Rapid screening of COVID-19 is key to controlling the pandemic. However, current nucleic acid amplification involves lengthy procedures in addition to the discomfort of taking throat/nasal swabs. Here we describe potential breath-borne volatile organic compound (VOC) biomarkers together with machine learning that can be used for point-of-care screening of COVID-19. Using a commercial gas chromatograph-ion mobility spectrometer, higher levels of propanol were detected in the exhaled breath of COVID-19 patients (N= 74) and non-COVID-19 respiratory infections (RI) (N= 30) than those of non-COVID-19 controls (NC)/health care workers (HCW) (N= 87), and backgrounds (N= 87). In contrast, breath-borne acetone was found to be significantly lower for COVID-19 patients than other subjects. Twelve key endogenous VOC species using supervised machine learning models (support vector machines, gradient boosting machines (GBMs), and Random Forests) were shown to exhibit strong capabilities in discriminating COVID-19 from (HCW + NC) and RI with a precision ranging from 91% to 100%. GBM and Random Forests models can also discriminate RI patients from healthy subjects with a precision of 100%. In addition, the developed models using breath-borne VOCs could also detect a confirmed COVID-19 patient but with a false negative throat swab polymerase chain reaction test. It takes 10 min to allow an entire breath test to finish, including analysis of the 12 key VOC species. The developed technology provides a novel concept for non-invasive rapid point-of-care-test screening for COVID-19 in various scenarios.
Keywords: COVID-19; acetone; biomarkers; exhaled breath; machine learning; propanol; volatile organic compounds (VOCs).
Creative Commons Attribution license.