Introduction: Acute respiratory distress syndrome (ARDS) is a common cause of acute respiratory failure but remains difficult to recognize. Exhaled breath contains metabolites reflective of pulmonary inflammation. We aimed to evaluate the diagnostic accuracy of breath metabolomics for ARDS.

Methods: In this prospective two-center observational study in patients receiving invasive ventilation, breath metabolites captured on sorbent material and quantified using gas chromatography and mass spectrometry. ARDS was diagnosed using the Berlin definition by three experts. Patients with ?certain? ARDS labels from one hospital were used to train a classifier based on the five most significant breath metabolites selected by a random forest model. The diagnostic accuracy of this classifier was tested in all patients from a second hospital.

Results: 357 patients were included in the derivation cohort and 142 patients in the validation cohort. 1-methylpyrrole, 1,3,5-trifluorobenzene, methoxyacetic acid, 2-methylfuran and 2-methyl-1-propanol were included in the classifier, which had an area under the receiver operating characteristics curve (AUROCC) of 0.71 (CI: 0.63-0.78) in the derivation cohort and 0.63 (CI: 0.52-0.74) in the validation cohort. Combining the breath test with the lung injury prediction score (LIPS) increased the AUROCC to 0.75 (CI: 0.68-0.82) and 0.64 (CI: 0.53-0.75) respectively.

Conclusion: The classifier had moderate diagnostic accuracy for ARDS after external validation, combination with the LIPS slightly improved diagnostic accuracy, which is insufficient for use in clinical practice.