Asthma and chronic obstructive pulmonary disease (COPD) are heterogeneous diseases with high pathological burden and healthcare costs [1-3]. In outpatient clinical practice, an accurate differential diagnosis is often very difficult, particularly in adult smokers, requiring specific lung function tests [4, 5]. Since nuclear magnetic resonance (NMR)-based metabolomics of exhaled breath condensate (EBC) discriminates adults with COPD [6-8] or asthma [9] from healthy subjects, we hypothesised that it is also able to differentiate asthma and COPD patients of different severities. After approval by the Maugeri ethics committee, we recruited prospectively patients with a new diagnosis of asthma (n=31) and COPD (n=44) according to current Global Initiative for Asthma and Global Initiative for Chronic Obstructive Lung Disease guidelines. Six patients with asthma and nine COPD patients were excluded because of relevant comorbidities that could potentially affect the analysis (nine for the presence of coronary or valvular heart disease, five for the presence of diabetes mellitus, one for hypothyroidism). In addition, the EBC samples obtained from five asthma and three COPD subjects were technically unsuitable for NMR analysis. The final 20 asthma and 32 COPD subjects were used to build the reference statistical model (table 1). Since no a priori analysis was possible, we could only evaluate the adequacy of our sample size a posteriori, estimating a sample size of 17±3 asthma and 23±3 COPD patients. A second cohort was also enrolled for an external blind validation comprising 13 asthma and 20 COPD patients. EBC collection was achieved with a TURBO-DECCS condenser (Medivac, Pilastrello, Italy) as reported [10]. NMR spectra were recorded at 27°C in a randomised sequential order on a 600-MHz Bruker Avance-III spectrometer (Bruker BioSpin GmbH, Rheinstetten, Germany) equipped with a CryoProbe using standard experiments. Metabolites were identified by resorting to two-dimensional experiments. Within-day, between-day, and technical repeatability, and detection limit were assessed as reported [7, 9]. Proton NMR spectra were automatically data reduced to 390 integral segments ("buckets"), each of 0.02 ppm, using the Bruker AMIX 3.6 software package. Unsupervised principal component analysis (PCA) was first applied. However, to better identify clustering, we used orthogonal projections to latent structures discriminant analysis (OPLS-DA), and the obtained model showed improved predictive and interpretive abilities, and in a permutation test (n=300) revealed no overfit. The model quality was evaluated via the goodness-of-fit (R 2) and the goodness-of-prediction (Q 2) parameters [11]. Metabolite quantification was obtained using the corresponding normalised buckets. Metabolite statistical significance was determined by parametric (t-test) or non-parametric (Mann−Whitney U-test) tests according to the results of a normality test performed to evaluate each distribution (Shapiro−Wilk, Kolgomorov−Smirnov test). p-values <0.05 were conside...