COVID-19 (coronavirus disease 2019) is a contagious infectious disease caused by the new SARS-CoV-2 (severe acute respiratory syndrome coronavirus 2) virus, which was declared a global pandemic on 11 February 2020, by the World Health Organization (WHO) (World Health Organization, 2020). According to information provided by the Center for Systems Science and Engineering (CSSE) at Johns Hopkins University (JHU), accessed on 31 July 2022, almost 548 million people have been infected and more than 6.4 million people have died in the last two years. In Brazil, more than 33 million cases have been reported (with more than 678,000 deaths), and in Canada, more than 4 million cases have been reported (and almost 43,000 deaths) (Johns Hopkins University, 2022).

Clinical manifestations of COVID-19 infection vary from asymptomatic to symptomatic. This infection affects the respiratory system and includes symptoms such as fever, dry cough, dyspnea, headache, sputum production, hemoptysis, myalgia, fatigue, nausea, vomiting, diarrhea, abdominal pain, and loss of smell and taste (Oliveira et al., 2020; Rai et al., 2021). About 25% of patients with mild-to-moderate COVID-19 have been found to have dysphonia since multiple body structures are linked to human voice generation, such as the lungs, vocal folds, and laryngeal muscle (Suppakitjanusant et al., 2021). Thus, respiratory sounds such as cough, speech, and breath may be an important biomarker for COVID-19 diagnosis.

Audio signals generated by body structures and organs have been widely explored for diagnosis and monitoring of diseases by clinicians and clinical researchers (Brown et al., 2020). In Pramono et al. (2016), a pertussis identification algorithm is proposed by using cough and whoop sounds. They extracted several features, such as Mel-frequency cepstral coefficients (MFCCs) and the zero crossing rate (ZCR), and further used a logistic regression (LR) model-based classifier. Pahar et al. (2021) utilized cough sounds produced by patients with tuberculosis and other lung ailments to distinguish both diseases by MFCCs and the ZCR.

In the same direction, various research studies employing respiratory sounds were conducted for COVID-19 screening (Brown et al., 2020; Casanova et al., 2021; Schuller et al., 2021; Verde et al., 2021; Pahar et al., 2022; Pleva et al., 2022; Sharma et al., 2022; Villa-Parra et al., 2022). Cough and breathing sounds from COVID-19, asthmatic, and healthy individuals were utilized by Brown et al. (2020). A total of 733 dimensional features were extracted by several methods, including MFCCs and the ZCR, which were reduced afterward by principal component analysis (PCA) and tested in classifiers such as LR, gradient boosting trees, and support vector machines (SVMs). In Verde et al. (2021), only voice alterations due to COVID-19 infection were estimated by the main discriminant features used in clinical practice to assess the voice quality and for voice classification, such as MFCCs and spectral roll-off (SR). Experimental research is presented by Pahar et al. (2022) to detect COVID-19 with cough, breath, and speech sounds by using deep transfer learning and bottleneck features, employing the convolution neural network (CNN), long short-term memory network (LSTM), and ResNet50 architecture for classification. On the other hand, in Sharma et al. (2022), a model from textural features and the k-nearest neighbor (KNN) classifier was utilized in COVID-19 screening.

In this work, we propose a texture-based method for COVID-19 respiratory sound classification, which captures texture-related information from its spectrogram representation. We hypothesize that using texture-based features may increase class discrimination with low computational complexity. Although texture-based approaches are widely used in image classification, a lack of research using them in biomedical signal processing is observed. It is to be noted that the various aforementioned research studies utilized acoustic (i.e., ZCR) or deep features (i.e., deep transfer learning) (Sharma et al., 2022). Our study utilized cough, speech, and breath sounds from three different databases to evaluate our proposed methodology. The textural information is captured by employing local binary patterns and local ternary patterns.

This study is structured into four sections as follows. Section 2 describes the dataset, followed by the proposed system to detect the presence of SARS-CoV-2 through cough, speech, and breath analysis. Afterward, the results and discussions are presented in Section 3 and Section 4, respectively, in which the performance of our approach is analyzed. Finally, the conclusions about the proposed method are given in Section 5.

As shown in Figure 1, the workflow of our proposed system to infer COVID-19 by cough, speech, and breath sounds is presented. The following sections will detail each step.

As mentioned previously, two feature extraction methods were evaluated. Figure 4 shows the results of all nine experiments for both CCS and CSS and the Cambridge KDD database for cough, speech, and breath classification with the SVM as a classifier for COVID-19 screening.

In the case of CCS and CSS databases, the main objective of both C1 and C2 experiments is to perform a binary classification between COVID-19 positive and negative patients by using cough and breath signals, respectively. Although LBP and LTP-based features performed the same result for the validation set (UAR = 80.00%), employing cough sounds (C1), LTP-based features achieved better results for the test set (UAR = 74.16%). Regarding C2, where speech samples were evaluated, LBP-based features achieved better results than LTP in both validation (UAR = 74.24%) and test (UAR = 58.52%) sets. In addition to the original challenges, one more experiment with CCS and CSS datasets was also carried out to investigate if the combination of cough and speech sounds could improve the distinction between COVID-19 and healthy individuals. In this experiment (C3), the results obtained in the test set for both LBP and LTP-based features are close. However, they have a significant difference in the validation set for both the UAR and F1-score.

For the Cambridge KDD dataset, six experiments were conducted (K1 to K6) with cough and/or breath sounds. As previously mentioned, the Cambridge KDD dataset has cough and breath audio samples from individuals who reported cough as a symptom or not. For almost all experiments (except K2 and K6), the LBP-based features achieved better or equal results to LTP in both validation and test sets. Also, to perform a binary classification, we conducted experiments (K1–K4) to investigate the impact of COVID-19 screening in patients with cough as a symptom from those with a clean medical history (nonsmoking and no COVID-19 symptoms). Experiment K1 demonstrated good classification by using our approach in COVID-positive and non-COVID individuals (UAR = 88.69%). As expected, in experiment K2, a UAR of 100.00% was achieved using LTP-based features in COVID-positive patients with cough and non-COVID individuals with cough. On the other hand, experiments K3 and K4 demonstrated that cough as a symptom is not a major factor in inferring COVID-19, with a UAR of 94.11% for COVID-positive patients and non-COVID individuals, and 95.00% for COVID-positive patients with cough and non-COVID patients with cough since they obtained good performance using breath sounds.

As explained by Sharma et al. (2022), in experiment K5, we also conducted an experiment using cough and breath audios combined to perform a binary classification between COVID-positive patients with cough and non-COVID individuals with cough. A UAR of 97.37% and F1-score of 0.97 were achieved using our LBP-based feature approach. On the other hand, these results suggest that combining cough and breath sounds can distinguish better between COVID-19-positive and -negative patients than cough and speech sounds together.

Asthma is one of the most prevalent chronic diseases in the United States, and respiratory infections are frequently thought to be asthma triggers (Esmaeilzadeh et al., 2022). Patients with acute asthma attacks frequently describe upper respiratory symptoms, including cough, in the days before the exacerbation starts (Pattemore et al., 1992). Then, we conducted experiment K6 to evaluate if a cough sound from an individual with a respiratory disease (asthma) can be distinguished from a COVID-positive individual with cough. In this experiment, LTP-based features achieved a UAR of 100.00% and F1-score of 1.00 for both validation and test sets.

Since the COVID-19 outbreak, several research studies have been conducted to infer infection by COVID-19 (Brown et al., 2020; Schuller et al., 2021; Zhou et al., 2021; Pahar et al., 2022; Pleva et al., 2022; Sharma et al., 2022; Villa-Parra et al., 2022). From the experiments, we can find that the proposed audio texture feature extraction can achieve a good performance in COVID-19 screening. Coppock et al. (2022) presented a summary of the INTERSPEECH 2021 ComParE. A cough and speech UAR of 75.9% (Casanova et al., 2021) and 72.1% (Schuller et al., 2021) was achieved, respectively. Although we reached a slightly lower UAR for cough (UAR = 75.54%), it is worth noting that we did not use any data augmentation and deep learning methods. On the other hand, Solera-Ureña et al. (2021) achieved a UAR of 69.3% without data augmentation using SVM as a classifier. Unlike for cough, we did not achieve a good performance for speech tasks compared to the baseline shown by Schuller et al. (2021) (UAR = 72.1%). Casanova et al. (2021), when exploring the same approach utilized for cough, had achieved a UAR of 70.3%. Klumpp et al. (2021) explored Mel spectrograms and various classifiers, such as LSTM, CNN, SVM, and LR, with data augmentation, and a UAR of 64.2% was reached. Brown et al. (2020) explored acoustic features in their research, and AUC-ROC up to 0.82 was achieved for binary classification. It is worth mentioning that various research studies utilized the accuracy (ACC) and area under the curve of receiver operating characteristic (AUC-ROC) as metrics. Sharma et al. (2022) analyzed cough, speech, and breath sounds collected from smartphones by using spectrograms and textural features (LBP and Haralick’s), and an ACC of 98.9% and 72.2% for 2- and 5-class classification tasks were reached, respectively. The authors utilized 120-dimensional features for LBP and 14-dimensional Haralick’s features. In contrast, our research utilized 59-dimensional features only. A study conducted by Pahar et al. (2022) explored deep architecture for COVID-19 detection, and they achieved an AUC-ROC of 0.98, 0.94, and 0.92, respectively, for all three sound classes (cough, breath, and speech). Respiratory audio data are also explored for chronic obstructive pulmonary disease (COPD) detection. Srivastava et al. (2021) explored MFCC features and achieved an AUC-ROC of 0.89. In addition to research using respiratory audio in COVID-19 screening, various research studies employing chest X-ray images are being conducted, such as the one by Bhatt et al. (2021). In this work, progressive resizing and transfer learning techniques are explored in normal and COVID-19-infected X-ray images, and an accuracy of up to 100.00% was achieved.

It is to be noted that various research studies explored acoustic features (Brown et al., 2020; Pahar et al., 2022), although textural features are still not much explored in audio analysis, particularly for COVID-19 screening, thus opening avenues in respiratory sound analysis. In general, our results show that LBP achieved better performance than LTP. It is worth mentioning that the main advantage of using LBP textural features is the low computational complexity, thus facilitating a mobile implementation of the proposed system to infer COVID-19 using a smartphone.

In this work, we presented a framework to infer COVID-19 by using cough, speech, and breath audios using textural features. For each respiratory sound, the Mel spectrogram was computed, and two different methods for feature extraction were performed over the Mel spectrogram gray-scale image: local binary patterns and local ternary patterns. We have evaluated the features with an SVM classifier for three different databases. Our study with different feature extraction methods reveals that the LBP-based feature is superior to LTP in most of the experiments. In addition, we have noticed that in a binary classification, cough sounds are better to distinguish between COVID-19-positive and -negative individuals.

As a limitation to our study, we performed the analysis using a small sample size. Although good performance was obtained with our proposal, it may not be a good alternative to detect COVID-19. However, it creates new opportunities to develop COVID-19 screening tools for telemedicine and remote monitoring (Sharma et al. 2022; Villa-Parra et al. 2022). In future works, we will explore data augmentation techniques (Saldanha et al., 2022), transfer learning (Bhatt et al., 2021), and interpretable deep learning models (Joshi et al., 2021) to improve the interpretability and usability of our framework to help COVID-19 diagnosis.