LSTM-Based COVID-19 Detection Method Using Coughing

doi:10.21203/rs.3.rs-2106413/v1

Download PDF

Research Article

LSTM-Based COVID-19 Detection Method Using Coughing

https://doi.org/10.21203/rs.3.rs-2106413/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

COVID-19 has disrupted and irrevocably changed the everyday lives of people all around the world. This viral disease has created the necessity for a contact-free, non-invasive, and easy-to-use diagnostic device. In this paper, we propose a smartphone-based COVID-19 detection method that detects COVID-19 based on the coughing sound of patients. The proposed algorithm segments the coughing sounds collected from the raw audio signals acquired by a smartphone and then detects COVID-19 from the segmented coughing sounds. The proposed algorithm puts raw coughing sounds and the features extracted from the raw sounds into long-term short memory (LSTM), which is known to be effective in the regression and classification of periodic time series signals. Experimental results show that the proposed method applied to the Virufy dataset provides COVID-19 detection accuracy of 92% from the coughing segments. The proposed method has an advantage in pre-diagnosing COVID-19 since the proposed method only requires a smartphone Index Terms—COVID-19, LSTM., machine learning.

COVID-19

LSTM.

machine learning

According to the World Health Organization (WHO), COVID-19 led to approximately 15 million deaths in 2020 and 2021 [1]. Moreover, COVID-19 has devastated healthcare systems around the world [2]. Due to the highly contagious characteristics of COVID-19, a discrete, non-contact, and easy-to-use remote COVID-19 diagnosis tool is required.

Researchers have successfully detected respiratory diseases using audio signals [3]. If the audio sensor is placed near the subject's nose, i.e., at a stable position and a certain distance from the nose of the subject, a smartphone microphone can pick up tiny or loud audio sounds [4]. Audio signals have been used in monitoring sleep and diagnosing diverse diseases, including pulmonary tract infection, pneumonia, and chronic obstructive pulmonary disease (COPD) [5]. The audio signal gives the coughing sound, which is useful information for diagnosing respiratory disease, pharyngitis, laryngitis, sinusitis, otitis media, influenza-positive forms, and the common cold [6]. Especially, COVID-19 affects the sound of someone's coughing, breathing, and voice tone since it infects the respiratory system [7].

There have been studies that detect respiratory diseases using hardware such as external microphones [8], electronic patches [9], and mobile phones [10]. However, this additional hardware requires setup or installation, and users must pay to buy it. On the other hand, software-based respiratory disease detection methods detect diseases using time-/frequency-domain features [4, 5]. For example, Mel-Scaled Spectrogram, Mel frequency cepstral coefficients (MFCC), tonal centroid, chromagram, and spectral contrast were extracted [11], and these variables are put into an ensemble of machine learning to

detect COVID-19. As the other example of software-based COVID-19 detection methods, in [12], MFCC, first order MFCC, second order MFCC, and zero crossing rate (ZCR) features are extracted from breathing, coughing, and speech signals and then put these features as input into a neural network for COVID-19 detection purposes. In [13], raw audio signals are converted into spectrogram images, which are used to train their proposed machine learning algorithm to detect COVID-19. Machine learning algorithms have been applied to detect COVID-19.

Machine learning algorithms have been proposed to detect COVID-19. In [11], an ensemble-based multi- criteria decision-making (MCDM) approach was proposed to classify COVID-19 and healthy subjects. Here, the MCDM is adopted to select the feature based on the order preference or similarity. Logistic regression (LR), K-nearest neighbor (KNN), support vector machines (SVM), and decision tree algorithms were used as classifiers to detect COVID-19 in [12]. In [13], an ensemble-based deep neural network is adopted to identify COVID-19 and is tested on the breathing and coughing dataset. An ensemble-based model is also adopted in [14] to train with spectrogram images from raw coughing data.

Among the machine learning algorithms, a recurrent neural network (RNN) is strong when applied to sequential or time series data [15]. Traditional neural networks consider each data section independent [16], while RNN can predict the future data sequence based on the last data samples. Specifically, the RNN combines the previous sequence’s output with the next input [16]. In [12], the RNN was implemented to differentiate between COVID-positive and healthy people. However, the RNN has long-term dependencies and is challenging to train for longer sequences of data due to the vanishing and exploding gradients [17]. LSTM is a particular type of RNN with memory cells, which addresses these challenges. Due to this advantage of LSTM over RNN, LSTM has been widely used in regressing or classifying time-series signals [12].

COVID-19 coughing open datasets are available online in CoughVid [18], Coswara [19], Virufy [20], NoCoCoDa [21]. CoughVid has 150 patients’ signals divided into four categories - COVID-19, asthma, bronchitis, and healthy - further simplified and separated into COVID-19 and healthy subjects. The comprehensive data collected from these patients have 30,000 audio segments and 328 cough sounds. In [14], the coughing signal from the COUGHVID dataset and spectrogram images of the coughing signal are fed into VGG-13, convolutional recurrent neural network (CRNN), and gated convolutional neural network (GCNN).

In this paper, we propose an LSTM-based COVID-19 detection method that uses the raw coughing signal itself and its features. We adopt the raw coughing signals from the Virufy dataset [20]. Here, features are extracted from the raw signal in the time-/frequency-domain. We divide the coughing signal dataset into training and testing sets. Then, the coughing signal and its features from the training set were used to train the LSTM network. The rest of this paper is organized as follows: Section 2 describes the dataset and pre-processing used in this paper. The proposed algorithm is presented in Section 3. Section 4 describes our experimental results. Finally, Section 5 concludes this paper.

2.1. Dataset

The Virufy open-source database was used to test the proposed method [6]. It consists of coughing sound files collected from 16 subjects using a smartphone. Among 16 subjects, seven subjects are COVID-19 positive while 9 are healthy patients. The length of the coughing sound files is 15 to 17 seconds.

2.2. Preprocessing of Raw Signal in Dataset

Virufy's audio signal has a sampling frequency of 44,800 Hz. The proposed algorithm divides the audio signal into time intervals of 1 second for fast processing, called a segment, as shown in Fig. 1. Successive segments are overlapped with an overlap interval of 50 ms. 75% of the data is used for training, and the remaining 25% is used for testing. Data points from training and test data were used to train an LSTM network.

Figure 2 shows the flowchart of our overall proposed COVID-19 detection method which classifies COVID-19 and healthy subjects from coughing events using LSTM deep learning model.

3.1. Features extracted from audio segments for diagnosing COVID-19

LSTM models are trained by raw signals and extracted features. Since the characteristics found in speaking, coughing, and breathing are different, the features extracted from each sound wave are expected to be different. The feature extraction is performed on audio signals, as shown in Fig. 3. As features, we consider spectral centroid, spectral roll-off, zero crossing rate, and Mel frequency cepstral coefficients (MFCC). The details of these parameters are described in the following.:

Spectral Centroid: Spectral centroid is defined as the location of the center of the spectrum [22]. It is calculated by a weighted average of the frequencies, as shown in (1) [23]:

$$Spectral Centroid= \frac{{\sum }_{k=1}^{N}kF\left[k\right]}{{\sum }_{k=1}^{N}F\left[k\right]}$$

where k is the frequency value, N is the maximum frequency value, F[k] is the weight of the frequency [24]. The lower the spectral centroid, the higher the presence of coughing-related sounds in the cough signal [25]. Hence, the spectral centroid value of the coughing segment obtained from a COVID-19 positive patient is expected to be lower than that obtained from a healthy subject.

Spectral Roll-off

The spectral roll-off is defined to be the frequency below which the specific portion of the total spectral energy resides [26]. For each frame, the spectral roll-off value is calculated as

$$Spectral Roll-off= C*{\sum }_{n=1}^{N}{F}_{k}\left[n\right]$$

where ${F}_{k}$ is the amplitude of the audio signal in the kth bin, C is the percentage of the frequency. In this paper, we set the value of C to be $0.85$. The spectral roll-off value of the coughing segment obtained from COVID-19 patients is expected to be smaller than that of healthy subjects [26].

Zero Crossing

Zero crossing rate is defined to be the number of times that the signal crosses the value zero and is calculated for the segment as follows

$${Z}_{t}=\frac{1}{N-1}{\sum }_{t=1}^{N-1}ᴨ\left[y\left(t\right)y\left(t-1\right)\right]$$

where ᴨ(A) is an indicator function where it is only 1 for terms where A is less than 0, and otherwise it is equal to 0, $y\left(t\right)$ is the value of the signal $y$ at time index t, N represents the number of samples in the signal [27].

Mel frequency cepstral coefficients (MFCC)

MFCC is calculated using the corner frequency of 700 as in (4), which filters the frequency g above 700 in a logarithmic way while the frequency g below 700 is filtered linearly [25].

$$MFCC\left(g\right)=2595*log(1+\frac{g}{700})$$

3.2. LSTM Algorithm

The extracted features and raw signal segments are used to train our proposed bidirectional long short-term memory (LSTM) model. An LSTM unit comprises a memory cell and three gates: input, forget, and output. In step t, the LSTM unit accepts the input and the hidden unit outputs from the previous step, as shown in Fig. 5a [28]. The LSTM is able to remember information over the long term. The LSTM is a robust version of the RNN since the LSTM can prevent the vanishing or exploding gradient problem [29, 30] during backpropagation. In the case of long data sequences in training an RNN model, the gradient becomes too small or too large, making it difficult for the RNN to keep learning or updating the weights [29]. LSTM solves this problem by using a unique additive gradient structure that includes direct access to the forget gate's activations. It enables the network to encourage desired behavior from the error gradient using frequent gate updates at every learning process step [16].

Figure 5a shows a sample LSTM network which has a memory cell with the same structure as the hidden state${h}_{t}$ storing additional information. However, the memory cell can provide more functionalities than the hidden state by different types of gates (i.e., input, update, output, hidden gates) implemented inside the memory cell. An input gate${x}_{t}$ determines when to read data into the cell while an output gate reads the data from the cell. A forget gate controls how much information the memory cell will receive from the memory cell from the previous step. Such a design uses a specialized mechanism to select when to recall or ignore inputs in the hidden state. Figure 5a shows the different gates in the LSTM model. The LSTM model takes a sequence as input. Each sample is segmented, and each segment is a feature that is the input to the LSTM layer. We adopted bidirectional LSTM (Bi-LSTM) since the Bi-LSTM is widely used in regressing or classifying dependent sequential data.

3.3. Bi-directional LSTM Algorithm

The Bi-LSTM is an improvement on the traditional LSTM as it improves the classification accuracy. Bi-LSTM classifiers has both the forward and backward modes, which leads to higher accuracy. For example, in this paper, we apply the Bi-LSTM by processing coughing input signals heading from left to right (front) direction and from right to left (back) [31] as shown in Fig. 5b. The Bi-LSTM exploits the context and predicts its respective classes. Denoting considering time steps, output, internal state, hidden state, each state input, and the weight by t, y, c, h, x, and W respectively, they have relations as follows:

(${{y}_{t}}^{front}$,${{h}_{t}}^{front}, {{c}_{t}}^{front})=$${LSTM}^{front}$(${{c}_{t-1}}^{front}, {{h}_{t-1}}^{front},{{x}_{t};W}^{front}$) (5)

(${{y}_{t}}^{back}$,${{h}_{t}}^{back}, {{c}_{t}}^{back})=$${LSTM}^{back}$(${{c}_{t-1}}^{back}, {{h}_{t-1}}^{back},{{x}_{t};W}^{back}$) (6)

The LSTM layer considers frontal direction only as in (5) while the Bi-LSTM considers both (5) and (6). Bi-LSTM is more computationally expensive and slower than the LSTM due to the two layers or directions of training [32].

3.4. Bi-LSTM COVID-19 Classification Method

Our proposed LSTM network based on extracted features or the raw signal fragments can be implemented as shown in Fig. 6. The LSTM with features is expected to be effective if the features are time-dependent. One of the reasons to perform feature extraction is to deal with a large amount of data and to reduce complexity without losing any information. The feature extraction reduces the computational load while providing reliable accuracy. The features help to reduce the redundant information from the data.

4.1. Feature Values of Coughing from COVID-19 patients and Healthy Subjects

Figure 7. Comparison of the variation of time domain and frequency domain features between COVID-19 positive and negative coughing signals. (a-b) raw audio signals before processing (c-d) spectral centroid of the signals where COVID-19 positive cough is seen to have a lower amplitude overall (e-f) spectral roll-off is compared with COVID-19 positive signal having a lower roll-off band (g-h) zero crossing rate is compared and the COVID-19 positive cough has a lower number of peak (i-j) MFCC is also compared to show the distinct difference between the shape of signals between an healthy and COVID-19 patient.

A clear distinction between the positive and the healthy audio signal can be observed in the comparative analysis. The spectral frequency is shown to have a higher value when the signal is from a healthy person. This is also clearly visible when Figs. 6c and 6d are compared for the spectral centroids. The spectral roll-off provides a clear idea of the region or frequency range in which the energy is concentrated. Figures 6e and 6f show that the energy density is higher in the low-frequency region for the COVID-19 patient. Figures 10g and 10h show that the zero crossing rate values from a COVID-19 positive patient are lower than those from a healthy subject. The MFCC quantifies the spectrum space, including the amplitude and spacing information, as shown in Figs. 6i and 6j.

In the proposed method, as shown in Fig. 5, the extracted features were used to form a time series that shows how each feature progresses with time. The feature vector is then fed to the proposed LSTM architecture. In bidirectional LSTM, the coughing signal is fed to the LSTM block forward and backward. The number of hidden layers and the number of epochs in the LSTM layer are fixed based on trial-and-error methods to reach the optimum value, which does not increase the time complexity too much but increases accuracy sufficiently. The fully connected layer takes input from the LSTM layer. The fully connected layer for LSTM acts independently for each sample and assigns the weight. The SoftMax layer calculates the output for every sample.

4.2. Hyperparameters tuning for COVID diagnosis

Hyperparameters are parameters whose values control the network's learning process and determine the importance of model parameters that an algorithm ends up learning.

4.2.1. Minibatch

The data was processed in mini-batches (a consecutive subset of the dataset) to produce more stable gradient descent convergence and reduction in training time. Due to the data being a time series, a minibatch size of 35 was used to ensure lesser noise in the training process. There was a total of 9,300 data points for each raw signal segmented by 16 which is visible in Fig. 9 with the green and blue arrows. Minibatch size 35 means 35 data points in each batch, represented by the curly brackets in Fig. 9. It shows the usage of minibatch for the data segmented by 16.

4.2.2. Hidden Layers

The hidden layers of an LSTM model increase the model's depth and increase the model's prediction accuracy [33]. Both 50 and 75 hidden layers were used due to the size and nature of our data. The coughing signal is a time series data. The data had higher interdependence within the signal. As extraction of information from every segment of the data was done for the correct prediction of the proposed model, a larger remote unit was chosen to collect information from the time series coughing data. The fully connected layer can introduce any kind of dependence. Two fully connected layers have been used to introduce the interdependence of two different classes (COVID-19 vs. healthy), and classification has been done by using the SoftMax layer. SoftMax activation function has been used to classify multiple classes.

4.2.3. Epochs

The epoch size is a crucial component of the neural model. It measures how much the weight is updated for the training process [34]. It is key to find models with lower error [35]. The epoch size is a parameter that must be optimized to prevent both over and under-fitting of the model [35]. The accuracy varies with several epochs and hidden layer sizes. The epoch size value of 50 was chosen by evaluating our proposed model’s training and validation error.

The effect of these hyperparameters (Mini Batch size, Hidden layer number and epochs) on the accuracy is shown in Fig. 10.

4.2.4. Adam optimizer

Since the input signal was a long-sequenced time series data and there were an enormous number of data samples after segmentation, utilizing an optimizer that can deal with the larger dataset and is computationally efficient was necessary. Adam optimizer was used since it is easy to implement and works efficiently with larger datasets and larger parameters. The time complexities are compared to find the best fit. The Adam optimizer has been used during the training processes to lower the time complexity.

4.2.5. Learning rate

The learning rate defines how quickly the network learns its parameters. A learning rate of .001 was used. Except for the case when training the dataset with three consecutive periods (as the data is smaller) from the raw signals, the hidden layer number and epoch size were kept consistent.

4.3. COVID-19 Detection Accuracy of Proposed Method

The training of sequence to label learning takes longer to learn as it takes output only for the last sequence from the fully connected layer once. The time taken to converge is much longer, and the curve does not reach a steady or stable condition compared to the sequence. Sequence-to-sequence learns from every input to the fully connected layer. It converges faster and provides higher accuracy at a much lower epoch. The sorting of the data is done to make the padding process efficient. Since the data size is different for each audio signal, zero padding is necessary for data to make the size equal in small batches. The data is sorted and padded to improve accuracy. The time complexity analysis for sorting and balancing the dataset is shown in Fig. 12.

The lowest training time was for the balanced sequence to sequence training. The difference in balancing is not vast as the dataset used was relatively small. In larger datasets, the effect of balancing would further improve training time. The sorting process increased the training time as the segments were similarly sized due to even segmentation, so it did not further improve the time complexity. Table I shows the model parameters for training for the best possible outcome for the raw signals.

TABLE I. THE MODEL PARAMETERS WITH THE OPTIMIZED HYPERPARAMETER AND TEST RESULTS

Methods

Learning

Hidden layer (HL) and Epoch (E)

Training Accuracy

(A) and Time (T)

Test Accuracy

Segment with only 3periods, segmentation by 10) with unbalanced data

Sequence to Sequence

HL: 75

E: 70

A: 99%

T: 555s

75%

The model was trained with both the raw signals as well as the extracted features. The LSTM architecture was used to train the proposed model with the extracted features. The model uses 50 hidden units and 70 epochs depending on the small size of the data and the time-series nature. The sequence-to-sequence earning the highest test accuracy was 92% for all the intervals. The parameters and accuracy of the proposed model are shown in Table II.

Table 2

RESULTS FOR FEATURE EXTRACTION
Hidden Layer	Epochs	Training Accuracy (A) and Time (T)	Test Accuracy
50	70	A: 94% T :49 s	92%

4.4. Performance comparison of raw signals and features

Raw signals and features were both used for the training of the model. A comparison of accuracy between feature extraction with LSTM and proposed raw signal with LSTM is shown in Fig. 13.

Comparison between the proposed model was made with the two latest literature [36] and [17]. Both used the Virufy dataset but different classification methods. In [36], COVID-19 has been differentiated using feature extraction and classification techniques. First, they create a data pool of 121 segments from the cough samples of 16 subjects. After segmentation, there were 73 healthy and 48 COVID-19 samples. The efficient feature extraction method has used the power spectral density based on short-time Fourier transform and Mel-frequency cepstral coefficients (MFCC). Then SVM was applied to the processed signal. Regarding the result evaluation, SVM's radial basis function (RBF) kernel function had the highest accuracy. They applied cross-validation methods, i.e., leave-one-out (LOO) and hold-out (HO) methods. The highest accuracy obtained was 94.21% using RBF kernel and MFCC feature extraction with LO cross-validation. The second highest accuracy was 87.70% for SVM classification with MFCC feature extraction for two training datasets splitting HO cross-validation.

In [37], the author presented four machine learning methods, SVM, K nearest neighbor (KNN), Random Forest (RF), and recurrent neural network (RNN), to identify COVID-19 from recorded coughing sounds. They implemented their Model to publicly open-source datasets. One is Coswara, and the other is Virufy, the exact dataset we used for our Model. Their proposed method is comprised of five different steps. Firstly, they eliminate the silence segment of recorded audio and extract probable cough sounds. Coughs are then mixed in with the conversations between persons. Machine learning models were used to discriminate between coughing sounds and typical conversation. Secondly, useful features were extracted to train the Model, such as entropy of Mel-Frequency Cepstral Coefficient, Zero Crossing Rate, and Energy. In the third step, they implemented unsupervised learning to recognize the actual cough and remove the sounds that are not cough. The fourth step is to train the deep learning model with extracted features as the input and finally provide the diagnosis by identifying whether COVID-19 causes the cough or not. They implemented SVM, KNN, RF, and simple RNN. The highest obtained accuracy for the Virufy dataset was 81.25% using simple RNN.

Using the same dataset, the LSTM model proposed in this paper gives an accuracy of 92%, which outperforms the other two existing methods, as shown in Fig. 15. Additionally, the proposed method allows the recording of raw coughing sound for a few seconds and segments it automatically. Finally, it automatically diagnoses COVID-19 positive and COVID-19 negative by using the LSTM network.

A comparison of the accuracy using the same dataset between the existing method and our proposed method has been shown in Fig. 14.

"https://www.who.int/news/item/05-05-2022-14.9-million-excess-deaths-were-associated-with-the-covid-19-pandemic-in-2020-and-2021." (accessed 2022).
M. Salathé et al., "COVID-19 epidemic in Switzerland: on the importance of testing, contact tracing and isolation," Swiss medical weekly, vol. 150, no. 1112, 2020.
A. Srivastava, S. Jain, R. Miranda, S. Patil, S. Pandya, and K. Kotecha, "Deep learning based respiratory sound analysis for detection of chronic obstructive pulmonary disease," PeerJ Computer Science, vol. 7, p. e369, 2021.
Y. Nam, B. A. Reyes, and K. H. Chon, "Estimation of respiratory rates using the built-in microphone of a smartphone or headset," IEEE journal of biomedical and health informatics, vol. 20, no. 6, pp. 1493-1501, 2015.
E. Agu et al., "The smartphone as a medical device: Assessing enablers, benefits and challenges," in 2013 IEEE International Workshop of Internet-of-Things Networking and Control (IoT-NC), 2013: IEEE, pp. 48-52.
G. Botha et al., "Detection of tuberculosis by automatic cough sound analysis," Physiological measurement, vol. 39, no. 4, p. 045005, 2018.
"https://www.hopkinsmedicine.org/health/conditions-and-diseases/coronavirus/what-coronavirus-does-to-thelungs#:~:text=Acute%20Respiratory%20Distress%20Syndrome%20." (accessed 2022).
E. C. Larson, T. Lee, S. Liu, M. Rosenfeld, and S. N. Patel, "Accurate and privacy preserving cough sensing using a low-cost microphone," in Proceedings of the 13th international conference on Ubiquitous computing, 2011, pp. 375-384.
T. Elfaramawy, C. L. Fall, M. Morissette, F. Lellouche, and B. Gosselin, "Wireless respiratory monitoring and coughing detection using a wearable patch sensor network," in 2017 15th IEEE International New Circuits and Systems Conference (NEWCAS), 2017: IEEE, pp. 197-200.
C. Pham, "MobiCough: real-time cough detection and monitoring using low-cost mobile devices," in Asian Conference on Intelligent Information and Database Systems, 2016: Springer, pp. 300-309.
N. K. Chowdhury, M. A. Kabir, M. M. Rahman, and S. M. S. Islam, "Machine learning for detecting COVID-19 from cough sounds: An ensemble-based MCDM method," Computers in Biology and Medicine, p. 105405, 2022.
A. Hassan, I. Shahin, and M. B. Alsabek, "Covid-19 detection system using recurrent neural networks," in 2020 International conference on communications, computing, cybersecurity, and informatics (CCCI), 2020: IEEE, pp. 1-5.
B. W. Schuller, H. Coppock, and A. Gaskell, "Detecting COVID-19 from breathing and coughing sounds using deep neural networks," arXiv preprint arXiv:2012.14553, 2020.
S. Rao, V. Narayanaswamy, M. Esposito, J. Thiagarajan, and A. Spanias, "Deep Learning with hyper-parameter tuning for COVID-19 Cough Detection," in 2021 12th International Conference on Information, Intelligence, Systems & Applications (IISA), 2021: IEEE, pp. 1-5.
https://www.oreilly.com/library/view/neural-networks-and/9781492037354/ch04.htm. (accessed 2022).
"https://medium.datadriveninvestor.com/how-do-lstm-networks-solve-the-problem-of-vanishing-gradients-a6784971a577#:~:text=LSTMs%20solve%20the%20problem%20using,step%20of%20the%20learning%20process." (accessed 2022).
H. A. Rashid, A. N. Mazumder, U. P. K. Niyogi, and T. Mohsenin, "CoughNet: A Flexible Low Power CNN-LSTM Processor for Cough Sound Detection," in 2021 IEEE 3rd International Conference on Artificial Intelligence Circuits and Systems (AICAS), 6-9 June 2021 2021, pp. 1-4, doi: 10.1109/AICAS51828.2021.9458509.
L. Orlandic, T. Teijeiro, and D. Atienza, "The COUGHVID crowdsourcing dataset: A corpus for the study of large-scale cough analysis algorithms (Version 1.0)[Data set]," ed: Zenodo, 2020.
"https://www.kaggle.com/datasets/sarabhian/coswara-dataset-heavy-cough?resource=download." (accessed 2022).
"https://virufy.org/en/." (accessed 2022).
M. Cohen-McFarlane, R. Goubran, and F. Knoefel, "Novel coronavirus cough database: Nococoda," Ieee Access, vol. 8, pp. 154087-154094, 2020.
"https://www.sciencedirect.com/topics/engineering/spectral-centroid#:~:text=The%20spectral%20centroid%20(SC)%3A,transform%20frequency%20and%20amplitude%20information." (accessed 2022).
"https://ccrma.stanford.edu/~unjung/AIR/areaExam.pdf." (accessed 2022).
"https://en.wikipedia.org/wiki/Spectral_centroid." (accessed 2022).
R. Islam, E. Abdel-Raheem, and M. Tarique, "A study of using cough sounds and deep neural networks for the early detection of COVID-19," Biomedical Engineering Advances, vol. 3, p. 100025, 2022.
"http://www.egr.msu.edu/~aviyente/ece366pd09.pdf." (accessed 2022).
A. Pal and M. Sankarasubbu, "Pay attention to the cough: Early diagnosis of COVID-19 using interpretable symptoms embeddings with cough sound signal processing," in Proceedings of the 36th Annual ACM Symposium on Applied Computing, 2021, pp. 620-628.
M. You, W. Wang, Y. Li, J. Liu, X. Xu, and Z. Qiu, "Automatic cough detection from realistic audio recordings using C-BiLSTM with boundary regression," Biomedical Signal Processing and Control, vol. 72, p. 103304, 2022/02/01/ 2022, doi: https://doi.org/10.1016/j.bspc.2021.103304.
https://www.superdatascience.com/blogs/recurrent-neural-network s-rnn-the-vanishing-gradient-problem. (accessed 2022).
A. N. Shewalkar, "Comparison of rnn, lstm and gru on speech recognition data," 2018.
"https://www.baeldung.com/cs/bidirectional-vs-unidirectional-lstm#:~:text=As%20a%20result%2C%20the%20BiLSTM,handwritten%20recognition%2C%20and%20similar%20fields." (accessed 2022).
R. L. Abduljabbar, H. Dia, and P.-W. Tsai, "Unidirectional and Bidirectional LSTM Models for Short-Term Traffic Prediction," Journal of Advanced Transportation, vol. 2021, p. 5589075, 2021/03/26 2021, doi: 10.1155/2021/5589075.
"https://machinelearningmastery.com/stacked-long-short-term-memorynetworks/#:~:text=Stacking%20LSTM%20hidden%20layers%20makes,range%20of%20challenging%20prediction%20problems." (accessed 2022).
"https://towardsdatascience.com/epoch-vs-iterations-vs-batch-size-4dfb9c7ce9c9#:~:text=As%20the%20number%20of%20epochs,to%20optimal%20to%20overfitting%20curve." (accessed 2022).
"https://logicplum.com/knowledge-base/epoch/." (accessed 2022).
N. Melek Manshouri, "Identifying COVID-19 by using spectral analysis of cough recordings: a distinctive classification study," Cognitive Neurodynamics, vol. 16, no. 1, pp. 239-253, 2022.
K. Feng, F. He, J. Steinmann, and I. Demirkiran, "Deep-learning based approach to identify covid-19," in SoutheastCon 2021, 2021: IEEE, pp. 1-4.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

LSTM-Based COVID-19 Detection Method Using Coughing

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials

2.1. Dataset

2.2. Preprocessing of Raw Signal in Dataset

3. Methods

3.1. Features extracted from audio segments for diagnosing COVID-19

3.2. LSTM Algorithm

3.3. Bi-directional LSTM Algorithm

3.4. Bi-LSTM COVID-19 Classification Method

4. Result

4.1. Feature Values of Coughing from COVID-19 patients and Healthy Subjects

4.2. Hyperparameters tuning for COVID diagnosis

4.2.1. Minibatch

4.2.2. Hidden Layers

4.2.3. Epochs

4.2.4. Adam optimizer

4.2.5. Learning rate

4.3. COVID-19 Detection Accuracy of Proposed Method

4.4. Performance comparison of raw signals and features

5. Discussion And Conclusion

References

Additional Declarations

Status:

Version 1