Machine learning-enabled diagnosis of viral respiratory infections from exhaled volatile organic compound analysis.

BACKGROUND: The sensitivity and specificity of current breath biomarkers are often inadequate for effective requisite sensitivity for early-stage detection, thereby ignoring early-stage treatment in the patient. METHODS: In this study, we developed a screening model for viral respiratory infections using a combination of portable GC-MS and an artificial intelligence (AI) model. This platform employs machine learning algorithms to enhance the specificity and sensitivity of the model. Subsequently, we applied this platform to analyze 200 viral respiratory infections and normal exhaled samples. RESULTS: The diagnostic signatures, including 1-nonanethiol and 2-butanone, generated by the model effectively discriminated viral respiratory infection patients from normal controls with high sensitivity (90%), specificity (81%), and accuracy (AUC = 0.85). Furthermore, propionaldehyde and amylaldehyde, generated by the model, effectively discriminated COVID-19 from influenza A patients with sensitivity (87.5%), specificity (75%), and accuracy (AUC = 0.80). Data from UKBiobank indicated that in the volatile metabolite profiles exhaled by patients with viral respiratory infections, some characteristic components are related to the metabolic products of the host’s fatty acid β-oxidation pathway. CONCLUSION: This study presents a diagnostic model that can identify novel and feasible breath biomarkers for detecting early-stage viral respiratory infections. The promising results position the platform as an efficient noninvasive screening test for clinical applications, offering potential advancements in early detection for viral respiratory infections.