frequency Cepstral Coefficients (MFCC) [54]. The extracted attributes of MFCC for all sample outcome in the M*N matrix, where each row denotes the feature extraction and each column denotes one signal frame of MFCC for a particular frame. Frames can differ from one sample to another sample. There are many methods to practice the extracted attributes for the data classification process. In our proposed method, M*N MFCC dependent attribute vectors are extracted for every cough samples input and concatenate them into a final single attribute vector of 2M*1.
This method yields the mean of MFCC attributes parallel to every frame for the first attribute vector. Also, this method grabs the top Principle Component Analysis (PCA) PM*1MFCC attribute projections [55 -56] across every edge and concatenates them into a single vector M*1 by picking their magnitude. Then we combine 2M*1 single attribute vector and attribute vector. This method is demonstrated in Fig. 3.
Since the extracted cough audio attribute is multidimensional for visualizing the attributes, a reduction nonlinear dimensionality method named t-SNE (t-distributed Stochastic Neighbor Embedding) is used. It is appropriate for implanting the high dimensional data in two dimensions. In specific, this strategy designs a high dimensional object by two-dimensional points. This visualization permits feature interpretation in classes or clusters formation associated with classification decision limits. It can be monitored that different types of coughs hold unique features from each other. Hence, COVID-19 features are unique from other types of coughs. Thus, this observance implies the practical feasibility of Artificial Intelligence dependent cough based primary verdict for COVID-19 boosting us to move towards the strategy of Artificial Intelligence Engine for highly efficient and accruable implementation to permit mobile application dependent deployment.
The two phases of the solution for our proposed method: The first phase is the cough sound recognition from noncough, noisy sounds and mixed cough. The other one is the COVID-19 diagnoses from the cough sound. After trained the models for COVID 19 diagnosis and Cough detection, it deployed at the cloud server. The user interface provided by the application with pre-trained models. Cloud dependent implementation has the advantage of refining the model constantly until the data becomes obtainable.

The Architecture of the System
The entire architecture of the system is shown in Fig. 1 and the overall emphasizing steps are articulated in Fig. 2. The mobile application takes the sociodemographic and demographic survey. After completing the survey, it records the cough sound when incited by the voice recognition button. After clicking the diagnosis button, the noted audio is progressed to the server. In the server-side, the audio is nourished into a cough detector. If the audio file doesn't have a cough sound or the poor sound quality, then the server guides the application to re-record the cough sound. If the audio file has a cough sound, then the audio is progressed to three parallel and distinct classifier system.  The outcome of the classifiers is prompted onto the mediator. The application will show the diagnosis results only if all the three classifier yields similar classification results. If the results are dissimilar, the application yield the text of "Result indeterminate". To reduce the possibility of misdiagnosis, this centered tri-pronged mediator system is developed. Here, the opinions of the three diagnosis classifiers are taken with veto power. This technique helps to minimize the total rate of misdiagnoses (False positive and False-negative rate) of the COVID -19 cough detector application.  The noted audio (recorded cough sound) is progressed to a cloud-dependent server. Initially, the cough detector system evaluates its Mel spectrogram with 128 Mel components. Then it resized the image and changed into the grayscale image to merge the scaling intensity and minimize the image dimension resultant in a 320*240*1 dimensional image. The outcome of the image is nourished into Convolutional Neural Network (CNN) centered classifier to choose if the recorded audio input sample is cough sound or not. The high dimensional input of the Mel Spectrogram image is first forwarded through 2*2 max-pooling layer for minimizing the total complexity of the system. Then the two blocks of layers are progressed, each block consists of 2 Convolutional layers (CL) followed by 2 *2 max-pooling layer and 00.15 dropouts.

Detection of Cough Sound
The first block of CL uses 16 filters and 5*5 kernel size while the CL of the second block uses 32 filters. To prevent the overfitting, the trained complex attributes from this 4 CL are compressed and transmitted to 256 Fully Connected Layer neurons followed by 00.30 dropouts. Finally, "Cough" and "not Cough" is classified by the softmax activation function and output layer. The detection model is fulfilled by the binary cross-entropy loss-function.

Diagnosis of COVID-19
After detecting the cough sound from Cough Detector System, it is passed to Artificial Intelligence dependent centered tri pronged mediator for the COVID-19 cough diagnosis process. Here we used three classifiers for producing the output with high reliability [57].

CLASSICAL MACHINE LEARNING DEPENDENT MULTI-CLASS CLASSIFIER (CML-MC)
With the Mel spectrogram image as input, this is the first solution that influences a CNN-dependent 4 class classifier. The 4 classes of cough occurred by 1) Normal cough (no infection in respiratory tracts) 2) bronchitis 3) pertussis 4) COVID -19. The same type of CNN system is applied for detecting the cough but with a little variation to create a four-class classifier by altering the neuron number in the output-layer to 4 neurons for categorizing the input amongst 4 probable resultant classes. It helps to extenuate the overfitting of data. PCA based feature extraction techniques are applied here. These attributes are nourished into Multi-class SVM for data classification. Sampling attained the class balance from each class. With the help of Combined Attribute matrices as input, the k-fold SVM validation technique is performed for 1,00,000 iterations. This method is shown in Fig. 3.

Deep Transfer Learning dependent Multi-Class Classifier (DTL-MC)
This is a 2 nd parallel diagnosis test that contains deep-learning strategy instead of Classical Machine Learning. To shift the knowledge trained by the cough detection system by the same kind of diagnosis model Deep Transfer Learning model is used. This permit to train deep architecture using restraint amount of trained data as the primary attribute of Mel spectrogram input which classifies the pre-trained cough and it needs to be fine-tuned for more-subtle attributes. Here, we fixed the initial weights of the 1 st convolutional layer as a preliminary layer and train low-level latent attributes. For fine-tuning the weights, it permits the other layers. This method gives better performance than Scratch training of disease data.

Deep Transfer Learning dependent Binary-Class Classifier (DTL-BC)
This is a 3 rd parallel diagnosis test that contains deep transfer learning-dependent Convolutional Neural Network with an input of Mel spectrogram image. But only the binary classification is performed by it with similar input. It is used by CNN structure in Cough Detector System.

Experimental Results
To compute the model performance, we use the metrics of F1-score, specificity, sensitivity, overall accuracy, precision, and recall on validate and cross-validate set models. K-Fold cross-validation technique is performed to 1 0

Sadhana et al. / Procedia Computer Science 00 (2019) 000-000
compute the machine learning model's performance [58]. From Cross-Validation, Mean Confusion matrices are performed to compute these approaches. Moreover, Regularization methodology is used to prevent overfitting issues. Different hyper-parameters tuning such as dropout rate, activation functions, learning rate, number of latent layers are the functions reliant on deep neural network models that are computed based on cross-validation accuracy. Besides, to prevent the probability of overfitting, we evaluate the epochs numbers versus model delay.  Table 3 and Table 4 show the detection algorithm of Confusion Matrix and Cough Detection Performance Metrics respectively. Results showed that this method can separate cough and non-cough with total accuracy of cough detection is 95.76%. Fig. 4 shows the error graph of epochs of neural network dependent model versus mean loss for training data sets and validating the data sets. This approach has not been over-fitted because of the decay curve in training and testing data sets.

Diagnosis of COVID-19
Table 5 reports about the 1 st classifier of CML-MC classifier. It shows the 5-fold cross-validation of the standardized mean confusion matrix. Fig. 5 denotes the overall accuracy of the Cumulative Distribution Function (CDF) with different k's in k-fold cross validation datasets. Data availability and Performance matrices of this system are demonstrated in Table 6. The outcome of total accuracy is 88.13%.   Table 7 shows the 2 nd classifier DTL-MC performance metrics. Here, the total accuracy of the DTL-MC classifier is 92.11%. Table 8 shows the outcome of the normalized mean of the confusion matrix. For training and validating the datasets, the epochs vs mean loss of the DTL-MC classifier is reported in Fig. 6. After completing approximately 25 epochs, both curves are initiated to saturate that specifies the sensible learning time. Also indicates that it doesn't have overfitting.  Table 9 shows the 3 rd classifier DTL-BC performance metrics. Table 10 reports the normalized mean of the confusion matrix. 92.77% is the classification accuracy of this method. Fig. 7 illustrates the epochs vs mean loss of the DTL-BC classifier. After completing 20 epochs, both curves are initiated to saturate that specifies the sensible learning time. Also indicates that it doesn't have overfitting. While choosing the binary classification technique, the amount of non-COVID-19 cough sound samples is more than the COVID-19 cough sound samples. The outcome of DTL-BC and DTL-MC (deep learning-dependent classifier) is higher than CML-MC (manual attribute extraction dependent classifier). If the amount of data will increase, the classification accuracy for CML-MC classifiers will become increased.

Analysis of Overall Performance
After the performance analysis of 3 different kinds of classifiers, currently we examine the total performance of Artificial Intelligence (AI) power screening solution for SARS-CoV2 infection System with the help of mediator dependent architecture. This architecture gives the best performance when the classifier is entirely independent. Here, these 3 classifiers are not entirely independent but almost they are independent because it performs a unique training and testing process. Now we examine the overall performance of the methodology under the assumption of independent architecture. For comparison of the misdiagnosis rate of individual classifier vs mediator results, this needs to be performed. The prominence of the tri pronged mediator in this proposed method shows how the low-risk model can minimize the total false positive and false negative rate of Artificial Intelligence (AI) power screening solution for SARS-CoV2 is shown in equation (4) and (5). Here the rate of misdiagnosis of the total model is nearly zero. Tri pronged mediator counts over-sensitivity or under-sensitivity of each classifier by screening it with 'Result indeterminate (CI)'.
The possibility that the application predicts 'Result indeterminate (CI)' when the case can either infected by SARS-CoV2 or not infected by SARS-CoV2 is evaluated in (12) and (13). Now, the application would forecast the indeterminate result of 37.7% of the time of . By shifting the mediation approach for minimizing the percentage where the outcome of the application reflects a simple or weighted majority with M number of classifiers. Table 6 summarizes the outcome. The dependency factor can be captured by introducing the co-efficient factor.

Function of the Combiner
To get the results more accurate, Combiner evaluates the results of the IPC algorithm and Cough Detector, and generates the outcome as "Probably COVID-19" or "Probably not COVID-19" or "Result indeterminate" is shown in Table 11.

Conclusion
This paper proposed the Artificial Intelligence (AI) power screening solution for SARS-CoV2 infection with the help of Socio-Demographic Survey and Cough Detector through Mobile Application. The proposed method can help to detect the patient who might have any mild signs and symptoms. Also, acts as the test medium for COVID-19 cough diagnosis from different types of cough. We reported that COVID-19 cough contains the distinct latent attributes from other coughs. We evaluated the algorithm by latent attribute in COVID-19 cough patients and other coughs such as bronchitis, pertussis and non-infectious cough. From the perception of medical domain knowledge, we developed an application based on Artificial Intelligence centered tri-pronged mediator for cough-based analysis. The outcomes show this application can detect the COVID-19 people from the public. The accuracy of the application is reliable because we collected huge cough related data samples from various open source websites and applied our proposed methodology in it. Our proposed methodology is trustworthy and shows reliable accuracy. This application also reduced the possibility of misdiagnosis because we developed this structure as the risk-avert architecture. This application can't compete with clinical testing. But it is a distinct functional application for tracing, monitoring, tracking and cost-effective tool and also helps to reduce the rapid spread of this pandemic virus by the virtual environment. In future, we are going to do this proposed methodology in real time and compared our results with other experimental data.