Fast and noninvasive electronic nose for sniffing out COVID-19 based on exhaled breath-print recognition

Abstract

The reverse transcription-quantitative polymerase chain reaction (RT-qPCR) approach has been widely used to detect the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). However, instead of using it alone, clinicians often prefer to diagnose the coronavirus disease 2019 (COVID-19) by utilizing a combination of clinical signs and symptoms, laboratory test, imaging measurement (e.g., chest computed tomography scan), and multivariable clinical prediction models, including the electronic nose. Here, we report on the development and use of a low cost, noninvasive method to rapidly sniff out COVID-19 based on a portable electronic nose (GeNose C19) integrating an array of metal oxide semiconductor gas sensors, optimized feature extraction, and machine learning models. This approach was evaluated in profiling tests involving a total of 615 breath samples composed of 333 positive and 282 negative samples. The samples were obtained from 43 positive and 40 negative COVID-19 patients, respectively, and confirmed with RT-qPCR at two hospitals located in the Special Region of Yogyakarta, Indonesia. Four different machine learning algorithms (i.e., linear discriminant analysis, support vector machine, stacked multilayer perceptron, and deep neural network) were utilized to identify the top-performing pattern recognition methods and to obtain a high system detection accuracy (88-95%), sensitivity (86-94%), and specificity (88-95%) levels from the testing datasets. Our results suggest that GeNose C19 can be considered a highly potential breathalyzer for fast COVID-19 screening.

Fig. 1. Illustration of the fast and noninvasive COVID-19 detection utilizing portable electronic nose (GeNose C19) integrated with artificial intelligence (AI).

Four different machine learning algorithms (i.e., linear discriminant analysis (LDA), support vector machine (SVM), stacked multilayer perceptron (MLP), and deep neural network (DNN)) were employed to differentiate and classify the exhaled breath patterns of the patients, which were measured by 10 different metal oxide semiconductor gas sensors.

Fig. 2. Integrated GeNose C19 system and its components.

a Scheme of the off-line breath sampling pipeline for GeNose C19: (1) air inhaled through the nose and subsequently exhaled through the mouth to a sampling bag, (2) sealing or closing of the sampling bag cap to avoid the collected air leakage, and (3) direct plugging of the sampling bag into the electronic nose inlet (sample connector). b Diagram and c photograph of GeNose C19 integrated with a high-efficiency particulate air (HEPA) filter and an air sampling bag through a flexible medical-grade polytetrafluoroethylene (PTFE) tube with an outer diameter of 4 mm. The electronic nose consists of several main electronic and mechanical parts (i.e., power module, data acquisition system, 3-way-valve, micropump, and sensor module inside the sealed gas chamber). The sensor module comprises 10 different sensing devices arranged as an array. d HEPA filter for filtering out the particulate matters and trapping the SARS-CoV-2 available from the exhaled breath of a patient confirmed with positive COVID-19. e, f Scanning electron micrographs of the fiber filter used in the GeNose C19-filtering system.

Fig. 3. GeNose 19 sensor responses and their extracted features from the exhaled breaths of RT-qPCR-confirmed negative and positive COVID patients.

Typical sensing responses were obtained from 10 different conductometric gas sensors (S1–S10), which are integrated into a portable GeNose C19 system for the exhaled breaths of RT-qPCR-confirmed a negative and b positive COVID-19 patients. Boxplots of the distribution for negative and positive COVID-19 samples based on feature extraction: c maximum, d median, e standard deviation, and f variance values.

Fig. 4. Exhaled breath collection procedure performed in the profiling test.

Breath sampling procedure comprising patients with RT-qPCR-confirmed a positive (n_p(+)) and b negative (n_p(−)) COVID-19 infection in two hospitals (i.e., Bhayangkara General Hospital in Sleman District (RS Bhayangkara) and Bambanglipuro COVID-19 Special Field Hospital in Bantul District (RSLKC Bantul)). Both are located in the Special Region of Yogyakarta, Indonesia. The total breath samples were obtained by excluding the invalid ones measured by GeNose C19, in which the total confirmed positive and negative COVID-19 samples were n_b(+) = 333 and n_b(−) = 282, respectively.

Fig. 5. Measured breath data analysis using machine learning.

a Classification of two different exhaled breath samples (i.e., positive and negative COVID-19 samples) using the LDA model. b Overall accuracy (micro-averaged F1-score), c and d are sensitivity and specificity, respectively, of the training (10-fold cross-validation), testing, and all datasets obtained by four different machine learning algorithms (LDA, SVM, MLP, and DNN). e and f are the receiver operating characteristic (ROC) curves of the training and testing data, respectively, obtained by four different machine learning algorithms (LDA, SVM, MLP, and DNN). AUC and confidence intervals are also shown.