Combination of exhaled volatile organic compounds with serum biomarkers predicts respiratory infection severity

Abstract

Objective: During respiratory infections, host-pathogen interaction alters metabolism, leading to changes in the composition of expired volatile organic compounds (VOCs) and soluble immunomodulators. This study aims to identify VOC and blood biomarker signatures to develop machine learning-based prognostic models capable of distinguishing infections with similar symptoms.

Methods: Twenty-one VOCs and fifteen serum biomarkers were quantified in samples from 86 COVID-19 patients, 75 patients with non-COVID-19 respiratory infections, and 72 healthy donors. The populations were categorized into severity subgroups based on their oxygen support requirements. Descriptive and statistical analyses were conducted to assess group differentiation. Additionally, machine learning classifiers were developed to predict disease severity in both COVID-19 and non-COVID-19 patients.

Results: VOC and biomarker profiles differed significantly among groups. Random Forest models demonstrated the best performance for severity prediction. The COVID-19 model achieved 93% accuracy, 100% sensitivity, and 89% specificity, identifying IL-6, IL-8, thrombomodulin, and toluene as key severity predictors. In non-COVID-19 patients, the model reached 89% accuracy, 100% sensitivity, and 67% specificity, with CXCL10 and methyl-isobutyl-ketone as key markers.

Conclusion: VOCs and serum biomarkers differentiated HD, COVID-19, and non-COVID-19 patients, and enabled the development of high-performance severity prediction models. While promising, these findings require validation in larger independent cohorts.

Keywords: Machine learning; respiratory infections; serum biomarkers; severity prediction; volatile organic compounds.