This paper will review the different approaches used to discriminate between types of coughs through the limited technology of a micro-phone from a smartphone, as well as the medical issues surrounding coughs around the world. It will first describe a systematic method to identify papers that are relevant to the fields of artificial intelligence, coughs, and cough discrimination, and therefore worthy of review. The following section will discuss our step-by-step process of using the Google Scholar, PubMed, and MIT library search engines to iden-tify relevant papers. Following that, we will provide an overview of the field, including information on the biology of coughing, the different types of coughs, cough recording datasets, and propose methods for cough detection and discrimination. Then, we discuss the research done on COVID-19 related cough discrimination. After-wards, we describe other cough detection related techniques, such as principal component analysis (PCA) and loss change analysis (LCA) techniques, infection testing, pH monitoring, cough monitor-ing, and cough treatment. Finally, we discuss miscellaneous factors such as COVID-19 datasets, privacy, and applications. This review will provide insight into what can be done for early diagnosis of the COVID-19 cough through the rapid recognition of its symptomatic cough to lower the transmission of the virus.

The research papers chosen for this review had to meet the fol-lowing criteria: they discuss a prototype or algorithm aiding in detecting coughs; they use artificial intelligence or neural networks to discriminate between types of coughs; they consist of informa-tion about the biological origin of a cough or the recorded audio of a cough; they discuss chronic cough and its treatments; or they discuss the difference between multiple types of coughs, such as those of COVID-19, the common flu, lung cancer, or others. All these papers ultimately help in computer recognition of various kinds of coughs. Many were found through reviewing papers on coughs and AI, and then examining the papers that cited them. The more cita-tions the research paper had, the easier it was to find other relevant papers. The keywords used were the following, ranked by useful-ness: “cough,” “cough artificial intelligence,” “cough sound analysis,” “cough discrimination,” “covid 19 cough,” “chronic cough,” “suppura-tive airway disease,” “lower airway infection cough,” “cancer cough,” and “gastro oesophageal reflux cough."

As shown in Figure 1, we searched for “"cough” “artificial intel-ligence” symptom discrimination “detection” diagnosis” as well as “"cough” reflex larynx chronic wheeze “infection” “clinic"’ and finally “"chronic cough” treatment unmet needs,” finding a total of 1,461 hits.

We then selected the papers that had the most citations according to Google Scholar, reviewed them, and found 19 additional papers that were relevant. We selected the papers that matched the criteria set forth above.

A paper fit the criteria if it involved cough and treatment needs or cough and cough biology or cough and artificial intelligence or COVID-19 and cough.

All of these were filtered using the criteria described above, and a final list of the 204 most relevant papers was identified.

List 1 includes the reviewed papers on topics related to cough discrimination for various types of cough. List 2 includes the re-viewed papers for various methods used for cough discrimination. References in bold fonts refer to papers that are primarily related to that topic, while references in italic fonts refer to papers that focus on coughing more than its detection. For each topic, the paper with the highest accuracy is listed.

List 1: Method used in cough analysis (8 topics).

We now present 8 topics organized in three categories in relation to how cough issues may be solved:

• Cough Detection.

There have been many methods tested for recording coughs of the highest quality audio, but a microphone array has proven to be very successful. Salmi et al. (1988) analyzed the cough noise by a person on a static charge sensitive bed by passing it through a sampling rate of 30 Hz. The program used an algorithm where it calculated the mean noise levels of the signal. Then, the detection level was multiplied by 4 for acoustic signals and 3 for body movement signals. If both body movement and acoustic signals surpassed the threshold simultaneously, then a cough would be recognized. This algorithm can also be used for sleep and sleep-related apneas. Cough sounds are transients containing frequency components with a range of 80 Hz to around 4,000 Hz, at minimum. Therefore, high-pass filtering can cut off low-frequency noise, or different sounds. The results of recording the coughs over a long period of time with automatic analysis of cough has advantages in high sensitivity and specificity.

Murata et al. (1998) described an experiment used to discriminate between productive coughs, caused by excess airway secretions, and non-productive coughs. In the experiment, some subjects had chronic airway diseases and some were healthy. Coughs with spu-tum were described as productive coughs. The voluntary coughs, on the other hand, were compared with vowel sounds for recording. Then, a sound spectrogram and time-expanded waveform were cre-ated through the cough audio waves. The cough was analyzed in its second phase after the expulsion of sound from the cough. Phase 2 was the longest phase with a length of 105 msec, followed by phase 3 (the noise created by the vocal cords closing) with a length of 90 msec, and lastly by with phase 1 (the expulsion) in last with a length of 50 msec.

Doherty et al. (1997) analyzed the differences acoustically between healthy members when given an induced cough. By giving capsaicin to the set of subjects, a spontaneous cough will be produced. The researchers plotted the data on a spectrogram by overall spectral energy, and root means square pressure plots. The Root-Means-Square plots showed that the most common pattern is a cough with two energy peaks, at the beginning and end. Less common were a single peaked cough sound and more than three high energy peaks. The spectrograms varied greatly, as did the spectral energy chart. Capsaicin is able to produce reproducible cough noises without fear of tachyphylaxis.

Aerts et al. (2005) focused on cough detection in pig houses, where sound equipment was hooked up to the laboratory and recorded the number of coughs for the pigs. When doing this, underestimations of up to 94% in counting coughs were reported. The goal of Drugman et al. (2013) was to study different sensors for cough detection and then test them in an experiment with healthy subjects in a confined room. The performance of the ECG sensors, thermistor, chest belt, accelerometer, contact, and audio microphones outperformed the KarmelSonix system.

Drugman et al. (2013) studied different sensors for cough detection and then tests them in an experiment with healthy subjects in a con-fined room. The performance of the sensors ECG, thermistor, chest belt, accelerometer, contact and audio microphones outperformed the Karmelsonix system.

Moradshahi (2013) tested cough sound discrimination algorithms in noisy environments since reverberation can cause a great deal of inaccuracy in these algorithms. When white noise was added to the system, the success of the discriminator decreased significantly, and when the distance from the microphone was greater, the algorithm could not discriminate between two different cough types. When tested with varying volume of cough sounds, the success of the discriminator also changed. As the volume increased, the success of the discriminator increased, but when the volume reached a certain point, the saturation of the system was so high that the success of the discriminator dropped. To improve upon the single microphone, the researcher used a microphone array and beamforming techniques to improve the performance of the discriminator in these noisy scenarios. The success of the discriminator increased with these new adjustments.

Orlandic (2020) described methods for data collection that can be used for cough analysis algorithms around the world. Its first consideration was the best method for data collection, which asked the user to cough into an elbow with the microphone at arm’s length since coughing is a potentially dangerous activity during a pandemic. Since this dataset uses crowdsourcing, it ran into the problem that many samples will be unrelated to the database’s desired content. Classifiers were used for this database cleansing using power spectral density. This paper has public source code for the XGB classifier that performed this task.

Bagad et al. (2020) proposed an AI model that is able to predict the presence of COVID-19 from cough sounds alone, which have been recorded on a phone application. The data collection procedure can be described in the following steps: subject enrollment, where the users report their demographic information; cough-sound recording, where the users records three separate audio samples of themselves coughing; and testing. The main neural architecture described for this task was the CNN, and the training strategies described were augmentation, pre-training, and label smoothing. It turned out that the performance for this algorithm was similar for both male and female individuals.

Sharma et al. (2020a) discussed the differences between speech and cough noises for those with and without COVID-19 and for those without it. The sound of the cough, the pattern of breathing, the respiration rate, way of speech, and intervals of breathing are all subtle tells that could reveal a decently accurate diagnosis. In those with COVID-19, low pitches with popping and bubbling would be heard. There will be cough sounds for a continuous 30 min, with an episode lasting around half a minute or so. A trademark description of the COVID-19 cough is a dry, barking, hoarse sound. If lungs are injured, a wheezing sound could be heard.

Chaudhari (2020) talked about how datasets from crowdsourc-ing could be used with neural network algorithms to detect COVID-19 from cough audios with a reasonable accuracy. A deep neural net-work was used with the publicly available Coswara and Coughvid datasets of cough sounds. Datasets of smartphone-recorded coughs from South America were also used. MFCCs were used to extract audio features, and then an ensemble of Deep Neural Networks were used for classification. The accurate results of this algorithm demonstrates that crowdsourcing is a moderately accurate method of detecting COVID-19.

Other papers relating to cough detection that were less relevant are: Larson (2011).

In 2020, the main use of neural network architecture and machine learning algorithms for clinical diagnosis was to discriminate ac-curately whether a cough was caused by COVID-19 or some other lung-related disease, for example, through the method shown in Figure 2. Research papers already exist that generally use methods of spectral analysis combined with an array of neu-ral networks. However, the method of recording the data is also called into question, where the recording must be accurate enough to yield accurate results, and it must be affordable enough to be accessed by many around the globe in order to have a large scale impact. Papers Shuja (2010), Pinkas et al. (2020), and Chaudhari (2020) used smartphone applications for recording, which yield near 90% accuracy or above and are is accessible to millions of people around the world. Some of the classifiers that were used for these research papers are: DL-MC, CML-MC, DL-BC, logistic regression, gradient boosting trees, SVM, and RNN. Some of the datasets that were used include ESC-50, national COVID-19 data collection project, Koswara, and Coughvid.

Coppock et al. (2021) presented seven main concerns in using cough audio to train neural networks to detect whether or not one has COVID-19 or not. Firstly, the algorithms may not be specific to COVID-19 and may solely detect the general health of the subject. Secondly, the unfiltered environmental sounds in the background of the cough audio may tamper with the training of the neural networks and could introduce bias–perhaps a correct diagnosis for COVID-19 may be more likely indoors. Thirdly, the patient’s knowledge of whether they have COVID-19 could corrupt neural network training if emotions leak into the voice. Fourthly, many datasets used for this training are not extremely reliable or valid. Fifthly, there are not many public codebases or datasets that can be used for neural network training. Next, demographic characteristics can introduce bias as disease prevalence is not consistent among all regions. Finally, the participant population is largely uncontrolled.

Therefore, reappearing participants will cause the model to function with higher accuracy, inflating the success scores.

Commonly found with neural network techniques are PCA and LCA techniques, which aid in feeding the datasets into the neural network algorithms. Using these techniques correctly is vital in optimizing the performance of the deep learning algorithms. In general, PCA and LCA are used to convert data taken from the cough audio into a usable sequence of numbers that can then be used in a neural networking algorithm. For example, Imran et al. (Ali et al., 2020) created an app that predicts whether a user has COVID-19 based on the sound of their cough. PCA projections are used on the MFCC features from the audio to create feature vectors. Then, these feature vectors are fed into the neural networking algorithm to generate a diagnosis for the cough audio. This diagnosis will then tell the user of the app whether or not the user has COVID-19.

Derraz (2020) created a neural network that can classify a cough as one that is from COVID-19 or from another illness. A PNN, along with DSP and neural networks, was used to differentiate these coughs. However, PCA techniques were used to create feature vectors which are then fed into the neural networks.

Larson (2011) included a great example on how PCA can be used to optimize the deep learning algorithms. While this system used a lengthy random forest classifier, with the help of PCA, it could greatly optimize how the data was being fed into the neural network. In this example, the PCA used orthogonal components and eigenvectors to reduce dimensionality. Then, the components were ranked based on the amount of variation they could explain in the data. With this system, the data was being fed into a neural network in an efficient and relevant way.

Khomsay (2019) used deep learning networks and tensor flow for identifying coughs. However, it needed the help of PCA which performs feature extraction, then sending this data to the Deep Learning Network. LCA is important for statistical analysis and finding distinct subsets in a population with inherent heterogeneity. Spycher et al. (2008) used the LCA method on coughing children. It was able to group the children into various phenotypes based on their type of cough.

Equi (2001) conducted an experiment that compares cough swabs taken from children with cystic fibrosis and children with concomi-tant sputum. The cough swab was a strong predictor of sputum culture, while a negative result did not rule out infection.

The purpose of De Marco et al. (2007) was to discover whether chronic obstructive pulmonary disease (COPD) could be predicted by the presence of phlegm, chronic cough, and dyspnea. After a study was done, a correlation was found which showed that the presence of chronic cough or phlegm is linked with a high risk of developing COPD.

The aim of Blondeau et al. (2007) was to examine the association between cough and weakly acidic reflux by studying a large set of patients with unexplained chronic cough. After this study was done, a positive association was found between the two.

The purpose of Ing et al. (1991) was to examine the correlation between a chronic persistent cough and a gastro-oesophageal reflux. Subjects identified as having chronic cough underwent 24 h am-bulatory oesophageal pH monitoring. These results were compared with a matched control group.

Palombini et al. (1999) described an experiment where the plain chest radiographs of nonsmoking patients who complained of cough for over 3 weeks were examined. pH monitoring was also per-formed. Asthma, PNDS, and GERD were frequently found as culprits, earning them the expression “pathogenic triad of chronic cough.”

Harle (2006) conducted a study of the cough caused by lung cancer. The current treatments are not perfect, and there is not a lot of information about lung cancer and its associated cough. Cough is a big symptom and a frequently unmet clinical need. However, the study has the potential to improve the understanding of therapeutic options for the cough associated with lung cancer.

Birring et al. (2008) examined the effectiveness of the Leicester Cough Monitor (LCM). The LCM was used to measure cough fre-quency, with recordings of up to 6 h. The recording proved that the LCM is reliable enough to be trusted to assess cough frequency in patients for up to 24 h. It may even have the capabilities to work in clinics.

Leconte et al. (2011) examined the effectiveness of the LR102 device, which has the ability to identify cough episodes through a cough frequency meter with electromyography and audio sensors. The LR102 overestimated cough frequency, but still was relatively close and useful, and reduced the time needed for analysis.

Mcgovern et al. (2018) explained where the mechanisms in a cough originate, as well as novel therapeutic ideas for a troublesome cough. Cough hypersensitivity syndrome exists among patients because the peripheral pathological process affects the activity and sensi-tivity of vagal primary afferent nerve fibers. Tissue inflammation can stimulate an increase in sensory neuron activity, which height-ens cough reflex sensitivity. This sensitivity manifests from talking, laughing, or smelling perfumes. Neuroplasticity has an important relation with chronic cough and pain. Viruses cause inflammation in the respiratory system, which stands in the way of sensory neuron function, leading to cough. Plasticity is a contributor to the continu-ation of chronic cough and cough hypersensitivity syndrome among patients. Through understanding the neural pathways responsible for chronic cough, there are new methods of therapy emerging that could target the neural pathways such as targeting molecular pathways central to the development of excessive cough.

Kang et al. (2019) covered how Korean cough patients are not re-ceiving the required treatment that they need. After an observation of the Korean patients, half of them stated that there was a lack of treatment effects, and roughly 30% stated that they were given an unclear diagnosis. Other common unmet needs were that most patients reported difficulties in locating cough specialists or clinics, as well as wanting further information on cough treatment and pre-vention. This is a problem for Korean adult patients, as they describe their chronic cough as impairing their daily actions, frustrating their family and friends, and causing depression.

Fujimura (2003) examined the correlation between atopic cough and cough variant asthma through studying a set of patients with atopic cough and cough variant asthma, as well as the effectiveness of the treatment of inhaling beclomethasone dipropionate (BDP). In conclusion, the study demonstrated that the onset of typical asthma occurred significantly less frequently in those with atopic cough than in those with cough variant asthma. Additionally, the treatment significantly decreased the development of typical asthma in those with cough variant asthma. A significant number of patients with cough variant asthma end up developing typical asthma. Atopic cough has indeed been proposed as a cause of isolated chronic non-productive cough.

Gibson (2016) went into depth on the mistreatment of adult pa-tients with unexplained chronic cough (UCC). The research divided patients with UCC into cough categories of treatment-resistant, idio-pathic, and intractable. Another study divided them into refractory, unexplained, and idiopathic coughs. The paper offers some possible therapies, such as nonpharmacological therapies. An intervention included two to four sessions of education, cough suppression tech-niques, breathing exercises, and counseling, which resulted in a positive impact on cough severity. Inhaled corticosteroids (ICSs) targeted airway inflammation, also reducing cough severity in mul-tiple studies. However, if the patient tests negative for bronchial hyperresponsiveness and eosinophilia, then ICS’s should not be prescribed.

Bowen et al. (2018) explained the causes of UCC, which is a chronic cough that has an unknown origin through a review of patients with UCC. Tachyphylaxis and dependence in pharmacotherapy are suspected to be the main problems, as they are frequently observed among patients with UCC. Clinicians are advised to use tricyclic antidepressants or gabapentin for treatment initiation. Although the chances of a successful treatment will diminish over time, most patients will be treated after several trials.

Redding and Carter (2017) showed the dangers of bronchiectasis and the effect that it can have on children. It is very likely that children who have this condition are underrepresented and untreated.

The aims of Harle et al. (2020) were to examine the cough, its impact, and its severity in the case of lung cancer. By using many cough-specific validated tools in the United Kingdom, a study can be done with lung cancer patients. This study demonstrated that there is an urgent demand for more potent antitussive treatments.

Birring (2011) discussed the treatment of chronic cough. It discussed whether asthma, gastroesophageal reflux, and upper air-way disorders are the causes or aggravants of chronic cough. There is a high demand to understand why a heightened cough reflex sensitivity exists for some patients, and whether genetic, molecular, or physiological concepts have a role to play in that.

The aim of Ryan et al. (2010) was to understand chronic cough and how speech language pathology could manage or improve the condition. Some outcome measures were capsaicin cough reflex sensitivity, automated cough frequency detection, and cough-related quality of life. Speech language pathology management may be able to intervene in the issues of refractory chronic cough.

The main public dataset for COVID-19 is the Coughvid dataset. The Coughvid crowdsourcing dataset contains over 25,000 recordings with 1,155 being those of COVID-19. These coughs originate from all around the world, and this is the largest known public cough dataset that is related to diagnosing COVID-19 in existence. This dataset can be accessed here: https://doi.org/10.6084/m9.figshare.14377019.

However, besides this dataset, there is a shortage of cough recordings that could be used to train neural networks that diagnose COVID-19 with a high accuracy.

In terms of these studies done, publicizing a participant’s name, email address, phone number, or any information that was irrelevant to the data of the study was prohibited. Therefore, every participant maintained privacy in cough recording, cough analysis, or cough treatment. The FDA prohibits use of any cough suppressants in children under the age of 18.

The main commonalities within the research papers are that prepro-cessing is done to the cough audio, involving spectral analysis, and then a neural network is used. In this step, MFCCs and Fast Fourier Transforms are used to convert the audio into a suitable input for a neural network. These neural networks include CNN, Deep Neural Networks, k-NN, Random Forests, hybrids, and combinations of neural networks. In general, it appears that algorithms that use ensembles of neural networks or hybrids tend to be more accurate than those that solely use one type of neural network. It is clear that neural networks improve as the number and size of the datasets increase. Therefore, public datasets are a great help for training the neural networks. Researchers can use this information to cre-ate increasingly accurate neural network architecture that yields extremely accurate results. The more powerful and accessible the algorithm, the cheaper and more widespread an accurate diagnosis.

Public databases of cough recordings such as the one used in the Dicova challenge of this conference, the fkthecovid dataset, and the ESC-50 dataset exist. However, there is a shortage of data for training the neural networks.

1) More data is needed in the field to train the neural networks optimally.

Beyond a need for larger datasets, all areas of the processing pipeline need to improve as we have seen throughout this review. In sum-mary, improvement areas include:

1) First and foremost, a need for larger datasets as we just dis-cussed

Much progress has been made in the detection and discrimination of cough based on neural networks, where optimal algorithms have been found such as shown in Figure 3. Machine learning algorithms have correctly diagnosed many that have been afflicted with all kinds of lung diseases. Further work is needed to bring neural networks into medical practice, which require more public data for training. This will likely be achieved in the near future. Cough discrimination, once performed at a high enough accuracy, can be applied to diag-nosing COVID-19 in clinics as well as in people’s homes, which will greatly reduce the cost of screening as well as increase accessibility. Therefore, people will be able to test themselves faster and more conveniently than ever before. Since COVID-19 is incredibly con-tagious for a large part due to the lack of frequent testing, a more convenient, timely, method such as machine learning can improve testing and do its part in keeping COVID-19 under control.

Other papers that were instrumental to this literature review paper but were unrelated to cough specifically are: (Conference on Human Factors in Computing Systems, 2008), (Ablamowicz and Fauser, 2007), (Patricia, 2007), (Adya et al., 2004), (Akyildiz et al., 2002), (Akyildiz et al., 2007), (Using the amsthm Package, 2015), (Andler, 1979), (David, 2003), (Archer et al., 1984), (Bahl et al., 2004), (Bowman et al., 1993), (Braams, 1991), (Buss et al., 1987), (Clark, 1991), (Kenneth, 1985), (Cohen, 1996), (Cohen et al., 2007), (Conti et al., 2009), (CROSSBOW, 2008), (Dijkstra, 1979), (Douglass et al., 1998), (Dunlop and Basili, 1985), (Ian, 2007), (Simon Fear, 2005), (Gerndt, 1989), (Goossens et al., 1999), (Van Gundy et al., 2007), (Hagerup et al., 1993), (Harel, 1978), (Harel, 1979), (CodeBlue, 2008), (Heering and Klint, 1985), (Herlihy, 1993), (Hollis, 1999), (Hörmander, 1985a), (Hörmander, 1985b), (IEEE, 2004), (Kirschmer and Voight, 2010), (Knuth, 1981), (Knuth, 1981), (Knuth, 1984), (Knuth, 1997), (Kong, 2001a), (Kong, 2001b), (Kong and Thanasankit, 2002), (Kong and Thanasankit, 2003), (Kong, 2004), (Kong, 2005), (Kong, 2006), (Korach et al., 1984), (Jacob, 1994), (Kosiur, 2001), (Lamport, 1986), (Lee and Wexelblat, 1978), (Newton and Kinder, 2005), (Li et al., 2008), (McCracken and Golden, 1990), (Mullender, 1993), (Mumford, 1987), (Natarajan et al., 2007), (Nielson, 1985), (Novak, 2003), (Obama, 2008), (Petrie, 1986), (Poker- Edge .Com, 2006), (Reid, 1980), (SIGCOMM Comput, 1984), (Rous, 2008), (Sadiq et al., 2021), (Saeedi et al., 2010a), (Saeedi et al., 2010b), (Salas and Hille, 1978), (Joseph Scientist, 2008), (Smith et al., 2010), (Spector and Mullender, 1990), (Thornburg, 2001), (TUG, 2017), (Tzamaloukas and Garcia-Luna-Aceves, 2000), (Veytsman, 2022), (Wenzel, 1992), (Werneck et al., 2000), (Werneck et al., 2000), (Culler et al., 2004), (Geiger and Meek, 2005), (Zhou et al., 2008), (Zhou et al., 2010), (Anzaroot and McCallum, 2013), (Anzaroot et al., 2014), (Bornmann et al., 2019), and (R Core Team, 2019).