Smart Pre-Screening of Patients for COVID-19 using Depth-wise Separable Convolutional Neural Networks on Cough Sounds

Abstract

Hospitals have become a breeding ground for pathogens, as well as an overcrowded and overburdened environment. Even the testing mechanisms and kits are limited and expensive to produce. Coronavirus Disease 2019 (COVID-19) is an infectious respiratory and vascular condition (blood vessel). Coronavirus 2 (SARS-CoV-2) infection is a specific form of coronavirus that is a severe acute respiratory syndrome. Common symptoms are fever, cough, weakness alongside a lack of smell and taste. Further, breathing or respiratory problems are also common. The incubation period can vary from 1 to 14 days, which is between virus infection and symptoms.  COVID-19 condition has very similar symptoms to colds, flu, strep throat, and other viral and bacterial diseases. In general, the symptoms do not distinguish COVID-19 from other conditions. For example, a flu test or a strep throat test may be prescribed by a physician after diagnosing symptoms. In fact, the diagnosis entirely based on disease symptoms may be prone to errors. Cough is a common symptom found in many conditions of breathing illness. Many medical literatures emphasize the importance of automatic, objective and reliable cough audio signal analysing systems that promise to detect pulmonary disorders. Cough, when chronic, is very embarrassing, unpleasant, distressing and degrades the quality of life of patients. Healthcare costs associated with consultations on pulmonary medicine impose heavy financial burdens on the patients. The audio analysis of cough is non-invasive and cheap for the diagnosis of pulmonary disorders. This study investigated if sophisticated machine learning algorithms could help diagnose COVID-19 utilizing cough audio signals. The main challenge is finding large amounts of data for good accuracy and reduced complexity of system to result in the predictable outcome. The other challenges are extracting the right features from the data and is distinguishable to noise, finding the memory requirements of devices portable enough to sense, analyse the real time cough audio data and give predictions on the person’s health conditions.

For such diseases, we propose a mechanism that combines Machine Learning with Neural Network Techniques and Hardware. Cepstral coefficients can be determined by analysing cough audio signals, which is what this device is designed for. By pre-screening and only sending patients who are more likely to be infected to a hospital, we can reduce the burden on the healthcare system.

When it comes to COVID -19 detection, there are currently three approaches. [1] Using bidirectional gated recurrent unit (BiGRU-AT), long short-term memory (LSTM), and other mathematical modelling, we first screen the patient's face with image recognition and measure their breathing rate while wearing a mask. Asthma and tuberculosis were detected in the past using custom classification and clustering algorithms embedded in smart stethoscopes. Problematically, it does not take into account other factors such as external heat signatures and mask interference when determining anomalies. To improve accuracy and efficiency, some other methods for obtaining the respiration rate without physical contact involve using a microphone, speaker, and mathematical modelling. Doppler's effect is used to calculate results, which is dependent on the power spectrum density. Since no monitoring is required, only the test subject's movement during experimentation appears to be a problem. Heart rate and pulse readings can be obtained using Continuous Wavelet Transformation (CWT), which has been shown to be effective in detecting anomalies like Respiratory Sinus Arrhythmia (RSA).

On the other hand, X-ray images [5] and Computer Tomography (CT) scans of the lungs have been segmented by using methods such as lobe segmentation, and the likelihood of COVID has been predicted by comparing the segmented images to previous data. In the past, the use of 3dRD Nets and CNNS for tuberculosis detection has been proven successful [6]. Others include dual sampling using Grad-CAM [7], attention mapping and Moth Flame Algorithm (MFA) [8, Bacteria Foraging Algorithm [9], and Whale Optimization Algorithm and Locust Search Algorithm [10]. As long as resolution and segmentation can be improved without sacrificing accuracy, high threshold values won't work as effectively as lower threshold values. IMPA [11] (Improved Marine Predators Algorithm) has also been tried, and it has outperformed the other algorithms in terms of classification and disease screening.

According to a third method, coughing sounds are analysed in order to determine whether or not a person has infectious diseases, as coughing is the first sign of diseases such as tuberculosis and influenza. After classifying the cough as voluntary or involuntary, the next step is to determine whether it is dry or wet [12] and try to match it with the Disease Cough Sound Database. Also, there is a Leicester Cough Monitor based on Hidden Markov Models [13], which can be used to identify it. In addition, the authors [15] have improved and implemented HMM-based models to achieve an AUC of 92 percent, which is an impressive feat.

Our goal was to create a system that could not only extract and detect cough, but also classify diseases based on it using the previously mentioned factors. Aside from reducing overall complexity, our model is more deployable and efficient because it reduces overall complexity