Statistical analysis
Statistical analysis was performed in IBM SPSS Statistics for Windows version 28.0.1.0 (IBM Corp., NY, USA) for both GC-MS and SIFT-MS data. Kruskal-Wallis tests with pairwise comparisons were performed to determine significant differences between recovered, mild rattling, moderate rattling and severe rattling infants for each of the 32 VOCs detected by GC-MS. For SIFT-MS data, an additional dimension reduction was performed on all 167 detected features by means of principal component analysis. Again Kruskal-Wallis tests with pairwise comparisons were performed to determine significant differences between recovered, mild rattling, moderate rattling and severe rattling infants, for each of the principal components (PC). Discriminant analysis with leave-one-out cross-validation was performed with the significantly different VOCs for the GC-MS data and with the significantly different PCs for the SIFT-MS data. The quality of the discrimination models (DMs) was evaluated with receiver operating characteristics (ROC) curve analysis.
To investigate whether PRO parameters could improve the DM based on the VOCs for GC-MS data or the PCs for SIFT-MS data, Pearson correlation coefficients were used to assess the correlation between the paediatrician’s rattle score and the most clinically relevant parameters extracted from the questionnaires (age at onset of noisy breathing, age at first cold, number of colds, PROb and PROs. Furthermore, Kruskal-Wallis tests were used to assess if those clinically relevant parameters were significantly different between the four groups (recovered, mild rattling, moderate rattling and severe rattling infants). In case a PRO parameter correlated with the paediatrician’s rattle score or was significantly different between the four groups, is was added to the DMs. For comparison, a third DM including only the PRO parameter(s) was used.