From January 2021, data related to cough of people with COVID-19 was recorded and collected using the online site “sorfeh.com/sendcough/en.” In order to increase the number of data, the contact information of patients with COVID-19 was received through laboratories, hospitals and infectious specialists. During this process, the data was collected after the patient's consent and having the required conditions. The conditions for the patient to enter the study include a positive PCR test, or CT Scan diagnosis of lung involvement by CT scan, or a definite diagnosis of corona virus by a physician based on clinical examinations, and no more than 8 days have passed since the definite diagnosis of COVID-19. People admitted to the hospital were excluded from the study because usually more than 8 days have passed since the duration of the illness of the hospitalized people and their coughs could be pulmonary complications after COVID-19 and not related to COVID-19.

Inclusion criteria for healthy people into the study: (1) lack of symptoms of the COVID-19, (2) if the symptoms are present, the physician's diagnosis should not be COVID-19, and (3) None of the close people should have COVID-19 (Alqudaihi et al., 2021; Sadhana et al., 2021; Santosh et al., 2022). During data collection and recording of coughing, healthy people and patients were asked to cough in a safe environment without the presence of other people.

The link of the online site was provided to the patients. Patients first chose their symptoms on the online site and recorded their cough for 7 s in a quiet environment and without the presence of people. The presence of sounds other than coughing can cause misdiagnosis (Figure 1). Then the recorded sound will be played and if the quality is confirmed, it will go to the next stage. In the next step, the user must select the symptoms he has and select the time of occurrence. Then he will be asked additional questions, such as the current status, age, gender, previous history of infection, PCR test, CT scan of the lung, whether there is a conflict or not. The recorded sounds are not from a specific type of microphone and have been recorded by different types of smartphones, tablets and laptops. This case can have a positive effect on the study and avoid bias. All the data collected are from Tehran city, which are from different minorities and ethnicities with various accents. All audio files used in our study are in uncompressed Pulse-code modulation (PCM) 16-bit format with a sampling rate of 48 kHz and a fixed 7-s length (Pahar et al., 2022).

The total number of collected data is almost 40,353. Almost 17,690 data were manually removed due to incomplete information, low sound quality, silent audio content, presence of surrounding noise when recording cough, absence of cough sound in audio content. Among the remaining 22,663 cough sounds, there were 14,521 negative coughs and 8,142 positive coughs. On the website, in addition to recording the sound of coughing, patients also recorded information such as clinical symptoms, days of onset of symptoms, the status of PCR tests and CT scans, the presence of previous infections, age, gender, and the person's disease status. The information about the symptoms of the patients is shown in Table 1. According to Table 1, the most common symptom among people with COVID-19 is fever (51.75%). Dry cough and productive cough are among the most common symptoms of the patients participating in the study with prevalence of 25.01 and 21.22%. Also, 29.2% of patients had no clinical symptoms. The symptoms of healthy people were negligible, so it was ignored.

The age range and gender of patients and healthy people are also mentioned in Table 2. The age range of the participants was 5–90 years, most of the participants were between 15 and 30 years old and most of them were women. All participants were residents of Tehran city, which are from different minorities and ethnicities with various accents.

The analysis will only be done on the classification and diagnosis of the infected and healthy person based on the sound of the cough, so no classification was done on each of the files of each category of coughs. We used all the positive samples of COVID-19 in our data set and randomly used the same number of negative samples of COVID-19 for a balanced distribution of samples (8,142 person in each group). The patients in the case group were selected with the conditions of (1) a positive PCR test, or (2) diagnosis of lung involvement by CT scan or (3) definite diagnosis of corona virus by a physician based on clinical examinations, and (4) no more than 8 days have passed since the definite diagnosis of COVID-19 and healthy people were selected in the control group with the conditions of (1) lack of symptoms of the COVID-19, (2) if the symptoms are present, the physician's diagnosis should not be COVID-19, and (3) none of the close people should have COVID-19.

We used Handcrafted Features to extract audio features. The frequency of the recorded raw audio was 48 kHz and was stored without compression. We used the librosa library to extract features. Extracting more features does not always lead to beneficial results, so first we extracted the available features according to their previous applications in medical diagnosis. If no result is obtained, more features are extracted.

The most common and useful features for speech and audio recognition are Mel-Frequency Cepstral Coefficients (MFCC). This feature can create higher resolutions at lower frequencies. The main idea in extracting MFCC coefficients is the property of the human ear in receiving and understanding speech, and this issue has made these coefficients a powerful tool in all areas of audio processing and recognition. The number of coefficients used in voice recognition usually varies between 9 and 13. The coefficient of 0 indicates energy, which is referred to as the characteristic of Shimmer. In the first step of extracting these coefficients, the Fourier is converted into a signal. Then the obtained spectrum exponent is expressed in mel scale and logarithm is taken from the exponent at each mel frequency. In the last step, the logarithmic spectrum of Mel is returned to the time domain. The result of these transformations is the Capstral representation of the signal spectrum, which shows the spectral characteristics of a frame of the audio signal (Figure 2).

We apply the Discrete Fourier Transform (DFT) on each cough audio (Picone, 1993).

Where N denotes the number of samples in frame, y_i[t] is the discrete time domain cough signal, w(t) is the window function in time domain and Y(k) is the kth harmonic corresponding to the frequency.

f (k) = kFs/N where Fs is the sampling frequency. MFCCs use Mel filter bank or triangular bandpass filter on each cough audio DFT output, equally spaced on the Mel-scale. At last, we apply the Discrete Cosine Transform (DCT) on the output of the log filter bank in order to get the MFCCs:

Where i = 1, 2, …, l; l denotes the cepstrum order, E(m) and M are the filter bank energies and total number of mel-filters, respectively. An audio signal is constantly changing, so for ease of understanding, it is assumed that on short time scales the audio signal does not change statistically. For this reason, we frame the signal in 10–100 ms frames. If the frame is too short, we won't have enough samples to get a reliable spectral estimate, and if it's too long, the signal will change too much over the entire frame. Mel's scale was used to more accurately determine which frequency is present in the frame. The Mel scale relates the perceived frequency or pitch of a pure sound to its actual measured frequency. Using this scale makes the features more consistent with what humans hear.

In extracting features, we first set pre-emphasis equal to 0.97 and normalize audio for each audio file, and in extracting features, in addition to MFCC, we considered delta and delta2 MFCC. Where delta MFCC is the temporal differential (delta) of the MFCC and delta2 MFCC is the differential of the delta of the MFCC (acceleration coefficients). The number of coefficients considered 13 and 23 ms step was used to generate MFCC coefficients, and the number of 302 frames was obtained, and average numbers were used in each frame.

Also, during to 7 s considered extracting chroma stft, rmse, spectral centroid, spectral bandwidth, rolloff, zero crossing rate, tonnetz and melspectro cases. Finally, 914 features were obtained (3 ^* 302 + 8 = 914).

Where (chroma stft) is chromagram from a waveform or power spectrogram and (rmse) is the root-mean-square of the magnitude of a short-time Fourier transform which provides the power of the signal. (Spectral centroid) is the mean (centroid) extracted per frame of the magnitude spectrogram. (Spectral bandwidth) is the band width of light at one-half the peak maximum. (Rolloff) is the center frequency for a spectrogram bin so that at least 85% of the energy of the spectrum in this frame is contained in this bin and the bins below. (Zero crossing rate) is the rate of sign-changes of the signal. (Tonnetz) is tonal centroid features and (melspectro) is mel-scaled spectrogram.

From each control and case group, one sample was selected that were similar in terms of the number of coughs and phonetics, and their extracted characteristics are shown in Figure 3.

Twenty-three milliseconds step was used to generate MFCC coefficients, and the number of 302 frames was obtained, and average numbers were used in each frame. After generating the MFCC coefficients, they can be plotted on a spectrogram to visualize the sound. Heat maps can easily distinguish between two groups. In the healthy and sick group, the minimum, maximum and average values of the features that had a significant difference after extraction are shown in Figure 4.

In this study, after the data was collected and the initial pre-processing was done, the dataset of the samples was prepared in two train and test groups. Then, data augmentation was done using standard methods, and finally common and useful features were extracted from them.

We trained our models in two modes. The first case is the use of primary data without augmentation the data between two categories of positive and negative COVID-19 and the second case is the use of augmentation data that has been increased for both positive and negative COVID-19 categories.

We have trained and evaluated five machine learning classifiers in total. Supervised Learning classification algorithms have been used, which include Support Vector Machine (SVM), random forest, SVM classifiers have performed well in classifying cough events (Sharan et al., 2017). Also the artificial neural networks that were based on the standard “Fully Connected” neural network, Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) recurrent neural networks have been established. CNN is a popular deep neural network architecture primarily used in image classification. It has also performed well in the classification of breathing and speech of COVID-19 (Pahar and Niesler, 2021). An LSTM model is a type of recurrent neural network whose architecture allows it to remember previously-seen inputs when making its classification decision. It has been successfully used in automatic cough detection (Miranda et al., 2019).

We tested different techniques for creating classification models, first to examine how effective these methods were for classification ability, and finally to examine the success rate in each mode to see which type of classification more efficiency can be achieved.

The independent term in the kernel functions is chosen as a hyperparameter during the optimization of the SVM classifier. The networks was optimized using the Cross-Entropy Loss and Adam optimizer with default parameters.

We selected several standard evaluation measures such as F1-Score, Sensitivity/Recall, Specificity, Precision and Accuracy for each model to evaluate the best models and their results. Several iterations were performed and the results reported. Table 3 shows the results of each model with the following observations: It was observed that most of the models have accuracy, sensitivity and specificity higher than 70% and this shows that cough provides the necessary data about the respiratory system and the pathogens involved. The signal processing features enable the model to capture latent cough sounds and detect COVID-19 with sufficient sensitivity and specificity.

Although all the models were able to achieve the minimum required evaluations, but not all of them are necessarily acceptable, and in the existing conditions, each of them shows different functions, and the best performance in the final use of the model and method should be considered. It can also be seen that by using the data augmentation method, most of the parameters related to the evaluation of the models have been improved.