Fast and Non-Invasive Electronic Nose Devices for Screening Out COVID-19 Virus Infection Based on Exhaled Breath VOC Detection

Current gene-based PCR diagnostics involving reverse-transcription polymerase chain reaction (RT-PCR) require at least several hours, expensive tools, and complicated sample collection methods to obtain results. A test for detecting volatile organic compounds (VOCs) in exhaled breath is advantageous as a simple, non-invasive, and fast screening method. In this study, a VOC detection system of array sensors was applied for the classification of breath control and COVID-19 virus infection. The ability to classify VOCs in the breath with COVID-19 virus infection has been studied with two metal-oxide (MOX) gas sensor arrays, commercially available sensors, and in-house sensors. The dataset of gas response signals from the array-type semiconductive gas sensors of the VOC detection system was analyzed using machine learning; principal component analysis (PCA) was used as a dimensionality-reduction method, and random forest (RF) and a convolutional neural network (CNN) were used as classification methods for the VOC concentration patterns in each breath. For the RF model, the accuracy results for the classification by two gas sensor arrays was 0.917 and this was improved by CO2 calibration to 0.967, and the feature importance analysis revealed the importance of specific gas sensors. For the CNN, an input layer of a transformed gray-scale image with the shape of 12 data points × 8 sensors was used, and its accuracy reached 100% within a relatively small number of epochs, demonstrating a short training time, which is beneficial for breath detectors or e-nose devices.