Figure 1

Sample collection and data analysis. Exhaled breath was sampled and analyzed using an electronic nose with a side–stream connection distal from the endotracheal tube. This resulted in a response for the 32 polymer sensors in the nano–composite sensor array. The eNose was trained using sparse–partial least square (SPLS) logistic regression with 10.000–fold cross–validation. Data from the training cohort was split into a fraction for model building and model evaluation (10 cases and 10 controls). The algorithm that provided the best internally validated diagnostic accuracy, evaluated by the area under the receiver operating characteristics curve (ROC–AUC), was selected for blind testing in the validation cohort and the ROC–AUC with optimal sensitivity and specificity was reported. Differences in the predictive algorithm between different subgroups (severity of disease, pulmonary and non–pulmonary ARDS) were analyzed using non–parametric tests and the ROC–AUC was reported. Furthermore, the ROC–AUC for distinguishing CPE and pneumonia from ARDS and moderate/severe ARDS only was calculated. A sensitivity analysis was performed using logistic regression on comorbidities that are known to influence breath prints, the PaO2/FiO2 ratio, minute volume ventilation and APACHE II and SAPS II scores.