Machine-learning-assisted sensor array for detecting COVID-19 through simulated exhaled air

Abstract

Reverse transcription polymerase chain reaction (RT-PCR), the primary test for COVID-19, requires complicated sample collection and several hours to obtain results. Breath test for exhaled volatile organic compounds (VOCs) has gained substantial research attention as a simple non-invasive and fast screening method. However, a unique VOC fingerprint as a potential prognostic biomarker is still unavailable. Accordingly, this study prepared simulated VOC gases to test the classification performance of a sensing system. The simulated VOC gases were selected according to the actual composition of the exhaled breath of patients with acute respiratory diseases, including COVID-19. Two sets of metal oxide sensor arrays, comprising eight commercial sensors and eight Advanced Institute of Science and Technology (AIST) laboratory-made sensors, were prepared to test their sensing ability for the simulated gases. The principal component analysis (PCA) results revealed that the AIST sensors had better sensing ability than the commercial sensors. Moreover, the recursive feature elimination cross-validation (RFECV) of random forest (RF) further confirmed the superiority of the AIST sensors over the commercial sensors. An artificial neural network (ANN) with excellent prediction performance for gas concentration was developed. This study provides a promising method for rapidly screening respiratory diseases.