Machine learning enabled detection of COVID-19 pneumonia using exhaled breath analysis: a proof-of-concept study

Area of Science:

• Respiratory Medicine
• Analytical Chemistry
• Computational Biology

Background:

• Real-time-reverse-transcriptase polymerase chain reaction (RT-PCR) for SARS-CoV-2 has limitations, including high false-negative rates in lower airway infections.
• A safe, simple, and accessible method for sampling lower airways is needed for early COVID-19 pneumonia diagnosis.

Purpose of the Study:

• To evaluate the efficacy of laser absorption spectroscopy (LAS) analysis of exhaled breathprints, combined with machine learning (ML), for detecting SARS-CoV-2.

Main Methods:

• A prospective observational study enrolled patients with confirmed SARS-CoV-2 and a control group.
• Breath samples were analyzed using LAS to detect volatile organic compounds (VOCs).
• Machine learning models classified unique LAS-spectra patterns (breathprints) to identify SARS-CoV-2.

Main Results:

• LAS-breathprints with ML achieved 72.2%-81.7% accuracy in differentiating SARS-CoV-2 positive from negative individuals.
• Performance remained consistent across various patient subgroups and SARS-CoV-2 variants.
• This approach outperformed a VOC-trained classifier model (63%-74.7% accuracy).

Conclusions:

• ML-based LAS breath analysis offers a valuable, non-invasive method for detecting SARS-CoV-2 and potentially other respiratory pathogens.
• This technology is easily deployable, scalable, and versatile for rapid detection and future outbreak preparedness.